图神经网络在流体数值计算中的网格适应能力:分辨率、局部加密与拓扑改变下的可验证精度、稳定性与效率收益
问题界定与评估判据:GNN-CFD综述、算子理论视角与离散不变性/网格失配误差的形式化刻画
把三份分组中所有'给出判据而非给出模型'的文献合并为一组,因为它们共同回答'换网格后什么东西应当保持不变、用什么指标判定收益'。GNN在非结构CFD网格上的应用综述梳理相较CNN的适配优势与训练成本高、泛化不足、效率受限等瓶颈;图网络代理预测综述补充应用侧现状;Discretization invariance 与 mismatch errors 一文直接形式化讨论离散失配误差与离散化不变性;RONOM 给出无限维算子与离散算子之间的误差界;operator-theoretic learning 一文把图网络、神经算子、神经场统一视为算子在不同离散化下的有限维表示,主张分离'所学算子'与'数组表示';neural operator is data all you 提供'数据即一切'式立场作为反面参照。本组为后续所有实证分组提供跨分辨率迁移、失配界与零样本超分辨的评价尺度。
- Application of Graph Neural Networks in Computational Fluid Mechanics(An-Yi Xiao, 2026, Applied and Computational Engineering)
- Graph neural networks for surrogate prediction of hemodynamics in thoracic aortic aneurysm: a patient-specific study(Y. Alkhrijah, R. Alfilh, Narinderjit Singh Sawaran Singh, W. Rajhi, 2026, Yu-byeon-hag)
- Neural computation as operator-theoretic learning on measures.(Ming-Shu Wang, 2026, Neural Networks)
- Discretization-invariance? On the Discretization Mismatch Errors in Neural Operators(Wen-Han Gao, Rui-Chen Xu, Yue-Fan Deng, Yi Liu, 2025, International Conference on Learning Representations)
- RONOM: Reduced-Order Neural Operator Modeling(S. Dummer, Dongwei Ye, Christoph Brune, 2025, arXiv.org)
- Neural Operator: Is data all you need to model the world? An insight into the impact of Physics Informed Machine Learning(Hrishikesh Viswanath, Md Ashiqur Rahman, A. Vyas, Andrey Shor, Beatriz Medeiros, Stephanie Hernandez, Suhas Eswarappa, Prameela, Aniket Bera, 2023, No journal)
分辨率/保真度进入网络参数的方式:多尺度图池化、多保真层级结构与跨网格规模参数迁移
合并三份分组中'以尺度层级与参数映射实现跨网格迁移'的架构工作,共同点是尺度要么被显式做成层级(池化/超节点/虚拟参考网格),要么被做成可迁移的参数映射,因而直接对应'能在新网格上运行/需微调才恢复精度'的分层判定。SGUNET 用DFS池化+参数映射函数在网格规模与分辨率之间做预训练-微调迁移(明确给出'需重新标定才能恢复精度'的一档);ME-GNN 混合图消息传递与均匀网格U-Net多尺度特征;多保真Graph U-Net在粗/细分辨率图之间共享单一网络;点云U-Net用虚拟参考网格规避插值带来的精度损失;快速高斯混合卷积核+AMG图层次+U-Net在非结构网格与未见算例上运行;MeshGIN 提出通用图网络框架;PARTNO 用物理自适应区域划分把全局注意力降到线性复杂度。本组重点核查'单元体积/邻域尺度/层级结构是否进入可迁移参数'。
- Transfer learning in Scalable Graph Neural Network for Improved Physical Simulation(Si-Qi Shen, Yu Liu, Daniel Biggs, Omar Hafez, Jiandong Yu, Wentao Zhang, Bin Cui, Jiulong Shan, 2025, arXiv.org)
- A multi-scale feature enhanced graph neural network for fluid dynamics prediction in complex geometries(Li Xiao, Tian-Yu Li, Yi-Ye Zou, Ming-Jie Zhang, Xiao-Gang Deng, 2026, arXiv.org)
- A Multi-Fidelity Graph U-Net Model for Accelerated Physics Simulations(R. J. Gladstone, H. Meidani, 2024, arXiv.org)
- A point cloud-based mesh-independent convolutional neural network frameworks for flow field prediction on variable geometries(Wontae Hwang, Sooyoung Kim, Donghyun Park, Seongim Choi, 2025, The Physics of Fluids)
- A graph neural network-based framework to identify flow phenomena on unstructured meshes(Linshan Wang, Y. Fournier, J. Wald, Y. Mesri, 2023, The Physics of Fluids)
- MeshGIN: a graph neural network framework for transonic aerodynamic predictions(Saad Hussain, Yang Pei, Bagh Hussain, Fahad Nawaz, 2026, Journal of the Brazilian Society of Mechanical Sciences and Engineering)
- PARTNO: A Neural Operator based on Physically Adaptive Region Partitioning for Solving Large-Scale Partial Differential Equations(Ji-Yan Qiu, Wu Yuan, Jian Zhang, 2026, IEEE International Joint Conference on Neural Network)
不变性与离散一致性写进表示本身:规范不变谱变换、无网格离散微分算子与拓扑-几何分离
这一组保留为独立分组,因为其贡献不是'多尺度层级',而是把不变性/恒等式约束构造进表示,从而在理论上排除网格依赖,是网格适应能力最强的架构型证据。GIST 用规范不变谱Transformer显式保证离散化不变且随网格规模线性扩展(faster_by_design 为其在交互式气动设计上的落地版本);GITO 以注意力在任意查询点实现零样本超分辨;无网格多项式矩约束工作学习只依赖局部几何的离散微分算子,因而分辨率无关并可跨粒子构型复用;RHMP 把拓扑(余边界算子)与几何(学习度量)分离,保证微分恒等式与变网格CFD稳定性;enhancing Graph U-Nets 用GMM卷积算子与噪声注入实现mesh-agnostic时空预测并与单保真度GNN/插值方案对比。
- GIST: Gauge-Invariant Spectral Transformers for Scalable Graph Neural Operators(Mattia Rigotti, Nicholas Thumiger, Thomas Frick, 2026, arXiv.org)
- Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD(Nicholas Thumiger, A. Bartezzaghi, Mattia Rigotti, Cezary Skura, Thomas Frick, Elisa Serioli, Fabrizio Arbucci, A. C. I. Malossi, 2026, arXiv.org)
- GITO: Graph-Informed Transformer Operator for Learning Complex Partial Differential Equations(M. Ramezankhani, Janak M. Patel, Anirudh Deodhar, Dagnachew Birru, 2025, arXiv.org)
- Learning Mesh-Free Discrete Differential Operators with Self-Supervised Graph Neural Networks(Lucas Gerken Starepravo, G. Fourtakas, S. Lind, A. Harish, Tianning Tang, J. R. C. King, 2026, arXiv.org)
- Learning Discrete Riemannian Metrics for Physical Fields with Cochain-Frame Equivarianc(Dong-Zhe Zheng, Christine Allen-Blanchette, 2026, arXiv.org)
- Enhancing Graph U-Nets for Mesh-Agnostic Spatio-Temporal Flow Prediction(Sunwoong Yang, R. Vinuesa, Namwoo Kang, 2024, arXiv.org)
非图对照基线:神经算子、CNN与插值后学习的跨网格零样本迁移及横向比较框架
要判断图表示是否带来'可验证'收益,必须有非图基线;本组由两份分组中的对照类文献合并而成。ReBaNO 以约简基+贪心结构实现严格离散不变并宣称优于PCA-Net、DeepONet、FNO、CNO;双分支耦合FNO以截断谱核实现30×110×5→60×220×5零样本粗到细上采样(R²≈0.90)并保持约3倍加速;另有FNO方法类工作、图-卷积网络耦合工作(对照规则栅格CNN与图消息传递的耦合方式)、层级深度学习/算子学习(DeepONet、FNO、PINO、GNN)与经典求解器比较的系统综述、以及数值一致性数据驱动模型的会议摘要。它们用于界定图表示相对插值后学习与谱算子的边际收益。
- ReBaNO: Reduced Basis Neural Operator Mitigating Generalization Gaps and Achieving Discretization Invariance(Haolan Zheng, Yan-Lai Chen, Jie-Qun Han, Yue Yu, 2025, arXiv.org)
- A Dual-Branch Coupled Fourier Neural Operator for High-Resolution Multi-Phase Flow Modeling in Porous Media(Hassan Al Hashim, Odai A. Elyas, John R. Williams, 2025, Water)
- A Fourier neural operator-based method for rapid prediction of 3D indoor airflow dynamics(Xiao-Xia Ding, Hao-Tian Zhang, Wei-Rong Zhang, Wei-Jia Zhang, Ying-Li Xuan, 2025, Building Simulation)
- Graph and convolutional neural network coupling with a high-performance large-eddy simulation solver(A. Serhani, Victor Xing, D. Dupuy, C. Lapeyre, G. Staffelbach, 2024, Computers & Fluids)
- AI-based Computational Engineering using Hierarchical Deep Learning Neural Network(Kanika Singhal, Harshita, Inderpreet Kaur, Madhav Bansal, Kirti Kushwah, 2026, Social Science Research Network)
- Session H Abstracts Numerically Consistent Data-Driven Subgrid-Scale Model for Large-Eddy Simulation(Michael Lawrence, Garcia Mentor, H. J. Bae, 2023, No journal)
网格几何与尺度特征的编码设计:坐标、流向位置、壁面符号距离、超节点层次与方程项边特征
本组聚焦'模型如何表达物理距离、邻域尺度与连接关系'这一前置机制问题,是判定换网格后精度能否直接保持的根本。具体做法包括:节点坐标+流向归一化位置、边界类型one-hot与壁面符号距离;把网格空间相关特征显式注入图并设计加权损失;以向量化边界条件作为输入;用超节点聚类划分多尺度层次以捕捉非局部压力耦合;把Navier-Stokes方程项作为边特征;用虚拟节点加密提升空间粒度(ST-GPINN)。组内多篇在非结构网格与不同几何上验证,可比对'特征携带尺度信息后精度是否不再依赖具体网格'。
- Turbulent Flow–Thermal Field Prediction Around a Pin-Fin Using Geometry-Aware Multiscale Graph Neural Network(Riddhiman Raut, Evan M. Mihalko, Amrita Basak, 2026, International Journal of Thermofluid Science and Technology)
- Grid adaptive reduced-order model of fluid flow based on graph convolutional neural network(Jiang-Zhou Peng, Yi-Zhe Wang, Si-Heng Chen, Zhi-Hua Chen, Wei-Tao Wu, N. Aubry, 2022, The Physics of Fluids)
- Spectral Geometric Deep Learning for Predicting Hydrofoil Flow Fields(A. Venter, R. Laubscher, A. Bekker, 2026, Volume 6: Polar and Arctic Sciences and Technology; Computational Mechanics and Data Science)
- KAN GNN Modeling of 2D Vessel Flow Simulation(M. Değirmenci, Elif Kayaalp Ata, F. Ersoz, Y. Genç, 2026, 2026 11th International Conference on Computer Science and Engineering (UBMK))
- PI-GNN: Physics-Informed Graph Neural Network for Super-Resolution of 4D Flow MRI(Amirkhosro Kazemi, I. Abecassis, Amir A. Amini, 2024, IEEE International Symposium on Biomedical Imaging)
- ST-GPINN: a spatio-temporal graph physics-informed neural network for enhanced water quality prediction in water distribution systems(Tian-Wei Mu, Fei-Yu Duan, Bao-Kuan Ning, Bo Zhou, Jun-Yu Liu, Man-Hong Huang, 2025, npj Clean Water)
把离散结构与守恒/边界条件写进网络:有限体积、有限元型图网络与物理信息约束嵌入
与前一组的差别在于:这一组不是加特征,而是复刻数值离散结构并强制满足守恒与边界条件,因而其网格适应性主张可被守恒性、边界精度直接检验。FVGP-Net 以有限体积数值微分与边界处理替代自动微分;HydroGraphNet 把质量守恒嵌入损失;基于网格的图网络方法给出网格到图的转换规范;GNN增强有限元以节点坐标与逐节点边界条件为输入并由FEM前后处理读写;有限体积启发型GNN与PIGNN-CFD 把守恒型离散结构与边界条件显式写入网络;可解释物理信息GNN 讨论约束可解释性与精度权衡。共同支撑'换网格后退化是否源于物理约束缺失'的归因。
- Physics-informed graph neural network based on the finite volume method for steady incompressible laminar convective heat transfer(Hai-Ming Zhang, Xin-Lin Xia, Ze-dong Wu, Xiaolei Li, 2025, The Physics of Fluids)
- A Physics-Informed Graph Neural Network Approach for Multi-Time-Step Prediction of Indoor Temperature Fields(Ziang Pang, Rui Shi, Li-Guo Zhang, 2025, Youth Academic Annual Conference of Chinese Association of Automation)
- A mesh-based Graph Neural Network approach for surrogate modeling of Lagrangian free surface fluid flows(Federico Lanteri, M. Cremonesi, 2025, Computers & Fluids)
- Graph Neural Network enhanced Finite Element modelling(Rutwik Gulakala, B. Markert, M. Stoffel, 2023, Pamm)
- Finite Volume Informed Graph Neural Network for Myocardial Perfusion Simulation(Raoul Sallé de Chou, M. Sinclair, S. Lynch, Nan Xiao, L. Najman, I. Vignon-Clementel, Hugues Talbot, 2024, International Conference on Medical Imaging with Deep Learning)
- PIGNN-CFD: A physics-informed graph neural network for rapid predicting urban wind field defined on unstructured mesh(Xu-Qiang Shao, Zhi-Jian Liu, Siqi Zhang, Zijia Zhao, Chen-Xing Hu, 2023, Building and Environment)
- Interpretable physics‐informed graph neural networks for flood forecasting(Mehdi Taghizadeh, Zanko Zandsalimi, M. Nabian, M. Shafiee-Jood, Negin Alemazkoor, 2025, Computer‐Aided Civil and Infrastructure Engineering)
等变/不变对称性与图网络的数值稳定性限制:换网格后精度退化的架构层面解释
保留为独立小分组,因为它回答的是'什么样的建模选择让模型在网格改变时仍然稳定',而非特征或约束的写法。等变与不变表示的受控对比实验说明显式建模不变量可提升长期预测精度;在瞬态热传导上对多种GNN架构与超参的系统比较发现图网络仿真与经典离散格式共享类似的稳定性限制(类比Von Neumann判据)。两者合起来为'是否满足守恒、能谱与耗散指标'提供架构层面解释,也是判断换网格退化是否源自破坏离散恒等式的关键依据。
- Importance of equivariant and invariant symmetries for fluid flow modeling(Varun Shankar, S. Barwey, Zico Kolter, Romit Maulik, V. Viswanathan, 2023, arXiv.org)
- Graph neural network structural limitation for thermal simulation and architecture optimization through rating system(Pierre Hembert, C. Ghnatios, J. Cotton, F. Chinesta, 2025, Advanced Modeling and Simulation in Engineering Sciences)
变网格适应性的核心实测证据:拓扑变化、非结构节点分布与跨分辨率层级迁移(流动类型与边界条件基本固定)
这是直接回答核心问题的证据组:三份分组中'模型结构被设计为对网格离散不敏感并给出跨网格实测'的文献合并于此。超声速流动在几何与飞行条件变化时拓扑改变,需几何注意力维持泛化;多孔介质多相流GCN在未见过的非规则域与非结构网格上运行;PhysGNN 以物理边权实现跨网格拓扑与分辨率的积分器网格无关性;轴对称热Homann流在未见几何上验证;物理信息GCN从单圆柱推广到双圆柱(几何适应性);多分辨率层级代理实现粗网格到细网格的迁移(需层级迁移/微调才恢复精度)。本组按'能在新网格上运行—能直接保持精度—需重新标定/微调才恢复精度'三档分层呈现,并与其余分组的多因素试验形成对照。
- Addressing variable mesh challenges: Generalizable supersonic flow field density prediction via graph neural networks(Ji-Ping Luo, Yin Long, Xu-Jian Zhao, 2026, The Physics of Fluids)
- Graph Convolutional Networks for Simulating Multi-phase Flow and Transport in Porous Media(Jia-Min Jiang, B. Guo, 2023, arXiv.org)
- PhysGraph: Physics-Based Integration Using Graph Neural Networks(Oshri Halimi, E. Larionov, Zohar Barzelay, Philipp Herholz, Tuur Stuyck, 2023, arXiv.org)
- Accelerating Thermal Homann Flow Simulation With Mesh-Based Graph Neural Networks(D. R. Samii, M. Tembely, 2025, IEEE Access)
- Physics-informed graph convolutional neural network for modeling fluid flow and heat convection(Jiang-Zhou Peng, Yue Hua, Yu-Bai Li, Zhi-Hua Chen, Wei-Tao Wu, N. Aubry, 2023, The Physics of Fluids)
- Multi-Hierarchical Surrogate Learning for Structural Dynamical Crash Simulations Using Graph Convolutional Neural Networks(J. Kneifl, J. Fehr, S. Brunton, J. Kutz, 2024, arXiv.org)
未见网格上的多步推进:时空图网络、图卷积LSTM与扩散式全局感受野的误差累积与长期稳定性
本组集中回答评估要求中的'多步推进':三份分组的时空与滚动预测文献合并。GCN-LSTM与GAT-LSTM在非结构网格上做时空耦合外推并报告数百倍加速;图卷积LSTM与MeshGraphNet混合结构专门抑制时序误差累积,并泛化到训练集外的网格配置与边界条件;圆柱入水多步预测明确考察误差累积与插值/外推差异;图自由扩散Transformer以全局感受野替代图消息传递,从短轨迹外推到未见雷诺数与几何;另有CO2羽流在多断层介质中的长时迁移预测。共同点是稳定性与跨条件外推能力是其适应性主张的核心证据,但守恒与能谱指标覆盖有限。
- A Spatio-Temporal Graph Neural Network for Predicting Flow Fields on Unstructured Grids with the SUBOFF Benchmark(Wei Guo, Cheng Cheng, C. Huang, Zhiqing Lu, Kang Chen, Jun Ding, 2025, Journal of Marine Science and Engineering)
- Spatio-temporal graph neural network for multi-step prediction of cylinder water entry flow fields(Chongbin Shi, Gui-Yong Zhang, Teizhi Sun, Zi-Hao Wang, Heng Wang, 2025, The Physics of Fluids)
- Long-term prediction of cylinder water entry flow fields using a physics-informed spatiotemporal graph neural network(Chongbin Shi, Gui-Yong Zhang, Teizhi Sun, Yu-Tong Sui, 2026, The Physics of Fluids)
- Learning CO2 plume migration in faulted reservoirs with Graph Neural Networks(X. Ju, F. Hamon, G. Wen, Rayan Kanfar, M. Araya-Polo, H. Tchelepi, 2023, arXiv.org)
- Fluid-DiT: Graph-Free Diffusion Transformers for Fluid Flow Simulations Learning(Shentong Mo, Guo-Lin Ke, 2026, arXiv.org)
亚格子闭合中显式滤波尺度、网格尺度与学习目标的绑定:图与ANN型SGS模型的跨分辨率迁移
把两份分组中的SGS闭合文献合并为两组中的第一组:其共同做法是把滤波宽度/网格尺度显式作为输入或训练条件,因而能回答'换网格时图表示收益能否与未解析尺度变化分离'。多尺度GNN SGS用Gaussian/box双滤波构造训练集、以多尺度Leonard应力与速度涨落为输入、通过无量纲化获得尺度等变,并在训练Re上下方及近壁区做先验、再在通道流上做后验;若干ANN型SGS模型把滤波宽度直接作为特征,非线性代数模型(MANA)用各方向局部网格宽度归一化梯度与张量基,是少数在未见网格分辨率上做后验检验的工作;混合亚滤波尺度模型与strain-rate型模型提供尺度相似项与构造对照;滤波后组分生成率模型把滤波宽度与运行网格匹配、在不重训情况下跨滤波宽度与更粗分辨率检验误差有界,并与需重划网格的CNN基线对比。
- Multi-scale graph neural network modeling of near-wall turbulence with robust flow generalizability(L. Hutin, Ye Wang, M. Tanahashi, 2026, The Physics of Fluids)
- Subgrid-scale model for large-eddy simulation of isotropic turbulent flows using an artificial neural network(Z. Zhou, G. He, Shi-Zhao Wang, G. Jin, 2019, Computers & Fluids)
- Artificial neural-network-based subgrid-scale model for large-eddy simulation of isotropic turbulence(Lei Yang, Dong Li, Kai Zhang, Kun Luo, Jian-Ren Fan, 2024, The Physics of Fluids)
- Artificial-neural-network-based subgrid-scale models in the strain-rate eigenframe for large-eddy simulation of compressible turbulent channel flow(Xingsi Ren, Dehao Xu, Jianchun Wang, Shi-Yi Chen, 2025, Physical Review Fluids)
- Artificial-neural-network-based nonlinear algebraic models for large-eddy simulation of compressible wall-bounded turbulence(Dehao Xu, Jianchun Wang, Chang-Ping Yu, Shi-Yi Chen, 2023, Journal of Fluid Mechanics)
- ARTIFICIAL NEURAL NETWORK-BASED SUBGRID-SCALE MODELS FOR LARGE-EDDY SIMULATION OF TURBULENCE 1)(Chen-Yue Xie, Ze-Long Yuan, Jianchun Wang, M. Wan, Shi-Yi Chen, 2020, No journal)
- Mixed subfilter-scale models for large-eddy simulation of decaying isotropic turbulence using an artificial neural network(Dong Li, Lei Yang, Kai Zhang, Kun Luo, Jian-Ren Fan, 2025, Computers & Fluids)
- Modeling subgrid scale production rates on complex meshes using graph neural networks(Priyabrat Dash, M. Bode, K. Aditya, 2026, arXiv.org)
- A neural-network-based subgrid-scale model trained on backward-facing-step flow for large-eddy simulation of unseen turbulent flows(Jonghwan Park, 2026, The Physics of Fluids)
- Subgrid-scale stress model for large-eddy simulation of turbulence using an artificial neural network(Lei Yang, Dong Li, Kai Zhang, 2024, No journal)
CNN/非图LES模型的滤波核、滤波尺度与雷诺数外推受控评估及超分辨粗粒化重建
与前一组的区别在于这一组以非图(CNN/全连接)模型为主,提供'换尺度/换分辨率'的受控评估范式与可对照基线,正是分离'图表示效应'与'未解析尺度效应'所必需的对照材料。内容包括:在BFS上训练并加SGS旋转增强后迁移至槽道、圆柱与Ahmed体;CNN类LES在滤波尺寸、滤波核类型与雷诺数外推上的系统评估;对ANN小尺度速度模型在不同雷诺数下核对能谱与两时间相关;无监督-监督GAN流水线针对极粗网格vLES;GNN/学习型超分辨在结构化非均匀与非结构网格上重建未解析小尺度。
- Toward neural-network-based large eddy simulation: application to turbulent channel flow(Jonghwa Park, H. Choi, 2021, Journal of Fluid Mechanics)
- Large eddy simulation of turbulent flow over a backward-facing step with a neural-network-based subgrid-scale model(Jonghwan Park, H. Choi, 2025, The Physics of Fluids)
- Extrapolation Performance of Convolutional Neural Network-Based Combustion Models for Large-Eddy Simulation: Influence of Reynolds Number, Filter Kernel and Filter Size(Geveen Arumapperuma, N. Sorace, M. Jansen, Oliver Bladek, L. Nista, Shreyans Sakhare, L. Berger, H. Pitsch, T. Grenga, A. Attili, 2025, Flow Turbulence and Combustion)
- Large eddy simulation of flow over a circular cylinder with a neural-network-based subgrid-scale model(Myunghwa Kim, Jonghwan Park, H. Choi, 2024, Journal of Fluid Mechanics)
- An artificial neural network model for recovering small-scale velocity in large-eddy simulation of isotropic turbulent flows(Jiangtao Tan, Guodong Jin, 2024, The Physics of Fluids)
- Unsupervised and supervised machine learning pipeline for super-resolution-based subgrid scale modelling in coarse-grid large-eddy simulations(Soju Maejima, S. Kawai, 2025, Journal of Fluid Mechanics)
- Super-resolution of turbulent reacting flows on complex meshes using graph neural networks(Priyabrat Dash, K. Aditya, C. Frouzakis, M. Bode, 2026, arXiv.org)
以传统数值方法为参照的离散一致性:DG/FEM求解器、收敛阶验证与自适应网格加密/网格生成
本组提供'换网格后精度是否可直接保持'的方法论支点:用图神经网络实现非结构单纯形网格上的间断Galerkin有限元求解器并验证(p+1)阶理论收敛率与算子评估吞吐,把多重网格层次做成U-Net;学习型有限元收敛性建模研究关注精度阶与离散一致性;AdaptNet 分别学习初始网格生成与基于Hessian度量的各向异性自适应加密,在训练中获得分辨率无关的网格求解能力——这直接对应'局部加密方式改变'这一子问题;另有面向网络组分输运追踪的快速传统数值方法,作为离散方案在网络拓扑变化下的参照。
- Implementing the discontinuous-Galerkin finite element method using graph neural networks with application to diffusion equations(Linfeng Li, Jiansheng Xiang, Bo-Yang Chen, C. Heaney, S. Dargaville, C. Pain, 2024, Neural Networks)
- Learning finite element convergence with the Multi-fidelity Graph Neural Network(Nolan Black, A. Najafi, 2022, Computer Methods in Applied Mechanics and Engineering)
- Graph Neural Networks for Mesh Generation and Adaptation in Structural and Fluid Mechanics(Ugo Pelissier, A. Parret-Fréaud, F. Bordeu, Y. Mesri, 2024, Mathematics)
- A rapid method for composition tracking in hydrogen-blended pipeline using Fourier neural operator(Jun-Hua Gong, Guoyun Shi, Zi-Ying Fan, Bo Yu, Yu-Jie Chen, Bin Chen, Jing-Fa Li, Bo-Hong Wang, Zongze Li, Weixin Jiang, P. Varbanov, 2024, The Physics of Fluids)
跨雷诺数、跨边界条件与复杂几何的工程泛化:多因素同时变化下的联合归因证据
合并两份分组中的工程泛化文献:这类研究同时改变网格、雷诺数、边界条件或几何,因此只能支持联合归因,在本报告中作为延伸证据并明确指出其归因局限。内容涵盖机翼瞬态流场预测(ST-SGCN)、双驱方腔与后台阶(ResMLP-GCN,训练Re=100–2000、外推Re=3000)、图网络代理预测框架、城市风场扩散生成模型(消融检验对网格变化的鲁棒性)、城市风RANS重构在留出网格几何/风向/传感器配置上零重训、冰盖PDE代理的域分解与跨子域迁移学习、溃坝洪水多尺度水力GNN(不变性原理+ghost cell时变边界条件,未见非结构网格与地形)、全球海洋预报用非结构网格token化+GNN主干在动能谱保真与地转平衡上超过结构化网格基线并显著降低显存与FLOPs。
- Fast Spatiotemporal Sequence Graph Convolutional Network-based transient flow prediction around different airfoils(Hao-Ran Xie, Zhi-Qiao Wang, Yu-Bai Li, Qiang Li, Wei-Tao Wu, Jun Han, Jiang-Zhou Peng, Yong He, 2024, The Physics of Fluids)
- A novel residual multilayer perceptron-based graph convolutional network for modeling heat transfer and fluid flow(Muhammad Hasnain, N. Anand, A. Srinivasa, Pratyush Kumar, 2025, The Physics of Fluids)
- A Graph Neural Network Based Surrogate Model for Flapping-Driven Ocean Energy Harvesting(Aarshana R. Parekh, Rui Gao, R. Jaiman, 2025, Volume 4: CFD, FSI and AI)
- Generative Urban Flow Modeling: From Geometry to Airflow with Graph Diffusion(F. Giral, Álvaro Manzano, Ignacio Gómez, P. Koumoutsakos, S. Le Clainche, 2025, arXiv.org)
- GenDA: Generative Data Assimilation on Complex Urban Areas via Classifier-Free Diffusion Guidance(F. Giral, Álvaro Manzano, Ignacio Gómez, R. Vinuesa, S. Le Clainche, 2026, arXiv.org)
- Domain-Decomposed Graph Neural Network Surrogate Modeling for Ice Sheets(A. Propp, M. Perego, E. Cyr, A. Gruber, Amanda A. Howard, Alexander Heinlein, P. Stinis, D. Tartakovsky, 2025, arXiv.org)
- Multi-scale hydraulic graph neural networks for flood modelling(R. Bentivoglio, E. Isufi, S. Jonkman, R. Taormina, 2025, Natural Hazards and Earth System Sciences)
- OceanLight: Efficient Global Ocean Forecasting via Geometry-Adaptive Unstructured Mesh Representation(Wei Wu, Xiang Wang, H. Leng, Qing-Ye Min, Jun-Xing Zhu, Jun-Qiang Song, 2026, arXiv.org)
大规模网格上的效率收益:分布式消息传递与图神经网络加速线性求解器
本组直接对应研究问题中的'效率收益'与网格拓扑改变下的实用性。D-MGN 用顶点割分区的分布式消息传递缓解大网格图上的通信开销并给出多卡加速比;基于图卷积网络的代数多重网格光滑子直接面向非结构网格的不规则性,预测稀疏伪逆算子系数,并报告在比训练所见网格大128倍的网格上仍保持效率、能在AirfRANS等未见工业算例上加速收敛。两者构成'网格规模/拓扑变化时图表示带来可测效率收益'的最直接证据,同时提示效率收益须与精度-稳定性代价联合评价。
- D-MGN: A Distributed Mesh Graph Neural Network for Scalable Engineering Simulation(Fu Lin, Jia-Sheng Shi, Wei-Xiong Rao, Jia-Chuan Wang, Lei Chen, Chunyan Zhu, Haihua Wang, Suo Xie, 2025, International Conference on Distributed Computing Systems Workshops)
- Acceleration of an algebraic multigrid pressure solver using graph neural networks(Eric Chill'on, A. Lidtke, D. A. K. Nguyen, B. Font, 2026, arXiv.org)
部署侧外围证据:洪水、地下水与网格质量评估中的效率与长时滚动稳定性(缺乏单因素网格变化归因)
本组是应用与工具侧证据,网格分辨率/拓扑作为背景条件存在但未做'仅改网格'的受控对比,因此在报告中定位为延伸证据。FloodGNN-GRU 在LISFLOOD-FP结构化网格上报告100倍训练、1000倍推理加速;面向非结构网格城市排水的时空图网络报告R²=0.9913;地下水多层蒙特卡洛元模型展示元建模降本;MTGNet 以GNN做多标签网格质量评估,为变网格前处理侧提供线索。合并理由是它们共同说明图表示的工程价值,同时暴露'精度收益是否源于图表示'这一归因链尚未闭合。
- FloodGNN-GRU: a spatio-temporal graph neural network for flood prediction(Arnold Kazadi, James Doss‐Gollin, Antonia Sebastian, Arlei Silva, 2024, Environmental Data Science)
- Analyzing Spatio-Temporal Graph Neural Networks to Predict Urban Flooding on Unstructured Mesh(Jae-Geun Jang, Young-Woo Kwon, 2025, BigData Congress [Services Society])
- Groundwater Flow Meta-model for Multilevel Monte Carlo Methods(Martin Špetlík, J. Březina, 2021, Conference on Theory and Practice of Information Technologies)
- MTGNet: multi-label mesh quality evaluation using topology-guided graph neural network(Haoxuan Zhang, Hai-Sheng Li, Xiao-Qun Wu, Nan Li, 2024, Engineering computations)
合并三份初始化分组后共得15个一级分组、86篇文献(去重后全部覆盖),按"判据→架构→编码→实测→闭合→参照→泛化→效率"的证据链排列。第1组(综述、算子理论与离散失配误差形式化)界定"换网格后应保持什么、用什么指标判定收益";第2、3组处理架构层的离散无关性——前者以多尺度池化/多保真层级与参数映射实现跨网格规模迁移(SGUNET明确落在"需微调才恢复精度"一档),后者把不变性、微分恒等式与拓扑-几何分离构造进表示,属最强主张;第4组提供非图对照基线(ReBaNO的严格离散不变、双分支FNO的零样本粗到细上采样、图-卷积耦合与算子学习综述),用于界定图表示的边际收益;第5、6组解剖网格属性如何被表达——第5组是坐标/流向位置/壁面符号距离/超节点层次/方程项边特征的特征工程路径,第6组是复刻有限体积-有限元离散结构并强制守恒与边界条件的约束嵌入路径,二者与第7组的等变对称性与稳定性限制(类比Von Neumann判据)共同解释换网格后退化的机制来源。第8组是核心实测证据:在流动类型与边界条件基本固定、单独改变拓扑/节点分布/分辨率/几何的设定下给出分层结果,其中"能在新网格上运行"证据充分、"直接保持精度"仅见于把不变量或离散结构写进模型的少数工作、更多工作需要几何注意力、物理边权、层级迁移或重标定才能恢复精度。第9组补齐"未见网格上多步推进"的评估维度(时空图网络、图卷积LSTM、图自由扩散Transformer),第10、11组构成湍流闭合的完整对照链:图/ANN型SGS把滤波宽度与网格尺度显式作为输入或训练条件,通过双滤波、多滤波尺度、尺度化输入与Leonard/尺度相似项来补偿网格效应,并以不重训跨滤波宽度的检验和需重划网格的CNN基线做对照;CNN/非图LES组则提供滤波核、滤波尺度与雷诺数外推的受控评估范式与超分辨重建证据。第12组以DG/FEM求解器、收敛阶验证与Hessian度量自适应加密作为传统数值参照,是"局部加密方式改变"子问题的直接材料。第13组(跨雷诺数、跨边界条件、复杂几何)与第15组(洪水/地下水部署、网格质量评估)因同时改变多个因素或缺乏单因素网格对比,只能支持联合归因,报告中作为延伸证据并明确其局限;第14组给出分布式消息传递与GNN加速代数多重网格这两条最直接的大网格效率证据。总体结论:非结构图表示在"可运行性"和"效率"上证据较充分,"直接保持精度"分歧明显且多数依赖尺度化输入、物理约束或微调,而严格保持流动类型、雷诺数与物理边界条件一致、仅改变网格分辨率/局部加密/拓扑的受控实验,以及跨网格多步推进中的守恒、能谱与耗散评估,仍是最值得补齐的空白。
把三份分组中所有'给出判据而非给出模型'的文献合并为一组,因为它们共同回答'换网格后什么东西应当保持不变、用什么指标判定收益'。GNN在非结构CFD网格上的应用综述梳理相较CNN的适配优势与训练成本高、泛化不足、效率受限等瓶颈;图网络代理预测综述补充应用侧现状;Discretization invariance 与 mismatch errors 一文直接形式化讨论离散失配误差与离散化不变性;RONOM 给出无限维算子与离散算子之间的误差界;operator-theoretic learning 一文把图网络、神经算子、神经场统一视为算子在不同离散化下的有限维表示,主张分离'所学算子'与'数组表示';neural operator is data all you 提供'数据即一切'式立场作为反面参照。本组为后续所有实证分组提供跨分辨率迁移、失配界与零样本超分辨的评价尺度。相关文献:An-Yi Xiao et. al, 2026 等 6 篇文献
合并三份分组中'以尺度层级与参数映射实现跨网格迁移'的架构工作,共同点是尺度要么被显式做成层级(池化/超节点/虚拟参考网格),要么被做成可迁移的参数映射,因而直接对应'能在新网格上运行/需微调才恢复精度'的分层判定。SGUNET 用DFS池化+参数映射函数在网格规模与分辨率之间做预训练-微调迁移(明确给出'需重新标定才能恢复精度'的一档);ME-GNN 混合图消息传递与均匀网格U-Net多尺度特征;多保真Graph U-Net在粗/细分辨率图之间共享单一网络;点云U-Net用虚拟参考网格规避插值带来的精度损失;快速高斯混合卷积核+AMG图层次+U-Net在非结构网格与未见算例上运行;MeshGIN 提出通用图网络框架;PARTNO 用物理自适应区域划分把全局注意力降到线性复杂度。本组重点核查'单元体积/邻域尺度/层级结构是否进入可迁移参数'。相关文献:Si-Qi Shen et. al, 2025 等 7 篇文献
这一组保留为独立分组,因为其贡献不是'多尺度层级',而是把不变性/恒等式约束构造进表示,从而在理论上排除网格依赖,是网格适应能力最强的架构型证据。GIST 用规范不变谱Transformer显式保证离散化不变且随网格规模线性扩展(faster_by_design 为其在交互式气动设计上的落地版本);GITO 以注意力在任意查询点实现零样本超分辨;无网格多项式矩约束工作学习只依赖局部几何的离散微分算子,因而分辨率无关并可跨粒子构型复用;RHMP 把拓扑(余边界算子)与几何(学习度量)分离,保证微分恒等式与变网格CFD稳定性;enhancing Graph U-Nets 用GMM卷积算子与噪声注入实现mesh-agnostic时空预测并与单保真度GNN/插值方案对比。相关文献:Mattia Rigotti et. al, 2026 等 6 篇文献
要判断图表示是否带来'可验证'收益,必须有非图基线;本组由两份分组中的对照类文献合并而成。ReBaNO 以约简基+贪心结构实现严格离散不变并宣称优于PCA-Net、DeepONet、FNO、CNO;双分支耦合FNO以截断谱核实现30×110×5→60×220×5零样本粗到细上采样(R²≈0.90)并保持约3倍加速;另有FNO方法类工作、图-卷积网络耦合工作(对照规则栅格CNN与图消息传递的耦合方式)、层级深度学习/算子学习(DeepONet、FNO、PINO、GNN)与经典求解器比较的系统综述、以及数值一致性数据驱动模型的会议摘要。它们用于界定图表示相对插值后学习与谱算子的边际收益。相关文献:Haolan Zheng et. al, 2025 等 6 篇文献
本组聚焦'模型如何表达物理距离、邻域尺度与连接关系'这一前置机制问题,是判定换网格后精度能否直接保持的根本。具体做法包括:节点坐标+流向归一化位置、边界类型one-hot与壁面符号距离;把网格空间相关特征显式注入图并设计加权损失;以向量化边界条件作为输入;用超节点聚类划分多尺度层次以捕捉非局部压力耦合;把Navier-Stokes方程项作为边特征;用虚拟节点加密提升空间粒度(ST-GPINN)。组内多篇在非结构网格与不同几何上验证,可比对'特征携带尺度信息后精度是否不再依赖具体网格'。相关文献:Riddhiman Raut et. al, 2026 等 6 篇文献
与前一组的差别在于:这一组不是加特征,而是复刻数值离散结构并强制满足守恒与边界条件,因而其网格适应性主张可被守恒性、边界精度直接检验。FVGP-Net 以有限体积数值微分与边界处理替代自动微分;HydroGraphNet 把质量守恒嵌入损失;基于网格的图网络方法给出网格到图的转换规范;GNN增强有限元以节点坐标与逐节点边界条件为输入并由FEM前后处理读写;有限体积启发型GNN与PIGNN-CFD 把守恒型离散结构与边界条件显式写入网络;可解释物理信息GNN 讨论约束可解释性与精度权衡。共同支撑'换网格后退化是否源于物理约束缺失'的归因。相关文献:Hai-Ming Zhang et. al, 2025 等 7 篇文献
保留为独立小分组,因为它回答的是'什么样的建模选择让模型在网格改变时仍然稳定',而非特征或约束的写法。等变与不变表示的受控对比实验说明显式建模不变量可提升长期预测精度;在瞬态热传导上对多种GNN架构与超参的系统比较发现图网络仿真与经典离散格式共享类似的稳定性限制(类比Von Neumann判据)。两者合起来为'是否满足守恒、能谱与耗散指标'提供架构层面解释,也是判断换网格退化是否源自破坏离散恒等式的关键依据。相关文献:Varun Shankar et. al, 2023 等 2 篇文献
这是直接回答核心问题的证据组:三份分组中'模型结构被设计为对网格离散不敏感并给出跨网格实测'的文献合并于此。超声速流动在几何与飞行条件变化时拓扑改变,需几何注意力维持泛化;多孔介质多相流GCN在未见过的非规则域与非结构网格上运行;PhysGNN 以物理边权实现跨网格拓扑与分辨率的积分器网格无关性;轴对称热Homann流在未见几何上验证;物理信息GCN从单圆柱推广到双圆柱(几何适应性);多分辨率层级代理实现粗网格到细网格的迁移(需层级迁移/微调才恢复精度)。本组按'能在新网格上运行—能直接保持精度—需重新标定/微调才恢复精度'三档分层呈现,并与其余分组的多因素试验形成对照。相关文献:Ji-Ping Luo et. al, 2026 等 6 篇文献
本组集中回答评估要求中的'多步推进':三份分组的时空与滚动预测文献合并。GCN-LSTM与GAT-LSTM在非结构网格上做时空耦合外推并报告数百倍加速;图卷积LSTM与MeshGraphNet混合结构专门抑制时序误差累积,并泛化到训练集外的网格配置与边界条件;圆柱入水多步预测明确考察误差累积与插值/外推差异;图自由扩散Transformer以全局感受野替代图消息传递,从短轨迹外推到未见雷诺数与几何;另有CO2羽流在多断层介质中的长时迁移预测。共同点是稳定性与跨条件外推能力是其适应性主张的核心证据,但守恒与能谱指标覆盖有限。相关文献:Wei Guo et. al, 2025 等 5 篇文献
把两份分组中的SGS闭合文献合并为两组中的第一组:其共同做法是把滤波宽度/网格尺度显式作为输入或训练条件,因而能回答'换网格时图表示收益能否与未解析尺度变化分离'。多尺度GNN SGS用Gaussian/box双滤波构造训练集、以多尺度Leonard应力与速度涨落为输入、通过无量纲化获得尺度等变,并在训练Re上下方及近壁区做先验、再在通道流上做后验;若干ANN型SGS模型把滤波宽度直接作为特征,非线性代数模型(MANA)用各方向局部网格宽度归一化梯度与张量基,是少数在未见网格分辨率上做后验检验的工作;混合亚滤波尺度模型与strain-rate型模型提供尺度相似项与构造对照;滤波后组分生成率模型把滤波宽度与运行网格匹配、在不重训情况下跨滤波宽度与更粗分辨率检验误差有界,并与需重划网格的CNN基线对比。相关文献:L. Hutin et. al, 2026 等 10 篇文献
与前一组的区别在于这一组以非图(CNN/全连接)模型为主,提供'换尺度/换分辨率'的受控评估范式与可对照基线,正是分离'图表示效应'与'未解析尺度效应'所必需的对照材料。内容包括:在BFS上训练并加SGS旋转增强后迁移至槽道、圆柱与Ahmed体;CNN类LES在滤波尺寸、滤波核类型与雷诺数外推上的系统评估;对ANN小尺度速度模型在不同雷诺数下核对能谱与两时间相关;无监督-监督GAN流水线针对极粗网格vLES;GNN/学习型超分辨在结构化非均匀与非结构网格上重建未解析小尺度。相关文献:Jonghwa Park et. al, 2021 等 7 篇文献
本组提供'换网格后精度是否可直接保持'的方法论支点:用图神经网络实现非结构单纯形网格上的间断Galerkin有限元求解器并验证(p+1)阶理论收敛率与算子评估吞吐,把多重网格层次做成U-Net;学习型有限元收敛性建模研究关注精度阶与离散一致性;AdaptNet 分别学习初始网格生成与基于Hessian度量的各向异性自适应加密,在训练中获得分辨率无关的网格求解能力——这直接对应'局部加密方式改变'这一子问题;另有面向网络组分输运追踪的快速传统数值方法,作为离散方案在网络拓扑变化下的参照。相关文献:Linfeng Li et. al, 2024 等 4 篇文献
合并两份分组中的工程泛化文献:这类研究同时改变网格、雷诺数、边界条件或几何,因此只能支持联合归因,在本报告中作为延伸证据并明确指出其归因局限。内容涵盖机翼瞬态流场预测(ST-SGCN)、双驱方腔与后台阶(ResMLP-GCN,训练Re=100–2000、外推Re=3000)、图网络代理预测框架、城市风场扩散生成模型(消融检验对网格变化的鲁棒性)、城市风RANS重构在留出网格几何/风向/传感器配置上零重训、冰盖PDE代理的域分解与跨子域迁移学习、溃坝洪水多尺度水力GNN(不变性原理+ghost cell时变边界条件,未见非结构网格与地形)、全球海洋预报用非结构网格token化+GNN主干在动能谱保真与地转平衡上超过结构化网格基线并显著降低显存与FLOPs。相关文献:Hao-Ran Xie et. al, 2024 等 8 篇文献
本组直接对应研究问题中的'效率收益'与网格拓扑改变下的实用性。D-MGN 用顶点割分区的分布式消息传递缓解大网格图上的通信开销并给出多卡加速比;基于图卷积网络的代数多重网格光滑子直接面向非结构网格的不规则性,预测稀疏伪逆算子系数,并报告在比训练所见网格大128倍的网格上仍保持效率、能在AirfRANS等未见工业算例上加速收敛。两者构成'网格规模/拓扑变化时图表示带来可测效率收益'的最直接证据,同时提示效率收益须与精度-稳定性代价联合评价。相关文献:Fu Lin et. al, 2025 等 2 篇文献
本组是应用与工具侧证据,网格分辨率/拓扑作为背景条件存在但未做'仅改网格'的受控对比,因此在报告中定位为延伸证据。FloodGNN-GRU 在LISFLOOD-FP结构化网格上报告100倍训练、1000倍推理加速;面向非结构网格城市排水的时空图网络报告R²=0.9913;地下水多层蒙特卡洛元模型展示元建模降本;MTGNet 以GNN做多标签网格质量评估,为变网格前处理侧提供线索。合并理由是它们共同说明图表示的工程价值,同时暴露'精度收益是否源于图表示'这一归因链尚未闭合。相关文献:Arnold Kazadi et. al, 2024 等 4 篇文献
总计 87 篇相关文献
No abstract available
With the success of deep learning, researchers have developed Graph Neural Networks (GNNs) in the domain of engineering simulation. To represent simulation objects, GNN models are built on mesh graphs consisting of mesh elements. Unfortunately, due to the difference between engineering simulation and well-known GNN applications such as social networks, existing distributed GNN models do not perform well on largescale mesh graphs, suffering from intensive message passing and communication overhead across distributed workers. To tackle this issue, in this paper, we propose a novel distributed mesh graph network D-MGN. By a vertex-cut partitioning policy, D-MGN divides large mesh graphs into partitions. After the partitions are assigned to distributed workers, D-MGN allows local communication-free message passing within each worker, and performs distributed message aggregation with rather low communication cost. Evaluation on an example dataset demonstrates that D-MGN outperforms a recent work SAC by 38.66% lower errors, $\mathbf{5 4 . 6 7 \%}$ faster convergence time, and $\mathbf{1 1 9 . 2 2 \%}$ higher speedup ratio on 8 GPUs.
The finite element discretization of computational physics problems frequently involves the manual generation of an initial mesh and the application of adaptive mesh refinement (AMR). This approach is employed to selectively enhance the accuracy of resolution in regions that encompass significant features throughout the simulation process. In this paper, we introduce Adaptnet, a Graph Neural Networks (GNNs) framework for learning mesh generation and adaptation. The model is composed of two GNNs: the first one, Meshnet, learns mesh parameters commonly used in open-source mesh generators, to generate an initial mesh from a Computer Aided Design (CAD) file; while the second one, Graphnet, learns mesh-based simulations to predict the components of an Hessian-based metric to perform anisotropic mesh adaptation. Our approach is tested on structural (Deforming plate–Linear elasticity) and fluid mechanics (Flow around cylinders–steady-state Stokes) problems. Our findings demonstrate the model’s ability to precisely predict the dynamics of the system and adapt the mesh as needed. The adaptability of the model enables learning resolution-independent mesh-based simulations during training, allowing it to scale effectively to more intricate state spaces during inference.
No abstract available
The rapid development of deep learning has significantly influenced computational studies in convective heat transfer. To facilitate broader applications of deep learning models in convective heat transfer, this paper proposes a physics-informed graph neural network based on the finite volume method (FVGP-Net) for unsupervised training and prediction of steady incompressible laminar convective heat transfer problems. In this model, mesh data generated by the finite volume method (FVM) are converted into graph data, preserving the mesh's topological properties. This conversion allows FVGP-Net to utilize a graph convolutional network for information aggregation, capturing both local and global flow features and enhancing the model's geometric adaptability and predictive performance. The model incorporates physical laws directly into its loss function, ensuring compliance to these laws without reliance on training data. Unlike traditional physics-informed neural networks (PINNs), FVGP-Net replaces automatic differentiation with FVM-based numerical differentiation, balancing training efficiency with prediction accuracy. Boundary conditions are handled in accordance with the FVM, ensuring that the model strictly satisfies these constraints. We validated FVGP-Net using representative test cases, also examining the effects of different initialization methods on model training. The results demonstrate that FVGP-Net achieves high accuracy in predicting incompressible laminar steady convective heat transfer. Compared to traditional PINNs, this model inherits the conservation properties of the FVM, enhancing velocity prediction accuracy in convective heat transfer problems by 70.03%. Furthermore, the application of transfer learning markedly accelerates training, achieving approximately 70% faster results compared to Xavier initialization.
No abstract available
Computational engineering is undergoing a rapid transformation due to artificial intelligence (AI), which combines data-driven learning with physics-based modeling. An important advancement is the development of Hierarchical Deep Learning Neural Networks (HDLNNs), which employ layered architectures to capture multi-scale and multi-physics features. Operator learning models like DeepONet, Fourier Neural Operator (FNO), and Physics-Informed Neural Operator (PINO) built upon the ideas of early frameworks like Physics-Informed Neural Networks (PINNs), which incorporated governing equations into training, to enable resolution-invariant PDE surrogates. Hierarchical extensions like HiDeNN and Convolutional HiDeNN (C-HiDeNN) improve structural and material simulations, while Graph Neural Networks (GNNs) provide mesh-based multigrid-like learning. Despite the rapid progress, challenges remain in handling stiffness, generalization to complex geometries, and multi- fidelity integration. Through comparison with traditional solvers and a description of possible future directions for scalable, physics-consistent AI systems, this study investigates the role of HDLNNs in computational engineering.
Data-driven models often neglect the underlying physical principles, limiting generalization capabilities in water distribution systems (WDSs). This study presents a novel spatio-temporal graph physics-informed neural network (ST-GPINN) for water quality prediction in WDSs, integrating hydraulic simulations, physics-informed neural networks (PINNs), and graph neural networks (GNNs) to capture dynamics and graph-based network connectivity while approximating partial differential equations (PDEs). ST-GPINN discretizes WDSs using virtual nodes to enhance spatial granularity, employs an Encoder-Processor-Decoder architecture for predictions. Validated on Network A (a small-scale network with 9 junctions and 11 pipes) and Network B (a real large-scale WDS with 920 junctions and 1032 pipes), ST-GPINN outperforms others, achieving a MAE of 0.0073 mg/L, RMSE of 0.0121 mg/L, and R2 of 88.91% in Network A, and a MAE of 0.008 mg/L, RMSE of 0.0098 mg/L, and R² of 98.91% in Network B. Its scalability and accuracy highlight ST-GPINN’s potential for water quality predictions.
Abstract Classical approaches for flood prediction apply numerical methods for the solution of partial differential equations that capture the physics of inundation processes (e.g., the 2D Shallow Water equations). However, traditional inundation models are still unable to satisfy the requirements of many relevant applications, including early-warning systems, high-resolution (or large spatial domain) simulations, and robust inference over distributions of inputs (e.g., rainfall events). Machine learning (ML) approaches are a promising alternative to physics-based models due to their ability to efficiently capture correlations between relevant inputs and outputs in a data-driven fashion. In particular, once trained, ML models can be tested/deployed much more efficiently than classical approaches. Yet, few ML-based solutions for spatio-temporal flood prediction have been developed, and their reliability/accuracy is poorly understood. In this paper, we propose FloodGNN-GRU, a spatio-temporal flood prediction model that combines a graph neural network (GNN) and a gated recurrent unit (GRU) architecture. Compared to existing approaches, FloodGNN-GRU (i) employs a graph-based model (GNN); (ii) operates on both spatial and temporal dimensions; and (iii) processes the water flow velocities as vector features, instead of scalar features. We evaluate FloodGNN-GRU using a LISFLOOD-FP simulation of Hurricane Harvey (2017) in Houston, Texas. Our results, based on several metrics, show that FloodGNN-GRU outperforms several data-driven alternatives in terms of accuracy. Moreover, our approach can be trained 100x faster and tested 1000x faster than the time required to run a comparable simulation. These findings illustrate the potential of ML-based methods to efficiently emulate physics-based inundation models, especially for short-term predictions.
Climate change has intensified extreme weather events, with floods causing significant socioeconomic and environmental damage. Accurate flood forecasting is crucial for disaster preparedness and risk mitigation, yet traditional hydrodynamic models, while precise, are computationally prohibitive for real‐time applications. Machine learning surrogates, such as graph neural networks (GNNs), improve efficiency but often lack physical consistency and interpretability. This paper introduces HydroGraphNet, a novel physics‐informed GNN framework that, for the first time, integrates the Kolmogorov–Arnold Network (KAN) to enhance model interpretability in unstructured mesh‐based flood forecasting. The framework embeds mass conservation laws into the loss function, ensuring physically consistent predictions. Additionally, it employs an autoregressive encoder–processor–decoder architecture that captures spatiotemporal flood dynamics while mitigating error accumulation over long forecasting horizons. Validation on flood data from the White River near Muncie, Indiana, demonstrates a 67% reduction in prediction error, near‐zero mass balance error, and a 58% improvement in the critical success index for major flood events compared to a baseline GNN model. These results highlight the potential of the proposed framework to advance real‐time flood forecasting with improved physical consistency and interpretability.
Abstract. Deep-learning-based surrogate models represent a powerful alternative to numerical models for speeding up flood mapping while preserving accuracy. In particular, solutions based on hydraulic-based graph neural networks (SWE-GNNs) enable transferability to domains not used for training and allow the inclusion of physical constraints. However, these models are limited due to four main aspects. First, they cannot model rapid differences in flow propagation speeds; secondly, they can face instabilities during training when using a large number of layers, needed for effective modelling; third, they cannot accommodate time-varying boundary conditions; and fourth, they require initial conditions from a numerical solver. To address these issues, we propose a multi-scale hydraulic-based graph neural network (mSWE-GNN) that models the flood at different resolutions and propagation speeds. We include time-varying boundary conditions via ghost cells, which enforce the solution at the domain’s boundary and drop the need for a numerical solver for the initial conditions. To improve generalization over unseen meshes and reduce the data demand, we use invariance principles and make the inputs independent from coordinates' rotations. Numerical results applied to dike-breach floods show that the model predicts the full spatio-temporal simulation of the flood over unseen irregular meshes, topographies, and time-varying boundary conditions, with mean absolute errors in time of 0.05 m for water depths and 0.003 m2 s−1 for unit discharges. We further corroborate the mSWE-GNN in a realistic case study in the Netherlands and show generalization capabilities with only one fine-tuning sample, with mean absolute errors of 0.12 m for water depth, a critical success index for a water depth threshold of 0.05 m of 87.68 %, and speed-ups of over 700 times. Overall, the approach opens up several avenues for probabilistic analyses of realistic configurations and flood scenarios.
Driven by the abundant data generated from computational fluid dynamics (CFD) simulations, machine learning (ML) methods surpass the deterministic criteria on flow phenomena identification in the way that they are independent of a case-by-case threshold by combining the flow field properties and the topological distribution of the phenomena. The current most popular and successful ML models based on convolutional neural networks are limited to structured meshes and unable to directly digest the data generated from unstructured meshes, which are more widely used in real industrial CFD simulations. We proposed a framework based on graph neural networks with the proposed fast Gaussian mixture model as the convolution kernel and U-Net architecture to detect flow phenomena based on a graph hierarchy generated by the algebraic multigrid method embedded in the open-source CFD solver, code_saturne. We demonstrate the superiority of the proposed kernel and U-Net architecture, along with the generality of the framework to unstructured mesh and unseen case, on detecting the vortices once trained on a single backward-facing step case. Our proposed framework can be trivially extended to detect more flow phenomena in three dimensional cases, which is ongoing work.
Machine learning (ML) has benefited from both software and hardware advancements, leading to increasing interest in capitalising on ML throughout academia and industry. There have been efforts in the scientific computing community to leverage this development via implementing conventional partial differential equation (PDE) solvers with machine learning packages, most of which rely on structured spatial discretisation and fast convolution algorithms. However, unstructured meshes are favoured in problems with complex geometries. To bridge this gap, we propose to implement the unstructured Finite Element Method (FEM) on simplicial meshes with graph neural networks. This paper is the first to implement an unstructured mesh FEM solver using graph neural networks. All compute-intensive algorithms in the solver are represented with either convolutional or graph neural networks. Specifically, the FEM solver uses a discontinuous Galerkin formulation with an interior penalty method for spatial discretisation and a multigrid preconditioned Krylov solver as the linear solver. The multigrid method has been designed to suit the data structure within the ML package and adopts the commonly used U-Net architecture for this. A hierarchy of coarsened meshes is generated from p-multigrid and algebraic node agglomeration guided by either a space-filling curve or a smoothed aggregation algorithm. The solver is verified and assessed for solving the diffusion problems. The solver shows the theoretical convergence rate of (p+1) order. Compared with a highly optimised implementation, the solver running on GPU can reach promising throughput in terms of matrix operator evaluation at 6.8 MDOF/s. The method can easily extend to other PDEs and computing platforms beyond CPU and GPU.
Climate change and rapid urbanization have intensified extreme rainfall and urban flooding into critical global challenges, necessitating rapid and accurate flood prediction systems. While SWMM 1D offers efficient network-level flow calculations, Finite volume method (FVM)-based models required for detailed water depth prediction involve repeatedly solving 2D shallow water equations and other complex partial differential equations, resulting in substantial computational overhead that impedes real-time forecasting. Recent advances in graph neural networks (GNNs) have demonstrated superior performance for river and coastal flood prediction, yet these models typically assume structured mesh and thus fail to address the fundamental complexities of urban drainage systems including unstructured mesh topologies, complex junction networks, impervious surfaces, and potential flow reversals. This paper proposes a Spatio-Temporal Graph Neural Network specifically optimized for urban flood prediction on unstructured mesh structures. The proposed model employs a dual-branch spatial architecture combining GraphSAGE and Graph Attention Networks (GAT) to learn spatial features, an MLP-based temporal branch to capture temporal dynamics, and physics-informed edge weights integrating elevation and distance information to accelerate spatial learning. Evaluated on 144 physically simulated rainfall scenarios with 2,128 samples total in the Pohang Naengcheon watershed, the model achieves exceptional performance with $R^{2}=0.9913, R M S E=0.877 m$, and $M A E=0.679 m$ on the test set. Critically, short-term predictions within 60-minutes which are essential for operational flood warning systems consistently maintain $R^{2}>0.99$, clearly demonstrating practical applicability. By bridging the accuracy-efficiency gap between physicsbased models and deep learning, this work enables urban flood nowcasting on unstructured meshes and provides a novel approach for operational deployment.
In the interdisciplinary field of data-driven models and computational fluid mechanics, the reduced-order model for flow field prediction is mainly constructed by a convolutional neural network (CNN) in recent years. However, the standard CNN is only applicable to data with Euclidean spatial structure, while data with non-Euclidean properties can only be convolved after pixelization, which usually leads to decreased data accuracy. In this work, a novel data-driven framework based on graph convolution network (GCN) is proposed to allow the convolution operator to predict fluid dynamics on non-uniform structured or unstructured mesh data. This is achieved by the fact that the graph data inherit the spatial characteristics of the mesh and by the message passing mechanism of GCN. The conversion method from the form of mesh data to graph data and the operation mechanism of GCN are clarified. Moreover, additional relevance features and weight loss function of the dataset are also investigated to improve the model performance. The model learns an end-to-end mapping between the mesh spatial features and the physical flow field. Through our studies of various cases of internal flow, it is shown that the proposed GCN-based model offers excellent adaptability to non-uniformly distributed mesh data, while also achieving a high accuracy and three-order speedup compared with numerical simulation. Our framework generalizes the graph convolution network to flow field prediction and opens the door to further extending GCN to most existing data-driven architectures of fluid dynamics in the future.
No abstract available
Computational fluid dynamics (CFD) discretizes the Navier-Stokes equations on computational meshes to obtain flow-field quantities like velocity, pressure, and density through numerical iteration, but traditional solvers, though reliable in accuracy, are computationally expensive and time-consuming for high-Reynolds-number turbulence and complex unsteady flows. With geometric deep learning, graph neural networks (GNNs) overcome the limitation of conventional neural network (CNNs) that only handle grid-structured Euclidean data; by message-passing, they naturally fit unstructured CFD meshes and capture spatial correlations. This paper systematically reviews GNN fundamentals and core fluid mechanics, including turbulence characteristics and mesh type classifications, and then focuses on typical GNN applications in flow-field prediction, turbulence modeling, and complex-flow simulation. Existing challenges such as high training costs, insufficient generalization, and limited computational efficiency are analyzed, and promising future research directions are discussed. The review indicates GNNs are efficient surrogates for traditional solvers, offering a pathway to intelligent, fast, high-fidelity simulation.
Quantitative assessment of velocity and pressure fields in stenotic bifurcations is vital for understanding cardiovascular risks, yet conventional Computational Fluid Dynamics (CFD) simulations remain computationally expensive. We introduce GraphChebKANTransformer, a novel mesh-based hierarchical graph neural network surrogate for rapid, full-field hemodynamic predictions. The model is trained on a rigorously verified dataset of 128,400 COMSOL simulation samples, ensuring high-fidelity physical ground truths through flow-rate consistency and mesh-independence checks. Distinct from standard graph surrogates that rely on fixed multilayer perceptron (MLP) updates, our architecture replaces conventional transformations with Chebyshev Kolmogorov-Arnold Network (ChebKAN) layers. By exploiting learnable polynomial-basis kernels on unstructured meshes, the framework provides a highly flexible and compact functional basis for capturing intricate geometry-dependent flow behaviors. A major limitation of traditional node-to-node message passing is its inability to capture non-local pressure dependencies induced by localized geometric constrictions. To overcome this, our model implements a multi-scale spatial hierarchy that explicitly separates local mesh learning from long-range physical coupling. Unstructured meshes are dynamically partitioned into super-node clusters where local features are extracted via ChebKAN layers, while global interpartition reasoning is executed through attention-based pooling and self-attention mechanisms. Evaluated on 1,584 unseen test samples across 96 novel geometries, the framework achieves excellent generalization, yielding a median R2 score of 0.9980 for velocity and 0.9919 for pressure fields. This work represents the first application of ChebKAN-driven message passing to mesh-based hemodynamics, offering a fast and highly accurate paradigm for real-time vascular flow analysis.
The optimal design of marine hydrofoils remains a challenging task for naval architects, as it typically relies on computational fluid dynamics (CFD) analyses that are both time-consuming and computationally expensive. These limitations constrain the exploration of wide design spaces and slow the iterative optimization process. Accordingly, surrogate models capable of delivering rapid yet accurate flow field predictions are desired to support efficient design and evaluation workflows. This paper presents the development of a data-driven surrogate model for predicting hydrofoil flow fields in marine applications. Constructed within a spectral graph convolutional neural network (GCN) framework, the surrogate predicts distributions of velocity and pressure flow fields. The model processes graph-based representations of mesh domains together with vectorized boundary condition inputs, enabling effective application with complex, unstructured data. The model is trained on CFD-generated datasets encompassing diverse foil geometries and boundary condition combinations. The resulting surrogate achieves reasonable predictive accuracy and enables near-instant estimation of detailed flow fields for unseen marine designs. This work establishes a foundation for future integration of geometric deep learning techniques into marine design optimization and intelligent performance prediction systems.
No abstract available
To overcome the limitations of traditional convolutional and recurrent neural networks in capturing spatio-temporal dynamics in flow fields on unstructured grids, this study proposes a novel Spatio-Temporal Graph Neural Network (ST-GNN) model that integrates a Graph Neural Network (GNN) with a Long Short-Term Memory (LSTM) network. The GNN component captures spatial dependencies among irregular grid nodes via message passing, while the LSTM component models temporal evolution through gated memory mechanisms. This hybrid framework enables the joint learning of spatial and temporal features in complex flow systems. Two variants of ST-GNN, namely, GCN-LSTM and GAT-LSTM, were developed and evaluated using the SUBOFF AFF-8 benchmark dataset. The results show that GAT-LSTM achieved higher accuracy than GCN-LSTM, with average relative errors of 2.51% for velocity and 1.43% for pressure at the 1000th time step. Both models achieved substantial speedups over traditional CFD solvers, with GCN-LSTM and GAT-LSTM accelerating predictions by approximately 350 and 150 times, respectively. These findings position ST-GNN as an efficient and accurate alternative for spatio-temporal flow modeling on unstructured grids, advancing data-driven fluid dynamics.
This study introduces a Spatio-Temporal Graph Neural Network (ST-GNN) for the multi-step prediction of highly dynamic cylinder water entry flow fields. The proposed model integrates a Graph Neural Network (GNN) to capture complex spatial dependencies on unstructured meshes with a Long Short-Term Memory (LSTM) network to model temporal evolution. We generate a dataset from numerical simulations across different cases, varying both Froude numbers and initial entry angles. Extensive ablation studies are conducted to systematically optimize the model's architecture and hyperparameters, revealing that a single-step prediction strategy with an input sequence of eight timesteps yields the most robust performance, significantly reducing error accumulation in long-term prediction. The optimized model achieves high accuracy (R2>0.97) and successfully predicts key physical phenomena during water entry. Furthermore, generalization tests on unseen cases demonstrate the model's excellent interpolation capability within the training data range and reveal the challenges of extrapolation, particularly for the complex phase interface. Crucially, while the multi-step prediction of the velocity field degrades due to error accumulation, the inferred prediction of the water volume fraction remains remarkably stable. This indicates that the model has learned not just the instantaneous spatial correlations between the fields, but more importantly, has captured the underlying temporal dynamics that govern the evolution of the phase interface.
Assessment of cardiovascular and neurovascular disease requires precise measurement of hemodynamic parameters for effective diagnosis and prognosis. Variations in velocity gradients are correlated with changes in the pressure gradient, a key hemodynamic biomarker, necessitating a super-resolved velocity field. We introduce PI-GNN (Physics-Informed Graph Neural Networks) that combines the strengths of Graph Neural Network (GNN) and Navier-Stokes equations as physical laws of fluid dynamics to enhance the resolution of 4D flow MRI. We employed the capacity of Graph Neural Networks (GNNs) to process unstructured Computational Fluid Dynamics (CFD) data for super-resolution tasks. GNNs, inherently size-agnostic, effectively capture both local and global features. This unique suitability is further enhanced by incorporating terms from the Navier-Stokes equations as edge features within their message-passing mechanisms. We trained, evaluated, and tested PI-GNN using high-fidelity simulated Computational Fluid Dynamics (CFD) data in a phantom model of vascular stenosis. Subsequent testing of our proposed network on in vitro 4D flow MRI data exhibited commendable precision in velocity field estimation for super-resolution. Our approach has the potential to impact broader fluid dynamics research to reduce computation time with applications in cardiovascular and biological systems.
Unstructured meshes offer the flexibility to capture complex details in flow fields, and graph neural networks (GNNs) are regarded as a natural tool for processing such mesh data. However, in practical flow field simulations, the topology of computational meshes changes with variations in aircraft geometry and flight conditions. When confronted with such fundamental topological variations, the predictive performance of existing GNN methods declines significantly, severely limiting their ability to generalize to unseen mesh topologies. To address this challenge, we propose a graph convolution method based on a geometric attention mechanism, which integrates a learnable local flow direction encoding and a cosine similarity weighting strategy to actively learn geometric relational rules that are independent of specific mesh connectivity. This approach enhances feature extraction and representation in flow fields, maintains stable performance across different mesh structures, and improves the capture of high-gradient features such as shock waves. As a result, the model's robustness to mesh topological changes and its predictive accuracy are significantly improved. Numerical experiments on several typical supersonic flow cases demonstrate that, compared to existing GNN methods, the proposed approach achieves superior performance in both accuracy and generalization capability, effectively adapting to topological variations in unstructured meshes.
This study introduces a novel physics-informed spatiotemporal graph neural network (PIST-GNN) for rapid, physically consistent, and accurate long-term prediction of complex water entry flow fields. The model integrates a MeshGraphNet-based GNN for spatial feature extraction on unstructured meshes and a long short-term memory (LSTM) network for temporal evolution, critically enhanced by embedding mass conservation and volume of fluid (VOF) transport equations as soft constraints during training. Comprehensive evaluations demonstrate that the combined physics constraints provide an optimal balance, achieving robust predictions with only a slight increase in data loss while significantly reducing physical inconsistencies like velocity divergence and VOF advection residuals. The PIST-GNN exhibits superior long-term stability and generalization capabilities, consistently maintaining the lowest prediction errors and the best physical consistency over 30 autoregressive steps on unseen Froude number and initial angle cases. Furthermore, through both quantitative and qualitative analysis of prediction results, we elucidate the complex influence of initial conditions on prediction accuracy. Ultimately, this study provides a practical and feasible approach to address the persistent challenge of error accumulation in long-term predictions of water entry flow fields by purely data-driven models. Future research will focus on overcoming the limitations of the current two-dimensional slice approximation method by optimizing physical constraints and model architecture, thereby more directly capturing the complete three-dimensional dynamic characteristics of the water entry flow field.
No abstract available
No abstract available
This paper investigates a physics-informed surrogate modeling framework for multi-phase flow in porous media based on the Fourier Neural Operator. Traditional numerical simulators, though accurate, suffer from severe computational bottlenecks due to fine-grid discretizations and the iterative solution of highly nonlinear partial differential equations. By parameterizing the kernel integral directly in Fourier space, the operator provides a discretization-invariant mapping between function spaces, enabling efficient spectral convolutions. We introduce a Dual-Branch Adaptive Fourier Neural Operator with a shared Fourier encoder and two decoders: a saturation branch that uses an inverse Fourier transform followed by a multilayer perceptron and a pressure branch that uses a convolutional decoder. Temporal information is injected via Time2Vec embeddings and a causal temporal transformer, conditioning each forward pass on step index and time step to maintain consistent dynamics across horizons. Physics-informed losses couple data fidelity with residuals from mass conservation and Darcy pressure, enforcing the governing constraints in Fourier space; truncated spectral kernels promote generalization across meshes without retraining. On SPE10-style heterogeneities, the model shifts the infinity-norm error mass into the 10−2 to 10−1 band during early transients and sustains lower errors during pseudo-steady state. In zero-shot three-dimensional coarse-to-fine upscaling from 30×110×5 to 60×220×5, it attains R2=0.90, RMSE = 4.4×10−2, and MAE = 3.2×10−2, with more than 90% of voxels below five percent absolute error across five unseen layers, while the end-to-end pipeline runs about three times faster than a full-order fine-grid solve and preserves water-flood fronts and channel connectivity. Benchmarking against established baselines indicates a scalable, high-fidelity alternative for high-resolution multi-phase flow simulation in porous media.
No abstract available
Neural operators have shown strong potential for solving partial differential equations (PDEs). Among them, transformer-based models often achieve high accuracy, yet they face efficiency bottlenecks in large-scale settings because mesh discretization combined with global attention incurs quadratic computational cost. To address this limitation, we propose PARTNO, an efficient PDE solver based on physically adaptive region partitioning. Instead of operating on the full mesh globally, PARTNO adopts a physics-informed region decomposition strategy that partitions the computational domain into physically similar and relatively independent subregions, and models region-specific properties. Within each subregion, we introduce a dedicated Halo-Attention mechanism to jointly model intraregion physical representations and interactions with neighboring subregions, while maintaining continuity of physical quantities across region boundaries. By restricting attention computation to within subregions, PARTNO reduces the overall complexity to linear, improves parallelism, and delivers excellent scalability for large-scale inference. Extensive experiments show that PARTNO achieves state-of-the-art performance on multiple benchmarks; in particular, it obtains the best results on DrivAerStar, the largest industrial-scale automotive dataset. We further demonstrate its practicality in real-world industrial simulations, including automotive and aerospace design.
Blending hydrogen into natural gas for transportation is a crucial approach for achieving the widespread utilization of hydrogen. Tracking the concentration of the hydrogen within the pipeline is important for monitoring gas quality and managing pipeline operations. This study develops a rapid computational model to predict the hydrogen and natural gas concentrations within the pipeline during transportation based on the Fourier Neural Operator (FNO), an operator neural network capable of learning the differential operator in the partial differential equation. In the proposed model, the numerical method is employed to generate datasets, with the spline interpolation used to enhance data smoothness. The initial and boundary conditions are taken as the inputs to accommodate varying transportation scenarios. Comparison results indicate that the proposed model can notably reduce the time needed to predict the hydrogen and natural gas concentrations while maintaining prediction accuracy. The accuracy of the proposed model is validated by comparing its calculated results with the analytical solution and the concentrations of hydrogen and natural gas within the pipeline under two transportation scenarios, with relative errors of 0.49%, 0.31%, and 0.45%, respectively. Notably, the trained model demonstrates strong grid invariance, a type of model generalization. Trained on data generated from a coarse grid of 101 × 41 spatial-temporal resolution, the proposed model can accurately predict results on a fine grid of 401 × 81 spatial-temporal resolution with a relative error of only 0.38%. Regarding the prediction efficiency, the proposed model achieves an average 17.7-fold speedup compared to the numerical method. The positive results indicate that the proposed model can serve as a rapid and accurate solver for the composition transport equation.
This paper formulates neural computation as the finite-dimensional representation of operators acting on functions and measures. Dense layers, convolutions, attention mechanisms, graph networks, neural fields, and neural operators are treated as instances of kernel-based transformations evaluated through sampling, projection, quadrature, or basis truncation. The aim is to separate the operator being learned from the particular array representation used to compute it. This separation makes architecture design an approximation problem: one must choose a measure space, function space, kernel class, topology, and discretization scheme suited to the target operator. Numerical analysis provides stability and quadrature tools; approximation theory gives representation rates; harmonic analysis identifies bases and transforms adapted to structure; signal processing supplies efficient finite representations; measure theory treats empirical data, weak convergence, transport, and distributional functionals. The result is a unified account of neural architectures as computable approximations of operators over structured domains.
No abstract available
In this study, we introduce a Graph network‐enhanced Finite Element approach to accelerate Finite Element simulations. We utilize the discretized geometry from a Finite Element pre‐processor to establish the graph and use the Graph Neural Network to solve the boundary value problem of the discretized domain. The advantage of graph neural networks is that they have a similar structure as compared to a discretized domain with nodes and elements. The underlying dynamics of the system are computed via a learned message‐passing. The goal here is to enhance and accelerate the FEM simulations using the proposed GNN network by incorporating the underlying mechanics knowledge into the network to enhance the generalizing ability of the network on various loading and boundary conditions. All the proposed studies in the literature where graph networks are applied to Finite Element Methods use images as input and output. The advantage of the proposed model is that it takes inputs such as the nodal information, their corresponding edges, nodal coordinates and the boundary conditions for each particular node from a Finite Element pre‐processor and computes the von‐Mises stress at each node along with their edge connections as output that can be read by a Finite Element post‐processor.
Cantilevered elastic foils or panels can undergo self-induced large-amplitude flapping oscillations when submerged in a flowing fluid. This canonical phenomenon of fluid-structure interaction, commonly observed in nature, such as in fluttering leaves or fish fins, offers a promising mechanism for ocean energy harvesting. By leveraging the repetitive flapping motion, elastic foils can harness the abundant kinetic energy present in marine environments, including steady ocean currents and unsteady wave-driven flows. The efficiency of energy harvested by these foils is dictated by the amplitude and frequency of the oscillation, which are influenced by interactions with the surrounding flow. Optimizing these parameters is crucial for maximizing energy conversion efficiency and designing robust systems. However, high-fidelity simulations required for optimization over a wide parameter space are computationally expensive. To address this challenge, we propose a graph neural network-based surrogate model (GNN-ROM) tailored for the inverted foil problem. This model effectively handles mesh-structured data from full-order simulations without modifications, enabling efficient and accurate fluid-structure interaction simulations. The inverted foil is modeled as an elastically mounted rigid foil that undergoes a pitching motion around its trailing edge in uniform flow. The coupled fluid-structure system is simulated using a high-fidelity Petrov-Galerkin finite element approach, leveraging the arbitrary Lagrangian-Eulerian (ALE) formulation for precise interface tracking. The high-fidelity simulations serve as ground truth data for the training and evaluation of neural networks. Our deep learning model utilizes a rotation equivariant, quasi-monolithic GNN architecture based on the ALE formulation, wherein the coupled system dynamics are predicted with two sub-networks. First, essential coefficients describing mesh motion are extracted by proper orthogonal decomposition, which are predicted over time using a single multilayer perceptron. Moreover, the GNN-ROM evolves the flow field according to the state of the system. This approach effectively simulates coupled FSI dynamics, offering a robust surrogate model for rapid parametric optimization and control of energy harvesting devices based on inverted foil configurations.
No abstract available
A new graph neural network (GNN) subgrid-scale (SGS) stress modeling framework based on multi-scale self-similarity input features is investigated as an improvement in generalizability over a previous GNN model developed by the authors' group. Training and a priori testing are conducted with an incompressible direct numerical simulation (DNS) database comprising homogeneous isotropic turbulence (HIT), turbulent mixing layer, and turbulent channel flow (CH) cases, each across five Reynolds numbers. DNS fields are filtered at two different scales using both Gaussian and box kernels to create a diverse large eddy simulation training dataset. The multi-scale Leonard stress tensor and spatial velocity fluctuations are used as inputs to capture scale interactions from a broader portion of the turbulent energy spectrum. Galilean invariance and scale equivariance are ensured by input definition and nondimensionalization, respectively. Randomized data augmentation is used to improve model robustness to changes in orientation. Spline-based graph-convolution layers enable non-local feature extraction while maintaining model compatibility with unstructured grids. A priori analysis of four GNN SGS models shows that graph convolution and multi-scale input features provide significant improvement in accuracy for incompressible flows at Reynolds numbers below, within and above the training range, with particular impact in the near-wall region. Model generalizability to compressible reactive flows is high except in a wall-bounded case. In a posteriori testing with incompressible turbulent CH, one multi-scale GNN was found to better match reference DNS profiles than conventional eddy-viscosity models, with computational overhead comparable to that of a dynamic k-equation model.
A neural-network-based large eddy simulation (LES) is conducted for turbulent flow over a backward-facing step. A fully connected neural network (NN) is trained using filtered direct numerical simulation (DNS) data at the Reynolds number of Reh=5100 based on the free-stream velocity and step height, to develop an NN-based subgrid-scale (SGS) model. Three sets of input variables are considered for training. NN1 is based solely on the filtered strain-rate tensor. NN2 employs the same tensor scaled by the eddy viscosity introduced in the Vreman model, which accounts for the local velocity gradients and grid anisotropy. NN3 further augments the NN2 input with the Leonard stress from the scale-similarity approach. The output of NNs is the SGS stress tensor. For a priori tests, all the NNs show better predictions of the SGS stress and SGS dissipation than the dynamic Smagorinsky model. A posteriori tests (i.e., LES) conducted at the trained Reynolds number and grid distribution show accurate predictions of the reattachment length, mean velocity, and root-mean-square velocity fluctuations. Furthermore, NN2 and NN3 demonstrate good performances on unseen coarser grid resolutions, but NN1 does not. Especially, NN2 maintains predictive capability even at a higher Reynolds number (Reh=28000), whereas NN3 requires ad hoc clipping on negative SGS dissipation for a stable LES at this Reynolds number.
Abstract A neural-network-based large eddy simulation is performed for flow over a circular cylinder. To predict the subgrid-scale (SGS) stresses, we train two fully connected neural network (FCNN) architectures with and without fusing information from two separate single-frame networks (FU and nFU, respectively), where the input variable is either the strain rate (SR) or the velocity gradient (VG). As the input variables, only the grid-filtered variables are considered for the SGS models of G-SR and G-VG, and both the grid- and test-filtered variables are considered for the SGS models of T-SR and T-VG. The training data are the filtered direct numerical simulation (fDNS) data at $Re_d=3900$ based on the free-stream velocity and cylinder diameter. Using the same grid resolution as that of the training data, the performances of G-SR and G-VG (grid-filtered inputs) and T-SR-FU and T-VG-FU (grid- and test-filtered inputs with fusion) are better than those of the dynamic Smagorinsky model and T-SR-nFU and T-VG-nFU (grid- and test-filtered inputs without fusion). These FCNN-based SGS models are applied to untrained flows having different grid resolutions from that of training data. Although the performances of G-SR and G-VG are degraded, T-SR-FU and T-VG-FU still provide good performances. Finally, T-SR-FU and T-VG-FU trained at $Re_d = 3900$ are applied to higher-Reynolds-number flows ($Re_d = 5000$ and 10 000) and their results are also in good agreements with those of fDNS and previous experiment, indicating that adding the test-filtered variables and fusion increases the prediction capability even for untrained Reynolds number flows.
No abstract available
This study is concerned with accurately predicting the subgrid-scale (SGS) stress using an artificial neural network (ANN) with a linear eddy-viscosity term and a nonlinear term as the input variables. A priori and a posteriori tests are conducted to examine the prediction performance of the ANN-based SGS stress model in decaying homogeneous isotropic turbulence. In a priori test, the present ANN-based SGS model shows high correlation coefficients between the true and predicted SGS stresses, and excellent predictions of the SGS stress and dissipation. In a posteriori test, it is found that the ANN-based SGS model can predict the turbulence statistics more accurately than the traditional dynamic SGS models. The generalization capabilities of the model to untrained flow conditions and unstrained types of turbulent flow have been evaluated. It is found that the proposed ANN-based model can provide an accurate prediction of the SGS stress under different Reynolds numbers and flow types. A comparison among several existing ANN-based models with different input variables is presented, demonstrating a significant advantage of the present model.
Subgrid-scale (SGS) stress modeling based on filtered variables is one of the crucial scientific challenges in large-eddy simulation. With the rapid development of machine learning technologies in recent years, data-driven turbulence modeling methods have gained its popularity. In this study, an SGS stress model based on artificial neural network (ANN), with strain-rate tensor and modified Leonard tensor as inputs, is developed for incompressible isotropic homogeneous turbulence. The proposed ANN model demonstrates a substantial enhancement in the prediction of the SGS stress. Also, the ANN model could provide better predictions of turbulence statistics, as compared to the traditional models. It is suggested that the ANN methods exhibit obvious advantages and considerable potentials for the development of turbulence models with high accuracy.
A neural-network (NN)-based subgrid-scale (SGS) model, trained only with data from flow over a backward-facing step (BFS), is assessed in large-eddy simulation (LES) of three untrained flows: turbulent channel flow, flow over a circular cylinder, and flow over an Ahmed body. The NN is trained to predict the SGS stress tensor from the filtered strain-rate tensor scaled by a Vreman-type eddy viscosity. To account for different orientations of boundary layers and separated regions in the test flows, the training dataset is augmented through streamwise-axis rotations. During training, the input and output variables are normalized by the free-stream velocity of the BFS flow, whereas spatially local normalization based on the true SGS stress magnitude does not provide improved predictions in the present LES. For the channel flow at Reτ≈393 and 710, the NN predicts the mean velocity profiles accurately, with rms velocity fluctuations similar to those from the dynamic Smagorinsky model (DSM) and the dynamic global model (DGM). For the circular-cylinder flow at Red=3900, the NN provides improved predictions of the recirculation length and streamwise velocity fluctuations compared with DSM and DGM. Furthermore, for the Ahmed body flow with a slant angle of 25° at ReH = 96 000, the NN reproduces the separation bubble on the slanted surface, with predictions similar to those from DGM for the drag coefficient and instantaneous vortical structures. These results suggest that BFS training data, combined with rotation-based augmentation, provide useful SGS information for developing an NN-based SGS model applicable to wall-bounded and separated flows.
Small-scale motions in turbulent flows play a significant role in various small-scale processes, such as particle relative dispersion and collision, bubble or droplet deformation, and orientation dynamics of non-sphere particles. Recovering the small-scale flows that cannot be resolved in large eddy simulation (LES) is of great importance for such processes sensitive to the small-scale motions in turbulent flows. This study proposes a subgrid-scale model for recovering the small-scale turbulent velocity field based on the artificial neural network (ANN). The governing equations of small-scale turbulent velocity are linearized, and the pressure gradient and the nonlinear convection term are modeled with the aid of the ANN. Direct numerical simulation (DNS) and filtered direct numerical simulation (FDNS) provide the data required for training and validating the ANN. The large-scale velocity and velocity gradient tensor are selected as inputs for the ANN model. The linearized governing equations of small-scale turbulent velocity are numerically solved by coupling the large-scale flow field information. The results indicate that the model established by the ANN can accurately recover the small-scale velocity lost in FDNS due to filtering operation. With the ANN model, the flow fields at different Reynolds numbers agree well with the DNS results regarding velocity field statistics, flow field structures, turbulent energy spectra, and two-point, two-time Lagrangian correlation functions. This study demonstrates that the proposed ANN model can be applied to recovering the small-scale velocity field in the LES of isotropic turbulent flows at different Reynolds numbers.
The extrapolation performance of Convolutional Neural Network (CNN)-based models for Large-Eddy Simulations (LES) has been investigated in the context of turbulent premixed combustion. The study utilises a series of Direct Numerical Simulation (DNS) datasets of turbulent premixed methane/air and hydrogen/air jet flames to train the CNN models. The methane/air flames, which are characterised by increasing Reynolds numbers, are used to model the subgrid-scale flame wrinkling. The hydrogen/air flame, exhibiting complex thermodiffusive instability, is employed to test the ability of the CNN-based combustion models to predict the filtered progress variable source term. This study focuses on the influence of varying training Reynolds numbers, filter sizes, and filter kernels to evaluate the performance of the CNN models to out-of-sample conditions, i.e., not seen during training. The objectives of this study are threefold: (i) analyse the performance of CNN models at different Reynolds numbers compared to the one trained with; (ii) analyse the performance of CNN models at different filter sizes compared to the one trained with; (iii) assess the influence of using different filter kernels (i.e., Gaussian and box filter kernels) between training and testing, to emulate a posteriori applications. The results demonstrate that the CNN models show good extrapolation performance when the training Reynolds number is sufficiently high. Vice versa, when CNN models are trained on low-Reynolds-number flame data, their performance degrades as they are applied to flames with progressively higher Reynolds numbers. When these CNN models are tested on datasets with filter sizes not included in the training process, they exhibit sufficient interpolation capabilities, the extrapolation performance is less precise but still satisfactory overall. This indicates that CNN models can be effectively trained using data filtered with a limited range of filter sizes and then successfully applied across a broader spectrum of filter sizes. Furthermore, when CNNs trained on box-filtered data are applied to Gaussian-filtered data, or vice versa, the models perform well for smaller filter sizes. However, as the filter size increases, the accuracy of the predictions diminishes. Interestingly, increasing the quantity of training data does not significantly enhance model performance. Yet, when training data are distributed with greater weighting towards larger filter sizes, the model’s overall performance improves. This suggests that the strategic selection and weighting of training data can lead to more robust generalization across different filter conditions.
Abstract An artificial neural network (ANN) is used to establish the relation between the resolved-scale flow field and the subgrid-scale (SGS) stress tensor, to develop a new SGS model for large-eddy simulation (LES) of isotropic turbulent flows. The data required for training and testing of the ANN are provided by performing filtering operations on the flow fields from direct numerical simulations (DNSs) of isotropic turbulent flows. We use the velocity gradient tensor together with filter width as input features and the SGS stress tensor as the output labels for training the ANN. In the a priori test of the trained ANN model, the SGS stress tensors obtained from the ANN model and the DNS data are compared by computing the correlation coefficient and the relative error of the energy transfer rate. The correlation coefficients are mostly larger than 0.9, and the ANN model can accurately predict the energy transfer rate at different Reynolds numbers and filter widths, showing significant improvement over the conventional models, for example the gradient model, the Smagorinsky model and its dynamic version. A real LES using the trained ANN model is performed as the a posteriori validation. The energy spectrum computed by the improved ANN model is compared with several SGS models. The Lagrangian statistics of fluid particle pairs obtained from the improved ANN model almost approach those from the filtered DNS, better than the results from the Smagorinsky model and dynamic Smagorinsky model.
No abstract available
Graph and convolutional neural network coupling with a high-performance large-eddy simulation solver
No abstract available
Abstract This study proposes a machine-learning-based subgrid scale (SGS) model for very coarse-grid large-eddy simulations (vLES). An issue with SGS modelling for vLES is that, because the energy-containing eddies are not accurately resolved by the computational grid, the resolved turbulence deviates from the physically accurate turbulence. This limits the use of supervised machine-learning models commonly trained using pairs of direct numerical simulation (DNS) and filtered DNS data. The proposed methodology utilises both unsupervised learning (cycle-consistency generative adversarial network (GAN)) and supervised learning (conditional GAN) to construct a machine-learning pipeline. The unsupervised learning part of the proposed method first transforms the non-physical vLES flow field to resemble a physically accurate flow field. The second supervised learning part employs super-resolution of turbulence to predict the SGS stresses. The proposed pipeline is trained using a fully developed turbulent channel at the friction Reynolds number of approximately 1000. The a priori validation shows that the proposed unsupervised–supervised pipeline successfully learns to predict the accurate SGS stresses, while a typical supervised-only model shows significant discrepancies. In the a posteriori test, the proposed unsupervised–supervised-pipeline SGS model for vLES using a progressively coarse grid yields good agreement of the mean velocity and Reynolds shear stress with the reference data at both the trained Reynolds number 1000 and the untrained higher Reynolds number 2000, showing robustness against varying Reynolds numbers. A budget analysis of the Reynolds stresses reveals that the proposed unsupervised–supervised-pipeline SGS model predicts a significant amount of SGS backscatter, which results in the strengthened near-wall Reynolds shear stress and the accurate prediction of mean velocity.
No abstract available
Abstract In this paper, we propose artificial-neural-network-based (ANN-based) nonlinear algebraic models for the large-eddy simulation (LES) of compressible wall-bounded turbulence. An innovative modification is applied to the invariants and the tensor bases of the nonlinear algebraic models through using the local grid widths along each direction to normalise the corresponding gradients of the flow variables. Furthermore, the dimensionless model coefficients are determined by the ANN method. The modified ANN-based nonlinear algebraic model (MANA model) has much higher correlation coefficients and much lower relative errors than the dynamic Smagorinsky model (DSM), Vreman model and wall-adapting local eddy-viscosity model in the a priori test. The significantly more accurate estimations of the mean subgrid-scale (SGS) fluxes of the kinetic energy and temperature variance are also obtained by the MANA models in the a priori test. Furthermore, in the a posteriori test, the MANA model can give much more accurate predictions of the flow statistics and the mean SGS fluxes of the kinetic energy and the temperature variance than other traditional eddy-viscosity models in compressible turbulent channel flows with untrained Reynolds numbers, Mach numbers and grid resolutions. The MANA model has a better performance in predicting the flow statistics in supersonic turbulent boundary layer. The MANA model can well predict both direct and inverse transfer of the kinetic energy and temperature variance, which overcomes the inherent shortcoming that the traditional eddy-viscosity models cannot predict the inverse energy transfer. Moreover, the MANA model is computationally more efficient than the DSM.
Session H Abstracts Numerically Consistent Data-Driven Subgrid-Scale Model for Large-Eddy Simulation
No abstract available
Abstract A fully connected neural network (NN) is used to develop a subgrid-scale (SGS) model mapping the relation between the SGS stresses and filtered flow variables in a turbulent channel flow at $Re_\tau = 178$. A priori and a posteriori tests are performed to investigate its prediction performance. In a priori test, an NN-based SGS model with the input filtered strain rate or velocity gradient tensor at multiple points provides highest correlation coefficients between the predicted and true SGS stresses, and reasonably predicts the backscatter. However, this model provides unstable solution in a posteriori test, unless a special treatment such as backscatter clipping is used. On the other hand, an NN-based SGS model with the input filtered strain rate tensor at single point shows an excellent prediction capability for the mean velocity and Reynolds shear stress in a posteriori test, although it gives low correlation coefficients between the true and predicted SGS stresses in a priori test. This NN-based SGS model trained at $Re_\tau = 178$ is applied to a turbulent channel flow at $Re_\tau = 723$ using the same grid resolution in wall units, providing fairly good agreements of the solutions with the filtered direct numerical simulation (DNS) data. When the grid resolution in wall units is different from that of trained data, this NN-based SGS model does not perform well. This is overcome by training an NN with the datasets having two filters whose sizes are bigger and smaller than the grid size in large eddy simulation (LES). Finally, the limitations of NN-based LES to complex flow are discussed.
Graph neural networks are well suited for physics based simulation. Among other features, graphs can accurately represent thermal effects, with energy conservation operating on the nodes (vertices) and heat flow coursing through edges. Moreover, graph neural networks incorporate the data topology and allow simulation on a wide variety of geometries. In this paper, various graph neural networks are trained using analytical solutions to the transient heat equation. Several architectures and hyper-parameters are compared to empirically identify good practices when using a graph neural network for numerical simulation. In addition, a novel method to optimize the network architecture is proposed. It is based on rating systems used to evaluate players’ skill level in competitive games. The results show that graph neural network-based simulation shares mutual restrictions with classical discretizations techniques (such as the Von Neumann stability criterion) and also that graph neural networks can be a potent generalization tool for heat transfer problems.
No abstract available
Conventional computational fluid dynamics methods are computationally intensive for tasks such as design optimization and real-time control, and while convolutional neural network-based surrogates potentially offer faster alternatives, their reliance on structured grids, static neighborhood definitions, and limited extrapolation capabilities highlights the need for more flexible and generalizable frameworks. Graph convolutional networks (GCNs), with their ability to operate on unstructured meshes and dynamically encode spatial relationships, offer a compelling alternative. In this work, a novel residual multilayer perceptron-based GCN (ResMLP-GCN) architecture is proposed, and its utility is demonstrated by applying it to three benchmark problems: (i) transient non-linear diffusion–advection problem, (ii) double lid-driven cavity flow, and (iii) backward-facing step flow. The developed models are trained on data from numerical simulations, using nodal coordinates and physical field variables as node features, employing a finite element-inspired hypergraph ϕ topology. The heat transfer model, trained on data from t=0 to 1000 s, accurately predicts the spatiotemporal evolution of temperature fields up to t=2000 s, maintaining relative errors below 8%. In the double lid-driven cavity problem, the model trained on Reynolds numbers ranging from 100 to 2000 achieves mean percentage errors under 14% when evaluated on extrapolated Reynolds numbers up to 3000. Similarly, for the backward-facing step, the model predicts the reattachment lengths for extrapolated Re with relative errors below 3%. These results highlight the potential of ResMLP-GCNs to capture complex spatiotemporal dynamics and generalize across temporal and parametric domains, offering a robust surrogate modeling strategy for heat and fluid flow problems.
The numerical simulation of Homann flow, a classical problem in fluid dynamics involving axisymmetric flow over a solid plate, is essential for understanding various boundary layer phenomena. Due to the prominence of this flow and its computationally intensive nature when dealing with turbulent regimes in complex geometries, the present paper explores the use of machine learning (ML) to solve Homann flow with heat transfer both rapidly and accurately. However, the application of ML, especially deep learning (DL), to engineering-focused physics simulations remains a significant challenge. Traditional numerical approaches have consistently demonstrated higher accuracy in simulation tasks compared to DL models, which often suffer from a lack of generalizability when predicting outcomes for geometries that differ slightly from those used in training. In this study, Graph Neural Networks (GNNs) are employed to solve thermal Homann flow on substrates of various geometries. We demonstrate that the trained model accurately predicts outcomes even on entirely novel cases, and notably, the computational efficiency of the GNN model significantly surpasses that of conventional numerical solvers.
Pin-fins are widely used to enhance heat transfer in compact heat exchangers, turbine cooling passages, and electronic devices, but their complex geometries make accurate thermal–fluid prediction computationally expensive. This paper presents a geometry-aware multiscale (GAMS) graph neural network (GNN) for predicting steady turbulent flow and heat transfer in a two-dimensional channel containing arbitrarily shaped pin-fin geometries. An automated framework integrating geometry generation, meshing, and ANSYS Fluent simulations was developed to construct the training dataset. Pin-fin geometries were parameterized using piecewise cubic splines, generating 1000 unique configurations through Latin Hypercube Sampling. Each simulation was converted into a graph representation, where nodes contained spatial coordinates, normalized streamwise position, one-hot boundary indicators, and signed distance to the nearest wall. These graph-based features were used to train the GNN to predict the temperature, velocity magnitude, and pressure fields directly from geometry. The network achieved excellent predictive accuracy, successfully capturing boundary layers, recirculation zones, and upstream stagnation regions while reducing computational wall time by 2–3 orders of magnitude compared to conventional CFD simulations. Overall, the proposed GNN provides a fast, reliable surrogate modeling framework for complex thermal–fluid flow configurations.
No abstract available
A novel Spatiotemporal Sequence Graph Convolutional Network (ST-SGCN) data-driven model is proposed to predict transient fluid dynamics around airfoils using complex and unstructured flow field data, with the aim of reducing dimensions and expediting predictions. Graph Neural Networks directly interact with the flow field grid, capturing spatiotemporal physical features of grid nodes and their interconnections, while eliminating the need for complex preprocessing steps. The ST-SGCN model integrates a Graph Convolutional Network and a Graph Attention Network with a Deep Recurrent Neural Network that uses a Gate Recurrent Unit as the kernel, adeptly extracting spatial and temporal physical features of the flow field to accurately predict transient flow states. Preliminary airfoil flow experiments demonstrated the model's ability to continuously predict transient flow fields, achieving an average accuracy of 97% for both velocity and pressure field predictions, with a maximum error of approximately 10% in the testing dataset. Further experiments, varying angles of attack, airfoils, and Reynolds numbers, demonstrated the model's generalizability, extensibility, and adaptability, with prediction errors below 5% and a speedup of over 20 times.
Despite advancements in high-performance computing and numerical algorithms, Computational Fluid Dynamics (CFD) remains challenging for practical real-time applications, particularly in analysis and design tasks such as digital twin implementations. While traditional Reduced-Order Methods offer efficient and accurate predictions of entire flow fields, autoencoder Convolutional Neural Networks (CNNs) have demonstrated success in reconstructing CFD solutions due to their exceptional local feature extraction capabilities and scalability. However, their applicability is constrained to simple geometries because of the reliance on Cartesian or pixel-like grid structures. In this study, we propose a novel Point-based U-Net (PointUNet) framework incorporating Local Point Encoding (LPE) as a mesh-independent autoencoder model. The key functionality of LPE lies in its ability to transform point cloud data into a standard input array for conventional CNNs using a Virtual Reference Grid. This approach avoids data loss typically associated with interpolation or extrapolation, enabling greater flexibility in mesh generation and complex geometry handling. Verification was conducted using airfoil flows at transonic speeds and cylinder flows at low Reynolds numbers with various cross-sectional shapes, achieving minimal verification errors. The results were compared directly with other point cloud methods, demonstrating superior accuracy and efficiency in predicting highly nonlinear flows involving separation and shock waves, showing better agreement with full-order CFD solutions.
This paper introduces a novel surrogate model for two-dimensional adaptive steady-state thermal convection fields based on deep learning technology. The proposed model aims to overcome limitations in traditional frameworks caused by network types, such as the requirement for extensive training data, accuracy loss due to pixelated preprocessing of original data, and inability to predict information near the boundaries with precision. We propose a new framework that consists primarily of a physical-informed neural network (PINN) and a graph convolutional neural network (GCN). The GCN serves as the prediction module and predicts thermal convection in the two-dimensional computational domain by considering the mutual influence between unstructured nodes and their neighbors. On the other hand, the PINN acts as the physical constraint module of the framework by embedding the control equation of thermal convection into the loss function of the neural network, ensuring that the inference and prediction results of the GCN comply with the constraints of the control equation. The advantages of this framework lie in two aspects. First, the computation mechanism of the GCN is more in line with the actual evolution of temperature fields. Second, the PINN enhances the cognitive ability of the surrogate model toward the convection field information. It accurately describes the changes of temperature gradient information at the boundary position and reduces the model's demand for training data. To validate the advantages of the proposed model, we gradually analyzed the model's geometric adaptability and predictive accuracy from the single cylinder case to the double cylinder case. We also investigated the impact of the number of sampling points on model training and compared the model's prediction results with those of a purely data-driven model. The results show that the proposed model exhibits good geometric adaptability and stability. With only 20 training data, the mean error of the proposed model in predicting the velocity and temperature field is less than 1% and 0.6% for the single cylinder, and less than 2% and 1% for the double cylinder case, while the mean error of the purely data-driven GCN model in predicting the velocity and temperature field is 9.4% and 6.4% for the double cylinder case. These findings demonstrate the effectiveness of the proposed physics-informed graph convolutional neural network, allowing for more accurate prediction of fluid flow and heat convection using surrogate model.
Crash simulations play an essential role in improving vehicle safety, design optimization, and injury risk estimation. Unfortunately, numerical solutions of such problems using state-of-the-art high-fidelity models require significant computational effort. Conventional data-driven surrogate modeling approaches create low-dimensional embeddings for evolving the dynamics in order to circumvent this computational effort. Most approaches directly operate on high-resolution data obtained from numerical discretization, which is both costly and complicated for mapping the flow of information over large spatial distances. Furthermore, working with a fixed resolution prevents the adaptation of surrogate models to environments with variable computing capacities, different visualization resolutions, and different accuracy requirements. We thus propose a multi-hierarchical framework for structurally creating a series of surrogate models for a kart frame, which is a good proxy for industrial-relevant crash simulations, at different levels of resolution. For multiscale phenomena, macroscale features are captured on a coarse surrogate, whereas microscale effects are resolved by finer ones. The learned behavior of the individual surrogates is passed from coarse to finer levels through transfer learning. In detail, we perform a mesh simplification on the kart model to obtain multi-resolution representations of it. We then train a graph-convolutional neural network-based surrogate that learns parameter-dependent low-dimensional latent dynamics on the coarsest representation. Subsequently, another, similarly structured surrogate is trained on the residual of the first surrogate using a finer resolution. This step can be repeated multiple times. By doing so, we construct multiple surrogates for the same system with varying hardware requirements and increasing accuracy.
Large-eddy simulations (LES) require closures for filtered production rates because the resolved fields do not contain all correlations that govern chemical source terms. We develop a graph neural network (GNN) that predicts filtered species production rates on non-uniform meshes from inputs of filtered mass fractions and temperature. Direct numerical simulations of turbulent premixed hydrogen-methane jet flames with hydrogen fractions of 10%, 50%, and 80% provide the dataset. All fields are Favre filtered with the filter width matched to the operating mesh, and learning is performed on subdomain graphs constructed from mesh-point connectivity. A compact set of reactants, intermediates, and products is used, and their filtered production rates form the targets. The model is trained on 10% and 80% blends and evaluated on the unseen 50% blend to test cross-composition generalization. The GNN is compared against an unclosed reference that evaluates rates at the filtered state, and a convolutional neural network baseline that requires remeshing. Across in-distribution and out-of-distribution cases, the GNN yields lower errors and closer statistical agreement with the reference data. Furthermore, the model demonstrates robust generalization across varying filter widths without retraining, maintaining bounded errors at coarser spatial resolutions. A backward facing step configuration further confirms prediction efficacy on a practically relevant geometry. These results highlight the capability of GNNs as robust data-driven closure models for LES on complex meshes.
Urban wind flow reconstruction is essential for assessing air quality, heat dispersion, and pedestrian comfort, yet remains challenging when only sparse sensor data are available. We propose GenDA, a generative data assimilation framework that reconstructs high-resolution wind fields on unstructured meshes from limited observations. The model employs a multiscale graph-based diffusion architecture trained on computational fluid dynamics (CFD) simulations and interprets classifier-free guidance as a learned posterior reconstruction mechanism: the unconditional branch learns a geometry-aware flow prior, while the sensor-conditioned branch injects observational constraints during sampling. This formulation enables obstacle-aware reconstruction and generalization to held-out mesh geometries, wind directions, and sensor configurations within the studied urban-flow setting, without retraining. We consider both sparse fixed sensors and trajectory-based observations using the same reconstruction procedure. When evaluated against supervised graph neural network (GNN) baselines and classical reduced-order data assimilation methods, GenDA reduces the relative root-mean-square error (RRMSE) by 25-57% and increases the structural similarity index (SSIM) by 23-33% across the tested meshes. Experiments are conducted on Reynolds-averaged Navier-Stokes (RANS) simulations of a real urban neighborhood in Bristol, United Kingdom, at a characteristic Reynolds number of $\mathrm{Re}\approx2\times10^{7}$, featuring complex building geometry and irregular terrain. The proposed framework provides a scalable path toward generative, geometry-aware data assimilation for environmental monitoring in complex domains.
Accurate yet efficient surrogate models are essential for large-scale simulations of partial differential equations (PDEs), particularly for uncertainty quantification (UQ) tasks that demand hundreds or thousands of evaluations. We develop a physics-inspired graph neural network (GNN) surrogate that operates directly on unstructured meshes and leverages the flexibility of graph attention. To improve both training efficiency and generalization properties of the model, we introduce a domain decomposition (DD) strategy that partitions the mesh into subdomains, trains local GNN surrogates in parallel, and aggregates their predictions. We then employ transfer learning to fine-tune models across subdomains, accelerating training and improving accuracy in data-limited settings. Applied to ice sheet simulations, our approach accurately predicts full-field velocities on high-resolution meshes, substantially reduces training time relative to training a single global surrogate model, and provides a ripe foundation for UQ objectives. Our results demonstrate that graph-based DD, combined with transfer learning, provides a scalable and reliable pathway for training GNN surrogates on massive PDE-governed systems, with broad potential for application beyond ice sheet dynamics.
In recent years, Graph Neural Network (GNN) based models have shown promising results in simulating physics of complex systems. However, training dedicated graph network based physics simulators can be costly, as most models are confined to fully supervised training, which requires extensive data generated from traditional physics simulators. To date, how transfer learning could improve the model performance and training efficiency has remained unexplored. In this work, we introduce a pre-training and transfer learning paradigm for graph network simulators. We propose the scalable graph U-net (SGUNET). Incorporating an innovative depth-first search (DFS) pooling, the SGUNET is adaptable to different mesh sizes and resolutions for various simulation tasks. To enable the transfer learning between differently configured SGUNETs, we propose a set of mapping functions to align the parameters between the pre-trained model and the target model. An extra normalization term is also added into the loss to constrain the difference between the pre-trained weights and target model weights for better generalization performance. To pre-train our physics simulator we created a dataset which includes 20,000 physical simulations of randomly selected 3D shapes from the open source A Big CAD (ABC) dataset. We show that our proposed transfer learning methods allow the model to perform even better when fine-tuned with small amounts of training data than when it is trained from scratch with full extensive dataset. On the 2D Deformable Plate benchmark dataset, our pre-trained model fine-tuned on 1/16 of the training data achieved an 11.05\% improvement in position RMSE compared to the model trained from scratch.
Physics-based deep learning frameworks have shown to be effective in accurately modeling the dynamics of complex physical systems with generalization capability across problem inputs. Data-driven networks like GNN, Neural Operators have proved to be very effective in generalizing the model across unseen domain and resolutions. But one of the most critical issues in these data-based models is the computational cost of generating training datasets. Complex phenomena can only be captured accurately using deep networks with large training datasets. Furthermore, numerical error of training samples is propagated in the model errors, thus requiring the need for accurate data, i.e. FEM solutions on high-resolution meshes. Multi-fidelity methods offer a potential solution to reduce the training data requirements. To this end, we propose a novel GNN architecture, Multi-Fidelity U-Net, that utilizes the advantages of the multi-fidelity methods for enhancing the performance of the GNN model. The proposed architecture utilizes the capability of GNNs to manage complex geometries across different fidelity levels, while enabling flow of information between these levels for improved prediction accuracy for high-fidelity graphs. We show that the proposed approach performs significantly better in accuracy and data requirement and only requires training of a single network compared to other benchmark multi-fidelity approaches like transfer learning. We also present Multi-Fidelity U-Net Lite, a faster version of the proposed architecture, with 35% faster training, with 2 to 5% reduction in accuracy. We carry out extensive validation to show that the proposed models surpass traditional single-fidelity GNN models in their performance, thus providing feasible alternative for addressing computational and accuracy requirements where traditional high-fidelity simulations can be time-consuming.
Despite the plethora of deep learning frameworks for flow field prediction, most of them deal with flow fields on regular domains, and although the best ones can cope with irregular domains, they mostly rely on graph networks, so that real industrial applications remain currently elusive. We present a deep learning framework for 3D flow field prediction applied to blades of aircraft engine turbines and compressors. Crucially, we view any 3D field as a function from coordinates that is modeled by a neural network we call the backbone-net. It inherits the property of coordinate-based MLPs, namely the discretization-agnostic representation of flow fields in domains of arbitrary topology at infinite resolution. First, we demonstrate the performance of the backbone-net solo in regressing 3D steady simulations of single blade rows in various flow regimes: it can accurately render important flow characteristics such as boundary layers, wakes and shock waves. Second, we introduce a hyper-net that maps the surface mesh of a blade to the parameters of the backbone-net. By doing so, the flow solution can be directly predicted from the blade geometry, irrespective of its parameterization. Together, backbone-net and hyper-net form a highly-accurate memory-efficient data-driven proxy to CFD solvers with good generalization on unseen geometries.
Physics-based simulation of mesh based domains remains a challenging task. State-of-the-art techniques can produce realistic results but require expert knowledge. A major bottleneck in many approaches is the step of integrating a potential energy in order to compute velocities or displacements. Recently, learning based method for physics-based simulation have sparked interest with graph based approaches being a promising research direction. One of the challenges for these methods is to generate models that are mesh independent and generalize to different material properties. Moreover, the model should also be able to react to unforeseen external forces like ubiquitous collisions. Our contribution is based on a simple observation: evaluating forces is computationally relatively cheap for traditional simulation methods and can be computed in parallel in contrast to their integration. If we learn how a system reacts to forces in general, irrespective of their origin, we can learn an integrator that can predict state changes due to the total forces with high generalization power. We effectively factor out the physical model behind resulting forces by relying on an opaque force module. We demonstrate that this idea leads to a learnable module that can be trained on basic internal forces of small mesh patches and generalizes to different mesh typologies, resolutions, material parameters and unseen forces like collisions at inference time. Our proposed paradigm is general and can be used to model a variety of physical phenomena. We focus our exposition on the detail enhancement of coarse clothing geometry which has many applications including computer games, virtual reality and virtual try-on.
Mesh-free numerical methods provide flexible discretisations for complex geometries; however, classical meshless discrete differential operators typically trade low computational cost for limited accuracy or high accuracy for substantial per-stencil computation. We introduce a parametrised framework for learning mesh-free discrete differential operators using a graph neural network trained via polynomial moment constraints derived from truncated Taylor expansions. The model maps local stencils relative positions directly to discrete operator weights. The current work demonstrates that neural networks can learn classical polynomial consistency while retaining robustness to irregular neighbourhood geometry. The learned operators depend only on local geometry, are resolution-agnostic, and can be reused across particle configurations and governing equations. We evaluate the framework using standard numerical analysis diagnostics, showing improved accuracy over Smoothed Particle Hydrodynamics, and a favourable accuracy-cost trade-off relative to a representative high-order consistent mesh-free method in the moderate-accuracy regime. Applicability is demonstrated by solving the weakly compressible Navier-Stokes equations using the learned operators.
We present a novel graph-informed transformer operator (GITO) architecture for learning complex partial differential equation systems defined on irregular geometries and non-uniform meshes. GITO consists of two main modules: a hybrid graph transformer (HGT) and a transformer neural operator (TNO). HGT leverages a graph neural network (GNN) to encode local spatial relationships and a transformer to capture long-range dependencies. A self-attention fusion layer integrates the outputs of the GNN and transformer to enable more expressive feature learning on graph-structured data. TNO module employs linear-complexity cross-attention and self-attention layers to map encoded input functions to predictions at arbitrary query locations, ensuring discretization invariance and enabling zero-shot super-resolution across any mesh. Empirical results on benchmark PDE tasks demonstrate that GITO outperforms existing transformer-based neural operators, paving the way for efficient, mesh-agnostic surrogate solvers in engineering applications.
Reliable global ocean forecasting is critical for climate monitoring, marine navigation, and extreme event early warning. Physics-based ocean forecasting models impose prohibitive computational costs, while existing deep learning approaches predominantly rely on structured-grid architectures, incurring unnecessary computation on masked land cells and enforcing uniform resolution across dynamically heterogeneous ocean regions regardless of local flow complexity. Here we present OceanLight, an efficient global ocean forecasting framework innovatively combining geometry-adaptive unstructured mesh tokenization with a graph neural network (GNN) backbone. OceanLight achieves pointwise forecast accuracy and kinetic energy spectral fidelity exceeding both operational numerical analyses and state-of-the-art AI-based models, while surpassing all AI-based ocean models in geostrophic balance consistency. Furthermore, OceanLight demonstrates reliable mesoscale eddy representation, capturing coherent ocean structures beyond pointwise statistical optimization. These capabilities are delivered with a 62% reduction in GPU memory consumption and 70\% reduction in FLOPs relative to structured-grid baselines. Our unstructured mesh representation establishes a generalizable paradigm for scalable data-driven oceanography.
Simulating complex fluid flows requires capturing full equilibrium distributions rather than just mean trajectories, yet high-fidelity solvers remain computationally prohibitive. Recent advances, such as Diffusion Graph Networks (DGNs), have combined diffusion models with graph neural networks to sample equilibrium states directly from unstructured meshes, enabling distributional accuracy even from short simulations. However, graph-based diffusion approaches suffer from hand-crafted architectural constraints, limited receptive fields in message passing, and costly multi-scale designs, which restrict scalability to larger and more complex domains. We propose Fluid-DiT, a Graph-Free Diffusion Transformer that replaces graph message passing with attention-based denoising, eliminating explicit graph design while preserving the ability to model distributions of chaotic flows. Our framework introduces a latent-space formulation that disentangles geometric fidelity from distributional learning, reducing high-frequency artifacts and accelerating sampling. By leveraging the transformer's global receptive field, Fluid-DiT naturally captures both local flow structures and long-range correlations without requiring hierarchical graph coarsening. On canonical benchmarks including laminar cylinder wakes, ellipse-flow systems, and turbulent 3D wing experiments, Fluid-DiT consistently outperforms graph-based diffusion baselines in both sample quality and distributional accuracy, achieving higher $R^2$ correlations and lower Wasserstein distances. Moreover, it generalizes robustly from short, incomplete trajectories to unseen Reynolds numbers and geometries, demonstrating strong scalability.
Industrial design in fields such as vehicle and aerospace engineering often relies on large-scale numerical simulations to evaluate fluid dynamics performance, which can incur substantial computational costs. Deep neural networks have shown promise in improving simulation efficiency, especially graph neural networks (GNNs), which demonstrate great potential due to their flexibility with unstructured data. However, GNNs face challenges when dealing with tasks involving complex geometries and large-scale meshes. In this paper, we propose the Multi-scale Feature Enhanced Graph Neural Network (ME-GNN) to tackle these challenges. ME-GNN employs a graph neural network with a two-step message-passing mechanism to capture detailed local features effectively. Additionally, it integrates an Attention U-Net with uniform grid discretization, enabling the extraction of both fine and coarse features. The model also utilizes K-hop sampling to construct subgraphs, facilitating efficient training on large datasets while preserving detailed local features. We evaluated ME-GNN on three benchmark datasets and achieved state-of-the-art results: a relative L2 error of 0.0196 for the velocity field and 0.0556 for the surface pressure on ShapeNet-Car, a normalized mean squared error of 0.0033 for the flow field on AirfRANS, and a relative L2 error of 0.1416 for the surface pressure on DrivAerNet.
Solving the pressure-Poisson equation remains the primary computational bottleneck in incompressible unstructured flow solvers primarily due to the inherent sensitivity of traditional linear solvers to mesh irregularities. This work introduces a data-driven algebraic multigrid (AMG) smoother that uses a modified graph convolutional isomorphism network (GCIN). The graph neural network predicts optimal polynomial coefficients to construct a sparse pseudo-inverse operator across diverse grid topologies. The coefficients are optimized to reduce the residual after each V-cycle iteration. By directly capturing the algebraic structure of the system from the sparse coefficient matrix, the proposed method maintains the solver's linearity while adapting to local anisotropies in unstructured grids. Our framework demonstrates significant performance gains by reducing the number of V-cycles required for a given tolerance and delivering wall-clock speedups from 4% to 37% across diverse benchmarks. Notably, the model exhibits robust generalization by maintaining efficiency on meshes up to 128 times larger than those seen in training, and by accelerating the solver's convergence on unseen industry-relevant problems such as the AirfRANS dataset.
State-of-the-art deep learning models have been extensively utilized to reconstruct small-scale structures from coarse-grained data in turbulent flows. However, their application has predominantly been restricted to structured uniform meshes, limiting their applicability to data associated with complex geometries that are typically simulated on structured non-uniform or unstructured meshes. Machine learning (ML) models based on graph neural networks (GNNs), known for their ability to process unstructured data, offer a promising alternative. In this study, we leverage the inherent flexibility of GNNs featuring message passing layers to develop a methodology for reconstructing unresolved small-scale structures from low-resolution data on complex meshes. The accuracy of the proposed approach is demonstrated using two cases: a reacting channel flow on a structured non-uniform mesh, and a reacting hydrogen fueled internal combustion (IC) engine featuring an unstructured mesh. Evaluation of results based on visual agreement, statistical metrics, and cumulative error reduction indicates the effectiveness of the method in accurately reconstructing fine-scale features. Overall, this study provides a pathway for integrating data-driven small-scale reconstruction and subgrid-scale modeling to enhance the accuracy of coarse-grained simulations on complex meshes.
Urban wind flow modeling and simulation play an important role in air quality assessment and sustainable city planning. A key challenge for modeling and simulation is handling the complex geometries of the urban landscape. Low order models are limited in capturing the effects of geometry, while high-fidelity Computational Fluid Dynamics (CFD) simulations are prohibitively expensive, especially across multiple geometries or wind conditions. Here, we propose a generative diffusion framework for synthesizing steady-state urban wind fields over unstructured meshes that requires only geometry information. The framework combines a hierarchical graph neural network with score-based diffusion modeling to generate accurate and diverse velocity fields without requiring temporal rollouts or dense measurements. Trained across multiple mesh slices and wind angles, the model generalizes to unseen geometries, recovers key flow structures such as wakes and recirculation zones, and offers uncertainty-aware predictions. Ablation studies confirm robustness to mesh variation and performance under different inference regimes. This work develops is the first step towards foundation models for the built environment that can help urban planners rapidly evaluate design decisions under densification and climate uncertainty.
This study aims to overcome the limitations of conventional deep-learning approaches based on convolutional neural networks in complex geometries and unstructured meshes by exploring the potential of Graph U-Nets for unsteady flow-field prediction. We present a comprehensive investigation of Graph U-Nets, originally developed for classification tasks, now tailored for mesh-agnostic spatio-temporal forecasting of fluid dynamics. Our focus is on enhancing their performance through systematic hyperparameter tuning and architectural modifications. We propose novel approaches to improve mesh-agnostic spatio-temporal prediction of transient flow fields using Graph U-Nets, enabling accurate prediction on diverse mesh configurations. Key enhancements to the Graph U-Net architecture, including the Gaussian-mixture-model convolutional operator and noise injection approaches, provide increased flexibility in modeling node dynamics: the former reduces prediction error by 95\% compared to conventional convolutional operators, while the latter improves long-term prediction robustness, resulting in an error reduction of 86\%. We demonstrate the effectiveness of these enhancements in both transductive and inductive learning settings, showcasing the adaptability of Graph U-Nets to various flow conditions and mesh structures. This work contributes to the field of reduced-order modeling for computational fluid dynamics by establishing Graph U-Nets as a viable and flexible alternative to convolutional neural networks, capable of accurately and efficiently predicting complex fluid flow phenomena across diverse scenarios.
Numerical simulation of multi-phase fluid dynamics in porous media is critical for many energy and environmental applications in Earth's subsurface. Data-driven surrogate modeling provides computationally inexpensive alternatives to high-fidelity numerical simulators. While the commonly used convolutional neural networks (CNNs) are powerful in approximating partial differential equation solutions, it remains challenging for CNNs to handle irregular and unstructured simulation meshes. However, simulation models for Earth's subsurface often involve unstructured meshes with complex mesh geometries, which limits the application of CNNs. To address this challenge, we construct surrogate models based on Graph Convolutional Networks (GCNs) to approximate the spatial-temporal solutions of multi-phase flow and transport processes in porous media. We propose a new GCN architecture suited to the hyperbolic character of the coupled PDE system, to better capture transport dynamics. Results of 2D heterogeneous test cases show that our surrogates predict the evolutions of pressure and saturation states with high accuracy, and the predicted rollouts remain stable for multiple timesteps. Moreover, the GCN-based models generalize well to irregular domain geometries and unstructured meshes that are unseen in the training dataset.
Deep-learning-based surrogate models provide an efficient complement to numerical simulations for subsurface flow problems such as CO$_2$ geological storage. Accurately capturing the impact of faults on CO$_2$ plume migration remains a challenge for many existing deep learning surrogate models based on Convolutional Neural Networks (CNNs) or Neural Operators. We address this challenge with a graph-based neural model leveraging recent developments in the field of Graph Neural Networks (GNNs). Our model combines graph-based convolution Long-Short-Term-Memory (GConvLSTM) with a one-step GNN model, MeshGraphNet (MGN), to operate on complex unstructured meshes and limit temporal error accumulation. We demonstrate that our approach can accurately predict the temporal evolution of gas saturation and pore pressure in a synthetic reservoir with impermeable faults. Our results exhibit a better accuracy and a reduced temporal error accumulation compared to the standard MGN model. We also show the excellent generalizability of our algorithm to mesh configurations, boundary conditions, and heterogeneous permeability fields not included in the training set. This work highlights the potential of GNN-based methods to accurately and rapidly model subsurface flow with complex faults and fractures.
Graph neural networks (GNNs) have shown promise in learning unstructured mesh-based simulations of physical systems, including fluid dynamics. In tandem, geometric deep learning principles have informed the development of equivariant architectures respecting underlying physical symmetries. However, the effect of rotational equivariance in modeling fluids remains unclear. We build a multi-scale equivariant GNN to forecast fluid flow and study the effect of modeling invariant and non-invariant representations of the flow state. We evaluate the model performance of several equivariant and non-equivariant architectures on predicting the evolution of two fluid flows, flow around a cylinder and buoyancy-driven shear flow, to understand the effect of equivariance and invariance on data-driven modeling approaches. Our results show that modeling invariant quantities produces more accurate long-term predictions and that these invariant quantities may be learned from the velocity field using a data-driven encoder.
Physical fields on meshes require a separation between topology and geometry: differential identities should remain exact, while geometry, material response, and anisotropic coupling are learned from data. Existing neural surrogates often mix these roles inside unconstrained message passing. We introduce Riemannian Hodge Message Passing (RHMP), which turns this separation into an architectural principle. RHMP fixes the cellular coboundaries ($d_k$) determined by oriented incidence and learns symmetric positive-definite cochain metrics ($H_k$) for geometry-dependent propagation. The learned metric defines how cochain fields are measured and coupled; its invariance to orthogonal changes of the hidden feature basis yields cochain-frame-equivariant propagation. RHMP implements this principle with metric-weighted Hodge blocks ($d_k^\top H_{k+1}d_k$), yielding exact cochain-complex identities ($d_{k+1}d_k=0$), nonnegative Hodge energies, positive-semidefinite operators, and exact Abelian curvature invariance. Across seven physical benchmarks spanning fluids, electromagnetism, gauge fields, and variable-mesh CFD, RHMP performs strongly across these systems, with the largest gains when topology, learned geometry, and field structure interact.
Computational Fluid Dynamics (CFD) is central to race-car aerodynamic development, yet its cost -- tens of thousands of core-hours per high-fidelity evaluation -- severely limits the design space exploration feasible within realistic budgets. AI-based surrogate models promise to alleviate this bottleneck, but progress has been constrained by the limited complexity of public datasets, which are dominated by smoothed passenger-car shapes that fail to exercise surrogates on the thin, complex, highly loaded components governing motorsport performance. This work presents three primary contributions. First, we introduce a high-fidelity RANS dataset built on a parametric LMP2-class CAD model and spanning six operating conditions (map points) covering straight-line and cornering regimes, generated and validated by aerodynamics experts at Dallara to preserve features relevant to industrial motorsport. Second, we present the Gauge-Invariant Spectral Transformer (GIST), a graph-based neural operator whose spectral embeddings encode mesh connectivity to enhance predictions on tightly packed, complex geometries. GIST guarantees discretization invariance and scales linearly with mesh size, achieving state-of-the-art accuracy on both public benchmarks and the proposed race-car dataset. Third, we demonstrate that GIST achieves a level of predictive accuracy suitable for early-stage aerodynamic design, providing a first validation of the concept of interactive design-space exploration -- where engineers query a surrogate in place of the CFD solver -- within industrial motorsport workflows.
Neural operators on irregular meshes face a fundamental tension. Spectral positional encodings, the natural choice for capturing geometry, require cubic-complexity eigendecomposition and inadvertently break gauge invariance through numerical solver artifacts; existing efficient approximations sacrifice gauge symmetry by design. Both failure modes break discretization invariance: models fail to transfer across mesh resolutions of the same domain, and similarly across different graphs of related structure in inductive settings. We propose GIST (Gauge-Invariant Spectral Transformer), a scalable neural operator that resolves this tension by restricting attention to pairwise inner products of efficient approximate spectral embeddings. We prove these inner products estimate an exactly gauge-invariant graph kernel at end-to-end $\mathcal{O}(N)$ complexity, and establish a formal connection between gauge invariance and discretization-invariant learning with bounded mismatch error. To our knowledge, GIST is the first scalable graph neural operator with a provable discretization-mismatch bound. Empirically, GIST sets state-of-the-art on the AirfRANS, ShapeNet-Car, DrivAerNet, and DrivAerNet++ mesh benchmarks (up to 750K nodes), and additionally matches strong baselines on standard graph benchmarks (e.g., 99.50% micro-F1 on PPI).
We propose a novel data-lean operator learning algorithm, the Reduced Basis Neural Operator (ReBaNO), to solve a group of PDEs with multiple distinct inputs. Inspired by the Reduced Basis Method and the recently introduced Generative Pre-Trained Physics-Informed Neural Networks, ReBaNO relies on a mathematically rigorous greedy algorithm to build its network structure offline adaptively from the ground up. Knowledge distillation via task-specific activation function allows ReBaNO to have a compact architecture requiring minimal computational cost online while embedding physics. In comparison to state-of-the-art operator learning algorithms such as PCA-Net, DeepONet, FNO, and CNO, numerical results demonstrate that ReBaNO significantly outperforms them in terms of eliminating/shrinking the generalization gap for both in- and out-of-distribution tests and being the only operator learning algorithm achieving strict discretization invariance.
Time-dependent partial differential equations are ubiquitous in physics-based modeling, but they remain computationally intensive in many-query scenarios, such as real-time forecasting, optimal control, and uncertainty quantification. Reduced-order modeling (ROM) addresses these challenges by constructing a low-dimensional surrogate model but relies on a fixed discretization, which limits flexibility across varying meshes during evaluation. Operator learning approaches, such as neural operators, offer an alternative by parameterizing mappings between infinite-dimensional function spaces, enabling adaptation to data across different resolutions. Whereas ROM provides rigorous numerical error estimates, neural operator learning largely focuses on discretization convergence and invariance without quantifying the error between the infinite-dimensional and the discretized operators. This work introduces the reduced-order neural operator modeling (RONOM) framework, which bridges concepts from ROM and operator learning. We establish a discretization error bound analogous to those in ROM, and get insights into RONOM's discretization convergence and discretization robustness. Moreover, three numerical examples are presented that compare RONOM to existing neural operators for solving partial differential equations. The results demonstrate that RONOM using standard vector-to-vector neural networks can achieve comparable performance in input generalization and achieves superior performance in both spatial super-resolution and discretization robustness, while also offering novel insights into temporal super-resolution scenarios and ROM-based approaches for learning on time-dependent data.