GNN + Autograd 技术路径文献调研
Physics-informed GNN 总体范式与 PDE 软约束基线
这一组是本次调研的"背景与判别基准":它们都在图上做物理方程约束,但主要形态是 GNN 输出后(或结合离散弱形式/变分形式)把 PDE 残差作为损失项或约束项加入训练,并未把空间微分嵌入消息传递内部。把它们集中列出,一是为了界定"PDE 残差作 Loss ≠ 导数嵌入 Message Passing"这一关键区别(调研要求第三节第 5 条),二是为了识别其中少数确实使用了 Autograd 的工作(如 SGPINN 的时间导数、PGGNO 的可微物理残差)与仅用预计算算子/形状函数梯度的工作(如 data-free 弱形式 GNN)之间的分野,避免按关键词误归类。包含电磁 Maxwell 变分形式、Laplace-Beltrami 本征对 dPINN、结构力学 DeepONet、数据中心热场等应用。
- Unsupervised learning with physics informed graph networks for partial differential equations(Lin Lu, Yi-Ye Zou, Jing-Yu Wang, Shufan Zou, Laiping Zhang, Xiao-Gang Deng, 2025, Applied intelligence (Boston))
- Physics-informed graph neural network for predicting fluid flow in porous media(Haiyang Chen, Liang Xue, Li Liu, Gaofeng Zou, Jiang-Xia Han, Yu Dong, Meng-Ze Cong, Yue-Tian Liu, S. Hosseini-Nasab, 2025, Petroleum Science)
- 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)
- Combining physics-informed graph neural network and finite difference for solving forward and inverse spatiotemporal PDEs(Hao Zhang, Longxiang Jiang, Xinkun Chu, Yong Wen, Luxiong Li, Jianbo Liu, Yonghao Xiao, Liyuan Wang, 2024, Computer Physics Communications)
- Isogeometric analysis-based physics-informed graph neural network for studying traffic jam in neurons(Ang-Ran Li, Y. Zhang, 2023, Computer Methods in Applied Mechanics and Engineering)
- A Graph-Structured, Physics-Informed DeepONet Neural Network for Complex Structural Analysis(Guangya Zhang, Tie Xu, Jin-Li Xu, Hu Wang, 2025, Machine Learning and Knowledge Extraction)
- Physics-Informed Graph Neural Network for Electromagnetic Simulations(Ya-Wei Su, S. Zeng, Xu-Qing Wu, Yue-Qin Huang, Jie-Fu Chen, 2023, URSI General Assembly and Scientific Symposium)
- A Physics-Informed Graph Neural Network for Computing Laplace–Beltrami Eigenfunctions on Manifolds(D. Lazzaro, S. Morigi, P. Zuzolo, 2025, IEEE Access)
- SPATIOTEMPORAL GRAPH PHYSICS-INFORMED NEURAL NETWORKS FOR DATA CENTER THERMAL MODELING(Shang-Peng Lei, G. Balakayeva, 2026, Scientific Journal of Astana IT University)
- Physics-Guided Graph Neural Operators for Real-Time, Safe Control of District Heating Networks(Sanqing Li, Xinyu Zhao, 2025, 2025 6th International Conference on Smart Grid and Energy Engineering (SGEE))
GNN 上的 Autograd / 解析微分与梯度增强监督(高阶导数能力)
这一组回答"在 GNN 上通过自动微分到底能拿到什么阶数的导数、这些导数是否等于物理空间导数"这一前置问题。共同点是:都以 GNN(或等变图网络、图算子)为载体,用 Autograd 或解析梯度显式求取一阶梯度、二阶导数/Hessian 或灵敏度矩阵,但导数的用途主要是构造训练目标(梯度监督、Hessian 正则、场-梯度自洽、顺式灵敏度一致性),而非作为下一层消息传递的输入特征。因此它们为本次构想提供了"可行性证据 + 计算开销证据"(二阶导的显存/时间代价),同时也是需要严格区分的对照:对坐标求导/预测梯度场,并不自动等于在局部消息传递内部获得了连续物理场偏导数。
- Beyond Numerical Hessians: Higher-Order Derivatives for Machine Learning Interatomic Potentials via Automatic Differentiation(Nils Gönnheimer, K. Reuter, J. Margraf, 2025, Journal of Chemical Theory and Computation)
- Multi-task learning for electronic structure to predict and explore molecular potential energy surfaces(Zhuoran Qiao, Fei-Zhi Ding, Matthew Welborn, Peter J. Bygrave, Daniel G. A. Smith, Anima Anandkumar, F. Manby, T. Miller, 2020, arXiv.org)
- Rapid 2-D Magnetotelluric Forward Modeling via Automatic Differentiation Fourier Neural Operator With Gradient Supervision(Xin Zhong, Weiwei Ling, Ke-Jia Pan, Pinxia Wu, Jiajing Zhang, Zhiliang Zhan, 2026, IEEE Transactions on Geoscience and Remote Sensing)
- Geo-PhysNet: A Geometry-Aware and Physics-Constrained Graph Neural Network for Aerodynamic Pressure Prediction on Vehicle Fluid–Solid Surfaces(Bo-Wen Liu, Hao Wang, Liheng Xue, Yin Long, 2025, Applied Sciences)
- A Systematic Analysis of Automatic Differentiation versus Discretization-based Constraints for Physics-Informed PDE Solvers(Xing Guo, Hong-Wei Tang, Ze-Wei Meng, Yi-Dong Zhang, Shao-Qiu Xiao, Feng Liu, 2026, arXiv.org)
物理量/几何张量嵌入消息传递通道的 derivative-aware 近似工作
这一组最接近本次调研的核心技术构想,是"把物理/几何微分信息送进 message passing 而不是只送进 loss"的代表性探索:Navier-Stokes 项作为边特征参与聚合、以边局部参考系保证旋转等变与平移不变的守恒量消息、由隐式场景场梯度驱动并反传可微积分器的粒子图网络、以及在同一科学问题上直接对照"连续场 AD(F-PINN)"与"图消息传递(F-GNN)"两种范式的比较研究。集中为一组,是因为它们分别从"边特征注入物理微分项""边局部坐标系与等变性""隐式场梯度驱动""AD 与消息传递范式对照"四个角度,为"局部相对坐标 + 可微消息"的可行性提供了最直接的证据与实验参照;同时也要指出它们的共同局限——导数多为预计算或由外部连续场提供,尚未在每条边上把消息写成关于局部查询坐标的可微函数并在原点处求高阶导。
- 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)
- A physics-informed graph neural network conserving linear and angular momentum for dynamical systems(Vinay Sharma, Olga Fink, 2026, Nature Communications)
- Particle Simulation Study of Visual Interaction in Neural Network-Driven Interior Environment Design(Mingyang Li, Yan-Song Hu, 2026, International Conference on Computational Intelligence and Communication Networks)
- Deep Learning Fractal Superconductivity: A Comparative Study of Physics-Informed and Graph Neural Networks Applied to the Fractal TDGL Equation(Călin Gheorghe Buzea, F. Nedeff, Diana-Carmen Mirilă, M. Agop, Decebal Vasîncu, 2025, Fractal and Fractional)
局部连续重构与高阶导数一致性的数学工具基础(RBF / 再生核 / 局部 stencil / 可微有限元)
这一组提供实现"每条边消息为关于局部查询坐标的可微函数、并在节点自身位置稳定取得二阶导数"所必需的数学与算子基础:局部 RBF 与 RBF-Hermite 局部 stencil 对 Laplace 算子的高精度逼近、单位分解式局部 meshfree 方法在不规则域上的适用性、再生核/隐式梯度框架下的任意阶导数与多项式再生(polynomial reproduction)条件、基于有限元形状函数梯度而非 AD 的弱形式残差计算,以及按单纯形重心坐标构造、可通过自定义反向微分规则在同一计算图中同时获得函数值与空间梯度的稀疏连接网络表示。它们共同回答"多项式完备性、变节点密度下的导数精度、边界附近局部重构、非结构网格适用性"这些关键考察点,是本构想在消息级局部重构层的可直接借鉴技术。
- Generalized reproducing kernel peridynamics: unification of local and non-local meshfree methods, non-local derivative operations, and an arbitrary-order state-based peridynamic formulation(M. Hillman, M. Pasetto, Guo-Hua Zhou, 2020, Computational Particle Mechanics)
- A local RBF-based meshfree method for the three-dimensional space-fractional advection-dispersion equation on irregular convex domains(Xiao-Gang Zhu, 2025, Computational Mechanics)
- A High‐Order Local Meshfree Method for Solving the Two‐Dimensional Multiterm Time‐Fractional Partial Integro‐Differential Equation(Mohammad Ilati, S. Eslami, 2025, Mathematical methods in the applied sciences)
- A local meshfree method based on radial basis function partition of unity method for two-dimensional (2D) fractional Rayleigh-Stokes problem(Ö. Oruç, 2026, Hacettepe Journal of Mathematics and Statistics)
- Data-free neural PDE solvers based on Graph Neural Networks and weak forms(Mikel M. Iparraguirre, I. Alfaro, David González, E. Cueto, 2026, arXiv.org)
- Sparsely connected neural network representation of Lagrange finite element function(Jia-Xiong Hao, Yun-Qing Huang, Nian-Yu Yi, 2026, arXiv.org)
本组文献围绕"GNN + Autograd"技术路径展开,分为四条主线:(1) Physics-informed GNN 的总体范式与 PDE 软约束基线,用于界定"PDE Loss 作为损失项"与"导数是网络内部特征"的区别;(2) 通过 Autograd/解析微分在 GNN 上获取一阶、二阶乃至 Hessian 级别导数,并将其作为监督或结构一致性目标的梯度增强类工作;(3) 真正把 PDE 项、几何量或隐式场梯度放进消息传递通道的近似 derivative-aware message passing 工作;(4) 为"局部连续重构 + 稳定高阶导数"提供数学基础的 meshfree/RBF/再生核/局部 stencil 与可微有限元表示。注意到:现有文献中尚无一例完整实现"局部相对坐标依赖的 Message → 在节点自身位置 Autograd 求导 → 导数增强 Message Passing → PDE Loss"闭环;多数工作停留在其一端(要么导数只进 Loss,要么 PDE 项只是预计算边特征),这正是本调研需要明确指出的研究空白。
这一组是本次调研的"背景与判别基准":它们都在图上做物理方程约束,但主要形态是 GNN 输出后(或结合离散弱形式/变分形式)把 PDE 残差作为损失项或约束项加入训练,并未把空间微分嵌入消息传递内部。把它们集中列出,一是为了界定"PDE 残差作 Loss ≠ 导数嵌入 Message Passing"这一关键区别(调研要求第三节第 5 条),二是为了识别其中少数确实使用了 Autograd 的工作(如 SGPINN 的时间导数、PGGNO 的可微物理残差)与仅用预计算算子/形状函数梯度的工作(如 data-free 弱形式 GNN)之间的分野,避免按关键词误归类。包含电磁 Maxwell 变分形式、Laplace-Beltrami 本征对 dPINN、结构力学 DeepONet、数据中心热场等应用。相关文献:Lin Lu et. al, 2025 等 10 篇文献
这一组回答"在 GNN 上通过自动微分到底能拿到什么阶数的导数、这些导数是否等于物理空间导数"这一前置问题。共同点是:都以 GNN(或等变图网络、图算子)为载体,用 Autograd 或解析梯度显式求取一阶梯度、二阶导数/Hessian 或灵敏度矩阵,但导数的用途主要是构造训练目标(梯度监督、Hessian 正则、场-梯度自洽、顺式灵敏度一致性),而非作为下一层消息传递的输入特征。因此它们为本次构想提供了"可行性证据 + 计算开销证据"(二阶导的显存/时间代价),同时也是需要严格区分的对照:对坐标求导/预测梯度场,并不自动等于在局部消息传递内部获得了连续物理场偏导数。相关文献:Nils Gönnheimer et. al, 2025 等 5 篇文献
这一组最接近本次调研的核心技术构想,是"把物理/几何微分信息送进 message passing 而不是只送进 loss"的代表性探索:Navier-Stokes 项作为边特征参与聚合、以边局部参考系保证旋转等变与平移不变的守恒量消息、由隐式场景场梯度驱动并反传可微积分器的粒子图网络、以及在同一科学问题上直接对照"连续场 AD(F-PINN)"与"图消息传递(F-GNN)"两种范式的比较研究。集中为一组,是因为它们分别从"边特征注入物理微分项""边局部坐标系与等变性""隐式场梯度驱动""AD 与消息传递范式对照"四个角度,为"局部相对坐标 + 可微消息"的可行性提供了最直接的证据与实验参照;同时也要指出它们的共同局限——导数多为预计算或由外部连续场提供,尚未在每条边上把消息写成关于局部查询坐标的可微函数并在原点处求高阶导。相关文献:Amirkhosro Kazemi et. al, 2024 等 4 篇文献
这一组提供实现"每条边消息为关于局部查询坐标的可微函数、并在节点自身位置稳定取得二阶导数"所必需的数学与算子基础:局部 RBF 与 RBF-Hermite 局部 stencil 对 Laplace 算子的高精度逼近、单位分解式局部 meshfree 方法在不规则域上的适用性、再生核/隐式梯度框架下的任意阶导数与多项式再生(polynomial reproduction)条件、基于有限元形状函数梯度而非 AD 的弱形式残差计算,以及按单纯形重心坐标构造、可通过自定义反向微分规则在同一计算图中同时获得函数值与空间梯度的稀疏连接网络表示。它们共同回答"多项式完备性、变节点密度下的导数精度、边界附近局部重构、非结构网格适用性"这些关键考察点,是本构想在消息级局部重构层的可直接借鉴技术。相关文献:M. Hillman et. al, 2020 等 6 篇文献
总计 25 篇相关文献
The development of machine learning interatomic potentials (MLIPs) has revolutionized computational chemistry by enhancing the accuracy of empirical force fields while retaining a large computational speed-up compared to first-principles calculations. Despite these advancements, the calculation of Hessian matrices for large systems remains challenging, in particular because analytical second-order derivatives are often not implemented. This necessitates the use of computationally expensive finite-difference methods, which can furthermore display low precision in some cases. Automatic differentiation (AD) offers a promising alternative to reduce this computational effort and makes the calculation of Hessian matrices more efficient and accurate. Here, we present the implementation of AD-based second-order derivatives for the popular MACE equivariant graph neural network architecture. The benefits of this method are showcased via a high-throughput prediction of heat capacities of porous materials with the MACE-MP-0 foundation model. This is essential for precisely describing gas adsorption in these systems and was previously possible only with bespoke ML models or expensive first-principles calculations. We find that the availability of foundation models and accurate analytical Hessian matrices offers comparable accuracy to bespoke ML models in a zero-shot manner and additionally allows for the investigation of finite-size and rounding errors in the first-principles data.
The fractal extension of the time-dependent Ginzburg–Landau (TDGL) equation, formulated within the framework of Scale Relativity, generalizes superconducting dynamics to non-differentiable space–time. Although analytically well established, its numerical solution remains difficult because of the strong coupling between amplitude and phase curvature. Here we develop two complementary deep learning solvers for the fractal TDGL (FTDGL) system. The Fractal Physics-Informed Neural Network (F-PINN) embeds the Scale-Relativity covariant derivative through automatic differentiation on continuous fields, whereas the Fractal Graph Neural Network (F-GNN) represents the same dynamics on a sparse spatial graph and learns local gauge-covariant interactions via message passing. Both models are trained against finite-difference reference data, and a parametric study over the dimensionless fractality parameter D quantifies its influence on the coherence length, penetration depth, and peak magnetic field. Across multivortex benchmarks, the F-GNN reduces the relative L2 error on ψ2 from 0.190 to 0.046 and on Bz from approximately 0.62 to 0.36 (averaged over three seeds). This ≈4× improvement in condensate-density accuracy corresponds to a substantial enhancement in vortex-core localization—from tens of pixels of uncertainty to sub-pixel precision—and yields a cleaner reconstruction of the 2π phase winding around each vortex, improving the extraction of experimentally relevant observables such as ξeff, λeff, and local Bz peaks. The model also preserves flux quantization and remains robust under 2–5% Gaussian noise, demonstrating stable learning under experimentally realistic perturbations. The D—scan reveals broader vortex cores, a non-monotonic variation in the penetration depth, and moderate modulation of the peak magnetic field, while preserving topological structure. These results show that graph-based learning provides a superior inductive bias for modeling non-differentiable, gauge-coupled systems. The proposed F-PINN and F-GNN architectures therefore offer accurate, data-efficient solvers for fractal superconductivity and open pathways toward data-driven inference of fractal parameters from magneto-optical or Hall-probe imaging experiments.
Accurate thermal modeling is the precondition for improving energy efficiency and operational reliability of data centers (DCs). However, airflow circulation and workloads cause the temperature field of DCs to exhibit complex spatiotemporal variations. Conventional data-driven models formulate temperature prediction as a multivariate time-series task, without explicitly accounting for spatial heat interactions. In comparison, physics-based methods require detailed geometric information and high computational resources. To address these limitations, we propose a spatiotemporal graph physics-informed neural network (SGPINN) framework for multi-node and multi-step DC temperature prediction. Inspired by resistance-capacitance thermal modeling, SGPINN represents the thermal field as a graph, where graph aggregation captures spatial thermal interactions, and a physical residual combines temperature prediction with graph-guided thermal propagation. The temporal temperature derivative is obtained through automatic differentiation, while learnable equivalent thermal capacitance and heat-flow terms are introduced to characterize heterogeneous node dynamics without requiring directly measured physical parameters. A learnable multi-task strategy is adopted to adaptively balance the data loss and physics loss. The model is evaluated using both a real-world DC dataset and a simulated DC dataset and compared with four representative baseline models. SGPINN achieves the best overall performance on both datasets, with average MAEs of 0.1097 °C and 0.3043 °C, respectively. In the real-world case, it reduces average MAE and RMSE by 3.57% and 3.58% compared with APINN, while ablation results verify the contributions of graph modeling, physical constraints, and adaptive loss weighting. These findings suggest that SGPINN improves predictive accuracy while providing a physics-guided framework for DC thermal modeling.
No abstract available
No abstract available
A physics-informed graph neural network conserving linear and angular momentum for dynamical systems
Accurate, interpretable, and real-time modeling of multi-body dynamical systems is essential for predicting behaviors and inferring physical properties in natural and engineered environments. Traditional physics-based models face scalability challenges and are computationally demanding, while data-driven approaches like Graph neural networks (GNNs) often lack physical consistency, interpretability, and generalization. In this paper, we propose DYNAMI-CAL GRAPHNET, a Physics-Informed Graph Neural Network that integrates the learning capabilities of GNNs with physics-based inductive biases to address these limitations. DYNAMI-CAL GRAPHNET enforces pairwise conservation of linear and angular momentum for interacting nodes using edge-local reference frames that are equivariant to rotational symmetries, invariant to translations, and equivariant to node permutations. This design ensures physically consistent predictions of node dynamics while offering interpretable, edge-wise linear and angular impulses resulting from pairwise interactions. Evaluated on a 3D granular system with inelastic collisions, DYNAMI-CAL GRAPHNET demonstrates stable error accumulation over extended rollouts, effective extrapolation to unseen configurations, and robust handling of heterogeneous interactions and external forces. DYNAMI-CAL GRAPHNET offers significant advantages in fields requiring accurate, interpretable, and real-time modeling of complex multi-body dynamical systems, such as robotics, aerospace engineering, and materials science. By providing physically consistent and scalable predictions that adhere to fundamental conservation laws, it enables the inference of forces and moments while efficiently handling heterogeneous interactions and external forces. This makes it invaluable for designing control systems, optimizing mechanical processes, and analyzing dynamic behaviors in both natural and engineered systems. Learning complex dynamics from data often leads to unstable or unphysical predictions. Here, the authors introduce Dynami-CAL GraphNet, a physics-informed architecture that conserves linear and angular momentum and enables accurate rollouts across diverse dynamical systems.
No abstract available
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.
No abstract available
No abstract available
This study introduces the Graph-Structured Physics-Informed DeepONet (GS-PI-DeepONet), a novel neural network framework designed to address the challenges of solving parametric Partial Differential Equations (PDEs) in structural analysis, particularly for problems with complex geometries and dynamic boundary conditions. By integrating Graph Neural Networks (GNNs), Deep Operator Networks (DeepONets), and Physics-Informed Neural Networks (PINNs), the proposed method employs graph-structured representations to model unstructured Finite Element (FE) meshes. In this framework, nodes encode physical quantities such as displacements and loads, while edges represent geometric or topological relationships. The framework embeds PDE constraints as soft penalties within the loss function, ensuring adherence to physical laws while reducing reliance on large datasets. Extensive experiments have demonstrated the GS-PI-DeepONet’s superiority over traditional Finite Element Methods (FEMs) and standard DeepONets. For benchmark problems, including cantilever beam bending and Hertz contact, the model achieves high accuracy. In practical applications, such as stiffness analysis of a recliner mechanism and strength analysis of a support bracket, the framework achieves a 7–8 speed-up compared to FEMs, while maintaining fidelity comparable to FEM, with R2 values reaching up to 0.9999 for displacement fields. Consequently, the GS-PI-DeepONet offers a resolution-independent, data-efficient, and physics-consistent approach for real-time simulations, making it ideal for rapid parameter sweeps and design optimizations in engineering applications.
Physic-informed neural networks (PINNs) have been widely used in computational physics including electromagnetic simulations. Unlike conventional numerical solvers, PINNs leverage deep neural networks and optimization to find the solution of Maxwell’s equations given specific source and boundary conditions. When the material in the solving domain is heterogeneous, it’s not easy to enforce continuity of electromagnetic fields due to the meshless property of PINNs. On the contrary, graph neural networks (GNNs) explicitly construct a graph (mesh) and apply message passing between the nodes in the graph. In this work, GNNs and PINNs are combined to build physics-informed graph neural networks (PIGNNs) that are flexible in mesh discretization and straightforward to enforce continuity of electromagnetic fields. Moreover, PIGNNs get easier to be trained and converge by resorting to the variational formulation of Maxwell’s equations.
District heating networks (DHNs) are key assets in urban energy systems, yet their real-time operation remains challenging due to nonlinear thermo-hydraulics, time-varying boundary conditions, and stringent safety constraints. Classical model predictive control (MPC) hinges on calibrated simulators that are often too slow or brittle for online use, while purely data-driven models struggle with long-horizon rollouts and constraint satisfaction. We propose a physics-guided graph neural operator (PGGNO) that learns fast, topology-aware surrogates of DHN dynamics and embeds first-principles structure via a differentiable physics residual. PGGNO maps boundary inputs to spatio-temporal fields of temperature and pressure, scales to varying network topologies, and exposes gradients for control. We couple PGGNO with a safety-filtered MPC scheme that enforces operational limits through differentiable inequality penalties and a robust backup controller. On three DHN benchmarks with up to 850 nodes, PGGNO reduces multi-step prediction error by 25-40
Geophysical forward modeling is the cornerstone of inversion, with its accuracy and efficiency directly influencing the quality and efficiency of the inversion. For frequency-domain magnetotelluric (MT) problems, despite the excellent approximation capabilities of deep learning (DL) frameworks such as the Fourier neural operator (FNO) in fast forward modeling, they suffer from strong data dependence and insufficient physical interpretability due to inadequate utilization of sensitivity information inherent in physical equations. To address this, we propose the automatic differentiation FNO (ADFNO), a 2-D MT forward modeling architecture empowered by data-augmented gradient supervision. By leveraging automatic differentiation (AD), we explicitly integrate the 2-D MT sensitivity matrix into the operator’s training framework. This enforces consistency between the network-predicted sensitivity and the numerically computed reference, guiding the model to better capture the derivative-level relationship between resistivity perturbations and electromagnetic responses. Numerical experiments verify that ADFNO can adaptively predict electromagnetic fields at arbitrary grid positions and frequencies, demonstrating a strong regularization effect of sensitivity and opening up a new avenue for 2-D MT forward modeling.
The aerodynamic pressure of a car is crucial for its shape design. To overcome the time-consuming and costly bottleneck of wind tunnel tests and computational fluid dynamics (CFD) simulations, deep learning-based surrogate models have emerged as highly promising alternatives. However, existing methods that only predict on the surface of objects only learn the mapping of pressure. In contrast, a physically realistic field has values and gradients that are structurally unified and self-consistent. Therefore, existing methods ignore the crucial differential structure and intrinsic continuity of the physical field as a whole. This oversight leads to their predictions, even if locally numerically close, often showing unrealistic gradient distributions and high-frequency oscillations macroscopically, greatly limiting their reliability and practicality in engineering decisions. To address this, this study proposes the Geo-PhysNet model, a graph neural network framework specifically designed for complex surface manifolds with strong physical constraints. This framework learns a differential representation, and its network architecture is designed to simultaneously predict the pressure scalar field and its tangential gradient vector field on the surface manifold within a unified framework. By making the gradient an explicit learning target, we force the network to understand the local mechanical causes leading to pressure changes, thereby mathematically ensuring the self-consistency of the field’s intrinsic structure, rather than merely learning the numerical mapping of pressure. Finally, to solve the common noise problem in the predictions of existing methods, we introduce a physical regularization term based on the surface Laplacian operator to penalize non-smooth solutions, ensuring the physical rationality of the final output field. Experimental verification results show that Geo-PhysNet not only outperforms existing benchmark models in numerical accuracy but, more importantly, demonstrates superior advantages in the physical authenticity, field continuity, and gradient smoothness of the generated pressure fields.
The focus of this paper is to present a new fourth‐order local meshfree technique for numerically solving the multiterm time‐fractional partial integro‐differential equation. The time‐fractional derivative term, defined in the Caputo sense, is approximated using the weighted and shifted Grünwald difference (WSGD) formula. Also, a second‐order technique for approximating the integral term in the Riemann–Liouville sense leads to the semidiscrete formulation of the model. Applying the meshless radial basis function (RBF)‐Hermite finite difference (RBF‐HFD) technique for spatial discretization completes the model's discretization process. The novelty of this method lies in the approximation of the Laplace operator using the Hermite RBF interpolation over local stencils. This leads to RBF‐generated Hermite finite difference (RBF‐HFD) formula, which provides a significant improvement in the accuracy and computational efficiency. Its meshfree nature also makes it very suitable for solving partial integro‐differential equations (PIDEs) on irregular domains. In the subsequent discussion, the unconditional stability and convergence of the time‐discrete scheme are proven via the energy method.The accuracy and efficiency of the proposed scheme are investigated through numerical simulations on regular and irregular domains. Furthermore, the numerical outcomes are compared with existing methods in the literature to highlight the accuracy and efficiency of the present technique.
No abstract available
State-based peridynamics is a non-local reformulation of solid mechanics that replaces the force density of the divergence of stress with an integral of the action of force states on bonds local to a given position, which precludes differentiation with the aim to model strong discontinuities effortlessly. A popular implementation is a meshfree formulation where the integral is discretized by quadrature points, which results in a series of unknowns at the points under the strong-form collocation framework. In this work, the meshfree discretization of state-based peridynamics under the correspondence principle is examined and compared to traditional meshfree methods based on the classical local formulation of solid mechanics. It is first shown that the way in which the peridynamic formulation approximates differentiation can be unified with the implicit gradient approximation, and this is termed the reproducing kernel peridynamic approximation. This allows the construction of non-local deformation gradients with arbitrary-order accuracy, as well as non-local approximations of higher-order derivatives. A high-order accurate non-local divergence of stress is then proposed to replace the force density in the original state-based peridynamics, in order to obtain global arbitrary-order accuracy in the numerical solution. These two operators used in conjunction with one another is termed the reproducing kernel peridynamic method. The strong-form collocation version of the method is tested against benchmark solutions to examine and verify the high-order accuracy and convergence properties of the method. The method is shown to exhibit superconvergent behavior in the nodal collocation setting.
This study dedicated to provide numerical solutions of fractionalRayleigh-Stokes problem in two-dimension (2D). The problem has Riemann-Liouvilletype fractional derivative. To obtain time discretization for theconsidered problem, we use a difference scheme along with an integrationprocedure. Hence a time semi-discretized scheme is obtained. We examinestability and convergence of the time semi-discretized scheme. Thenwe developed a local radial basis function partition of unity methodfor discretization of space variables. By combining time and spacediscretization we acquire a full discrete scheme for solving the consideredproblem. We solve some test problems and compare obtained resultsfor the test problems with existing techniques in the literature,to show efficiency of the proposed method.
Eigenfunctions and the spectrum of the Laplace-Beltrami operator are fundamental tools in geometry processing, providing a robust framework for analyzing complex geometries. This paper introduces an incremental framework specifically designed for the computation of the leading eigenpairs of the Laplace-Beltrami operator of 2-manifolds represented by triangular surface meshes in <inline-formula> <tex-math notation="LaTeX">${\mathbb {R}}^{3}$ </tex-math></inline-formula>. The proposed approach relies on a discrete physics-informed neural network (dPINN), which leverages Graph Neural Networks to effectively model the intrinsic geometry and topology of the manifold. dPINNs are ideally suited for learning on a triangular surface mesh in <inline-formula> <tex-math notation="LaTeX">${\mathbb {R}}^{3}$ </tex-math></inline-formula> because they leverage its discrete representation and numerical differentiation. Conversely, continuous PINNs (cPINNs) prove inadequate when the spatial domain (manifold) is not readily available for arbitrary sampling. Numerical experiments on several 2-manifolds in <inline-formula> <tex-math notation="LaTeX">${\mathbb {R}}^{2} / {\mathbb {R}}^{3}$ </tex-math></inline-formula> demonstrate the accuracy, robustness to geometric and topological variations, and generalization capabilities of the proposed neural network approach in case of 2-manifolds with parametric representations. The presented method aims to demonstrate how learning-based approaches can improve upon traditional numerical methods under adverse meshing conditions or domain discretization changes.
Particle Simulation Study of Visual Interaction in Neural Network-Driven Interior Environment Design
Traditional interior design tools, relying on static parameter presets, cannot resolve the nonlinear mapping relationship between user visual interaction signals and spatial dynamic evolution. This paper proposes a differentiable graph neural particle network based on visual attention modulation. This method encodes eye-tracking trajectories into a two-dimensional Gaussian mixture attention field and embeds it as a spatial modulation gating mechanism into the message passing process of the dynamic graph. Particle nodes construct adjacency relationships based on radial basis functions, and their edge weights are modulated point-by-point by the attention field and coupled with the Lennard-Jones potential to achieve a redistribution of visually driven interaction intensity. Simultaneously, an implicit scalar field parameterized by SIREN (Sinusoidal Representation Networks) is introduced to represent the semantic attributes of the environment. Particle motion is guided by the field gradient and backpropagated through a differentiable integrator to optimize the mapping function from visual input to spatial parameters end-to-end. Experiments are validated on an interactive dataset containing 127 real-world interior scenes. This paper controls the mapping delay from visual signals to particle trajectories to 23.7 milliseconds, and maintains the particle system energy conservation deviation at $1.3 \times 10^{-6}$. On the order of ${ }^{-4}$, this method establishes a differentiable closed loop between visual perception and spatial generation, providing a perception-driven real-time dynamic generation paradigm for interior environment design.
We construct a mesh-induced sparsely connected neural network framework that exactly reproduces arbitrary-order Lagrange finite element spaces over simplicial meshes. Unlike conventional black-box neural surrogates, the proposed network architecture is fully dictated by finite element discretization rules: local computations stem from simplex geometry and barycentric coordinate transformations, while global consistency is enforced through shared degrees of freedom. For linear Lagrange elements, local basis functions are directly implemented via affine barycentric layers, and high-order polynomial bases are explicitly decomposed into barycentric product compositions realized by specially designed $\mathrm{ReLU}^p$ modules. Equipped with element indicator branches and multiplication units, these modular local components are globally assembled into a sparsely connected neural network whose function space coincides exactly with the target finite element space, thereby inheriting the complete classical finite element approximation theory. By prescribing customized backward differentiation rules for piecewise activations, function values and their spatial gradients can be simultaneously extracted via automatic differentiation within a unified computational graph, eliminating the separate gradient calculation subroutines required in standard finite element implementations. Numerical experiments verify the accuracy of the neural network representation of Lagrange finite elements. Furthermore, by virtue of the intrinsic mesh-free nature of this neural network representation, finite element functions can be interpolated between non-matching meshes, and the proposed scheme can be applied to adaptive finite element methods for solving parabolic partial differential equations. An open-source code implementation of the proposed architecture is made publicly available.
Physics-informed neural networks (PINNs) represent a growing frontier in using artificial intelligence to solve partial differential equations (PDEs). Automatic differentiation (AD) plays a central role in this paradigm, which is mesh-free and replaces traditional iterative solvers with gradient-based optimization in continuous space. However, the inherent limitations of AD, particularly in handling higher-order derivatives and discontinuous solutions, pose significant challenges for complex problems. This has motivated a growing number of researchers to explore discretization-based constraints as an alternative path. Yet, the respective applicability of these two paradigms remains largely unexplored. In this work, we conduct systematic experiments across a wide spectrum of problems, from simple linear Poisson to high-Mach hypersonic flows with strong discontinuities. Through a rigorous decomposition of approximation, optimization, and truncation errors, we systematically elucidate the fundamental trade-offs and error-governing mechanisms of both paradigms, as well as two representative network architectures: multi-layer perceptron (MLP) and graph neural network (GNN). Our results reveal a consistent trend: as nonlinearity strengthens, the accuracy advantage of discretization-based constraints becomes increasingly pronounced, with smaller optimization errors compensating for the truncation errors. Moreover, the more complex the nonlinearity and boundary conditions, the greater the advantage of GNN over MLP. These insights offer a robust practical guideline for configuring neural PDE solvers in demanding engineering applications. Our source data and code are available at https://github.com/guoxing0809/neuropde_analysis.
We present a physics-informed, data-free neural solver for partial differential equations, built on a graph neural network architecture that utilises message passing. By relying on the weak form of the problem, we use gradients of finite-element shape functions (which are therefore polynomials) rather than automatic differentiation operators to compute the residuals of the equation from the displacements predicted by the network itself. Our approach generalises to previously unseen load cases and geometries, achieving easily convergence errors in the residuals of less than 1% and being capable of scaling up to models of considerable size and arbitrary geometries. To ensure compliance with the laws of physics and provide guarantees regarding the inference, it is possible to use the residual itself as an error indicator for the inference, and thus perform a refinement at the testing stage if the residual tolerance set in advance by the user is not met. Examples are provided to demonstrate the performance of the proposed method. This results in a method that avoids the costly process of obtaining, curating and storing high-fidelity synthetic data for training the neural network. Whilst this is not unique to our method, it is the first time it has been combined with a geometric machine learning technique capable of providing the necessary geometric bias to overcome the well-known difficulties of physics-informed neural networks.
We refine the OrbNet model to accurately predict energy, forces, and other response properties for molecules using a graph neural-network architecture based on features from low-cost approximated quantum operators in the symmetry-adapted atomic orbital basis. The model is end-to-end differentiable due to the derivation of analytic gradients for all electronic structure terms, and is shown to be transferable across chemical space due to the use of domain-specific features. The learning efficiency is improved by incorporating physically motivated constraints on the electronic structure through multi-task learning. The model outperforms existing methods on energy prediction tasks for the QM9 dataset and for molecular geometry optimizations on conformer datasets, at a computational cost that is thousand-fold or more reduced compared to conventional quantum-chemistry calculations (such as density functional theory) that offer similar accuracy.