部分观测流动的记忆效应、动力学闭合与长时预测:历史信息改善后续演化预测的条件、补充信息与有效作用时长
严格投影形式与记忆核的可测性:Mori–Zwanzig/Nakajima–Zwanzig 框架、广义朗之万与主方程、记忆项退化为马尔可夫的判据
这一组提供记忆项的精确定义与判据,是回答"历史补充了什么信息"的前提。电荷输运两篇在精确 GLE/GME 框架下给出记忆核与马尔可夫近似的定量对照,可作为"有效记忆时长"的可测参照;MZ 随机建模与图示算例说明未解析自由度如何被分裂为记忆核、噪声与正交项,并给出慢流形近似良好、记忆项可略去而噪声主导的条件;快振荡随机整流一文给出慢流形近似失效时记忆项的生成机制。它们是区分"严格投影闭合"与"受其启发的模型"的判据来源。
- Theoretical study of charge carrier transport in organic molecular crystals using the Nakajima-Zwanzig-Mori generalized master equation.(Yaming Yan, Meng Xu, Yanying Liu, Q. Shi, 2019, Journal of Chemical Physics)
- Generalized master equation for charge transport in a molecular junction: Exact memory kernels and their high order expansion.(Xiao-Han Dan, Meng Xu, Yaming Yan, Qiang Shi, 2022, Journal of Chemical Physics)
- The Mori-Zwanzig Formalism and Stochastic Modelling of Multiscale Dynamical Systems(F. Witteveen, 2016, No journal)
- A An Illustrative Example of Mori-Zwanzig Formalism(2020, No journal)
- Stochastic rectification of fast oscillations on slow manifold closures(M. Chekroun, Hong-Hu Liu, J. McWilliams, 2021, Proceedings of the National Academy of Sciences of the United States of America)
粗粒化实现与记忆注入技术:广义多尺度坐标/Schur 补消元、粗网格方程中的记忆效应、辅助变量与随机未解析通量闭合
这组把抽象记忆核落成可计算的粗粒化流程,回答"记忆项能否被构造而非仅被假设"。广义多尺度坐标/Schur 补给出细尺度消元产生记忆核的显式途径与算子稀疏判据,并厘清其与 Mori–Zwanzig 记忆核的对应;粗网格方程中的记忆效应讨论其数值实现;聚合物隐式溶剂模型用辅助变量把非马尔可夫动力学等价改写为扩展态空间中的马尔可夫动力学,是"向闭合模型注入记忆"的标准技术路径;MZ 降阶建模把框架用于流体 ROM,涡动力学粗粒化模拟把路线带入流体语境;未解析涡通量的随机建模提供统计层面的闭合替代。
- Building a Coarse-Grained Model Based on the Mori-Zwanzig Formalism(Hee Sun Lee, Surl-Hee Ahn, Eric F Darve, 2015, MRS Proceedings)
- How do we choose the right generalized multiscale coordinates for transport closure?(J. Segura, 2026, Transport Phenomena)
- Incorporation of memory effects in coarse-grained modeling via the Mori-Zwanzig formalism.(Zhen Li, X. Bian, Xian-Tao Li, G. Karniadakis, 2015, Journal of Chemical Physics)
- Implicit-solvent coarse-grained modeling for polymer solutions via Mori-Zwanzig formalism.(Shu Wang, Zhen Li, Wenxiao Pan, 2019, Soft Matter)
- Coarse-grained simulations of vortex dynamics and transition in complex high-Re flows(F. Grinstein, 2011, No journal)
- Mori-Zwanzig formalism based reduced-order modeling for decision-making in marine autonomy(Mengxue Hou, 2022, No journal)
- Stochastic modelling of unresolved eddy fluxes(M. Zidikheri, J. Frederiksen, 2010, Geophysical & Astrophysical Fluid Dynamics)
受投影启发的模态截断闭合与残差记忆回归:LSTM 记忆积分、autoencoder–Galerkin 子空间外误差、能量谱过稳定化与截断可预测时限
这组把 Mori–Zwanzig 记忆积分当作建模目标而非严格投影结果,构成"受理论启发"一类,并在一批工作中提供"仅当前状态 vs. 增加历史"的受控比较。RNN/LSTM 闭合以若干时间步的模态系数为输入回归记忆积分并检验超出训练区间的推进;autoencoder–Galerkin、分区 ROM 与"突破 Kolmogorov 屏障"讨论子空间外截断误差、闭合引起的能量谱过稳定化;MZ 记忆核被并入 Koopman 线性模型与部分可观测下的信念动力学;随机模态消减在同一低维系统中比较线性反馈、随机参数化与噪声膨胀的短期/长期效果并预测隐藏极端事件;Boussinesq→Lorenz 截断分析给出闭合假设与可预测时限 T~1/λ1 的分析范例,把模态截断与预测时限直接联系起来。
- Recurrent neural network closure of parametric POD-Galerkin reduced-order models based on the Mori-Zwanzig formalism(Qian Wang, N. Ripamonti, J. Hesthaven, 2020, Journal of Computational Physics)
- Hybrid AutoEncoder/Galerkin approach for nonlinear reduced order modelling(Nicolas Lepage, Samir Beneddine, C. Fiorini, I. Mortazavi, D. Sipp, N. Thome, 2025, Computers & Fluids)
- Breaking the Kolmogorov Barrier in Model Reduction of Fluid Flows(Shady E. Ahmed, O. San, 2020, Fluids)
- Mori-zwanzig approach for belief abstraction with application to belief space planning(Mengxue Hou, Tony X. Lin, Enlu Zhou, Fumin Zhang, 2024, Autonomous Robots)
- Predicting observed and hidden extreme events in complex nonlinear dynamical systems with partial observations and short training time series.(Nan Chen, A. Majda, 2020, Chaos)
- Boussinesq Navier Stokes to A Modification on the Lorenz-63 Model(Kimm Ryan E. Roldan, Qatryn MieR E. Villanueva, 2026, International journal of research and innovation in social science)
- Memory-dependent dynamic modeling of cable-driven soft robots using the Mori-Zwanzig Koopman operator(Larry Catalasan, Tianyi He, 2026, Nonlinear dynamics)
时延嵌入与 Hankel/Koopman 增广坐标:延迟深度与可观秩、最小实现、Strouhal 锚定,及闭环多步推进的误差放大
组内共享平移嵌入内核,以坐标增广而非显式记忆核表达历史。时延/锁相 DMD 处理仅有局部测点与含差分快照的情形;输入感知 HAVOK、最小实现时延 Koopman 与 Subspace DMD 把延迟长度与可观秩、最小实现联系起来并可接 Kalman 滤波做稀疏含噪实时估计;时空增广 STDMD 同时提取频率与波数;神经振荡模态恢复一文用 Hankel 矩阵分解给出延迟嵌入提升可观秩、扰动提升轨迹秩的互补约束与"有效记忆深度对应可估计模态数"的明确证据;矩形柱 Hankel-ROM 显式扫描嵌入维数;钝尾体 HAVOK 把线性算子特征值与极低频振荡及不同 Strouhal 数挂钩,是"记忆时间能否锚定涡脱落周期"最直接的可检验案例;扫描射流作动器与航天器姿态的 Hankel-DMD 研究给出多周期预测增益与闭环多步误差放大上千倍的量化反例。
- A snapshot difference and amplitude optimization based dynamic mode decomposition method for reconstruction and prediction of flow fields(Xuhui Luo, Yanhao Chen, Yan-Song He, Zhi-Fei Zhang, 2025, The Physics of Fluids)
- A data-driven analysis of near-wake flow dynamics for a blunt-based axisymmetric body under low-speed conditions(Khiem Van Pham, H. T. Tran, Anh Dinh Le, Thanh Van Do Thi, 2025, The Physics of Fluids)
- Wind field and pressure prediction of a ground-mounted rectangular cylinder based on Hankel-inspired reduced order model and temporal convolutional network-transformer(Wei Lu, Zhi-Bo Lu, Gang Hu, Hong-Fu Zhang, Jie Chen, Jia-Wei He, 2026, The Physics of Fluids)
- Time-delay-locked dynamic mode decomposition for extracting flow patterns of compressor in experiment(Jiaqi Wang, Jin Chen, Mo-Ru Song, Guangming Dong, 2022, Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering)
- Data-Driven Koopman Based System Identification for Partially Observed Dynamical Systems with Input and Disturbance(Patinya Ketthong, Jirayu Samkunta, N. Mai, M. Kamal, I. Murakami, Kou Yamada, 2024, The Scientist)
- Minimal realization time-delay Koopman analysis for stochastic dynamical system identification(Bi-Qi Chen, Ying Wang, 2026, Advances in Structural Engineering)
- ON AUGMENTED SPATIO-TEMPORAL DYNAMIC MODE DECOMPOSITION(G. Nedzhibov, 2024, MATTEX 2024, CONFERENCE PROCEEDINGS, Volume 1)
- Recovering Hidden Neural Oscillatory Modes from Partial Observations through Delay Embedding and Perturbation(Raku Miyoshi, Mikito Ogino, Daiki Sekizawa, Masafumi Oizumi, 2026, bioRxiv)
- Reduced-Order Modeling of Sweeping Jet Actuators Using Eigenvalue-Sorted Dynamic Mode Decomposition(Shafi Al Salman Romeo, Mobashera Alam, Kursat Kara, 2026, Aerospace)
部分观测系统的可辨识性与归因框架:稀疏动力学辨识、代数相空间补全与部分观测数据驱动建模
这组界定"历史信息是否必要"的问题边界与比较设计。压缩感知稀疏动力学识别从稀疏测量反演控制方程;代数重构明确指出部分观测须先补全相空间再辨识方程,把"用时间序列恢复未观测变量"作为独立步骤;Lorenz-63 仅观测两分量的工作证明学习到的隐变量与观测分量各阶导数相关,说明历史补充的正是缺失分量的动力学信息;部分观测生物系统的数据驱动建模与机器学习降阶建模综述提供总体框架。它们共同提醒:单凭预测误差下降无法区分"隐藏当前状态估计"与"未解析自由度闭合"。
- Modelling Sparse Dynamical Systems with Compressed Predictive State Representations(William L. Hamilton, M. M. Fard, Joelle Pineau, 2013, International Conference on Machine Learning)
- Algebraic Method for the Reconstruction of Partially Observed Nonlinear Systems Using Differential and Integral Embedding(A. Karimov, E. Nepomuceno, A. Tutueva, D. Butusov, 2020, Mathematics)
- Data-driven reconstruction of partially observed dynamical systems(P. Tandéo, P. Ailliot, F. Sévellec, 2023, Nonlinear Processes in Geophysics)
- Data-Driven Modeling of Partially Observed Biological Systems(Weiwei Su, Ching-Shan Chou, D. Xiu, 2024, Communication on Applied Mathematics and Computation)
- A reduced order modeling-based machine learning approach for wind turbine wake flow estimation from sparse sensor measurements(Zhao-Hui Luo, Long-Yan Wang, Jian Xu, Zilu Wang, Jianping Yuan, A. C. Tan, 2024, Energy)
历史用于估计隐藏的当前状态:信息状态理论的最优性条件、递归贝叶斯滤波与最优平滑
这组对应第一种作用机制,且预测误差下降几乎全部来自观测信息的在线注入而非模型自身闭合,因此是与"记忆闭合"做归因对照的必需对照组。信息状态两篇严格证明:由过去若干观测与控制输入构成的信息状态可把部分观测非线性问题等价转换为全观测问题,并给出求解确定最优解的条件,是"需要多长历史才足够"的理论边界;ENKF 流场状态估计器、流场特征估计、稀疏流场状态估计、集成滤波方案与 Belief Flow Filter(极端单传感器、生成式后验跟踪)代表贝叶斯同化路线;非线性最优平滑与最小阶控制导向流场估计提供算法基线与压缩瓶颈维度讨论。
- Information-state based Approach to the Optimal Output Feedback Control of Nonlinear Systems(Raman Goyal, Mohamed Naveed Gul Mohamed, Ran Wang, Aayushman Sharma, S. Chakravorty, 2025, American Control Conference)
- Information-State-Based Reinforcement Learning for the Control of Partially Observed Nonlinear Systems(Raman Goyal, Mohamed Naveed Gul Mohamed, Ran Wang, Aayushman Sharma, S. Chakravorty, 2025, IEEE Transactions on Neural Networks and Learning Systems)
- An EnKF-based Flow State Estimator for Aerodynamic Problems(D. Silva, Andre Fernando de Castro, 2019, No journal)
- Estimation of Flow Features in the Wake of a Circular Cylinder Using Artificial Neural Network(B. Şahin, Cetin Canpolat, M. Bilgili, 2024, The Arabian journal for science and engineering)
- Fluid flow state estimation from sparse sensor measurements using convolutional neural network(Taichi Nakamura, Kai Fukami, K. Fukagata, 2021, The Proceedings of Mechanical Engineering Congress Japan)
- An ensemble filter estimation scheme for Lagrangian trajectory reconstruction(Yin Yang, D. Heitz, É. Mémin, 2018, No journal)
- Bayesian Filtering in Physical Systems via Test-time Trained Flow Matching(Rui-Qi Feng, Chong-Yi Wang, Tao Zhang, Tai-Lin Wu, 2026, arXiv.org)
- Efficient nonlinear optimal smoothing and sampling algorithms for complex turbulent nonlinear dynamical systems with partial observations(N. Chen, A. Majda, 2020, Journal of Computational Physics)
- Least-order representation of control-oriented flow estimation exemplified for the fluidic pinball(Song-Qi Li, Wen-Peng Li, B. R. Noack, 2022, Journal of Physics, Conference Series)
- Energy networks for state estimation with random sensors using sparse labels(Y. Kumar, S. Chakraborty, 2022, arXiv.org)
稀疏传感全场重构与传感器/模态子空间的可观测性设计
在不做递归同化的前提下,用少量传感器读数一次性重构全场,并回答"传感器放在哪、用哪个子空间"才能最好地观测被隐藏的自由度。SHRED/UQ-SHRED 用超稀疏传感器历史重建全时空场并给出不确定性量化与设置消融;自编码器+交叉注意力、Voronoi/Senseiver 对比与 AE+ROM 给出稀疏含噪重构的精度与效率基准;壁面压力恢复涡结构与聚变诊断代理模型提供跨领域验证;SPOD vs POD 的 LSTM 估计器在同一双柱 flip-flop 流动中比较模态子空间,是子空间选择的受控实验;POD 多时延 LSE 把多时刻历史直接写入观测模型;风电场与贝叶斯传感器选择把可观测性作为优化目标。
- UQ-SHRED: uncertainty quantification of shallow recurrent decoder networks for sparse sensing via engression(Mars Liyao Gao, Y. Bao, A. Rude, Xing-Yu Zhang, Xin-Wei Shen, J. Kutz, 2026, arXiv.org)
- Flow reconstruction from sparse sensors based on reduced-order autoencoder state estimation(A. Tan, 2023, The Physics of Fluids)
- Machine-Learning-Based Reconstruction of Turbulent Vortices From Sparse Pressure Sensors in a Pump Sump(Kai Fukami, B. An, M. Nohmi, Masashi Obuchi, K. Taira, 2022, Journal of Fluids Engineering)
- Surrogate models for nuclear fusion with parametric Shallow Recurrent Decoder Networks: applications to magnetohydrodynamics(M. Lo Verso, C. Introini, E. Cervi, L. Savoldi, J. Kutz, A. Cammi, 2026, arXiv.org)
- Reconstructing flow field from sparse sensor data: A deep learning framework combining autoencoder and cross-attention mechanism(X. Kong, Qipei Fan, Jin-Zhao Li, Wei-Wei He, 2025, The Physics of Fluids)
- Sparse sensor-based flow estimation with spectral proper orthogonal decomposition(H. Gambassi, P. Ziadé, C. Morton, 2022, AIP Advances)
- POD-BASED ESTIMATION OF THE FLOW FIELD FROM FREE-SURFACE VELOCITY IN THE BACKWARD-FACING STEP(Thien Nguyen, T. X. Dinh, J. Wells, P. Mokhasi, D. Rempfer, 2011, Proceeding of Seventh International Symposium on Turbulence and Shear Flow Phenomena)
- Sparse-Sensor Placement for Wind Farm Control(J. Annoni, T. Taylor, C. Bay, K. Johnson, L. Pao, P. Fleming, K. Dykes, 2018, Journal of Physics, Conference Series)
- Optimization of sensor selection and Bayesian estimation for compressed sensing dynamic mode decomposition(Wen-Wei Jiang, Hua-Yu Liu, Jin‐Xing Ding, Chen-Hao Tan, Yun-Tao Zhou, Kai Yang, Xiaowei Gao, 2026, Acta Mechanica Sinica)
滤波尺度未解析通量:SFS 记忆参数化、神经亚格子模型与涡黏/动态/隐式 LES 等无记忆闭合下界
这组对应"历史描述未解析自由度对保留变量后续演化影响"这一作用在物理闭合层面的实现,同时给出"不引入历史、仅调整当前状态的通量闭合能做到多好"的下界基准。iGLASS 以积分完整的 SFS 湍动能与位能方程显式保留记忆,并与涡扩散、动态重构在同一算例对比,是"有记忆 vs. 无记忆闭合"最直接的受控比较;神经亚格子模型在槽道流中检验局部映射闭合的 a priori/a posteriori 差异与数值不稳定;a priori/a posteriori 分析把亚滤波应力分解为 Leonard/Cross/Reynolds 并逐项给出能量与拟涡能收支,是识别"稳定但过度耗散"的可操作框架;动态 SGS 模型、二维涡黏理论与隐式 LES 代表以数值耗散承担闭合的经典路线。
- An Implicit Algebraic Turbulence Closure Scheme for Atmospheric Boundary Layer Simulation(Xiao-Ming Shi, Rica Mae Enriquez, R. Street, G. Bryan, F. Chow, 2019, Journal of the Atmospheric Sciences)
- Large-eddy simulation subgrid modelling using neural networks(R. Stoffer, C. V. van Leeuwen, Damian Podareanu, V. Codreanu, M. Veerman, C. V. van Heerwaarden, 2020, No journal)
- A priori and a posteriori analysis in Large eddy simulation of the two-dimensional decaying turbulence using Explicit filtering approach(P. Perezhogin, A. Glazunov, 2021, No journal)
- A dynamic subgrid-scale model for LES based on the Mori-Zwanzig formalism(E. Parish, K. Duraisamy, 2016, No journal)
- Theory of Eddy Viscosity for Two-Dimentional Inviscid Barotropic Fluid(Takahiro, No journal)
- Computing Turbulent Flow Dynamics With Implicit Large Eddy Simulation(F. Grinstein, D. Drikakis, 2007, Journal of Fluids Engineering-transactions of The Asme)
未解析自由度的物理记忆时间尺度:湍流寿命与拉格朗日输运记忆、准地转与近壁亚格子参数化
这组把"有效记忆多长"锚定到可独立测量的结构时间尺度。湍流有限生成—耗散速率导致的边界应力记忆、横向腔体中由相干结构控制的非局域输运记忆(以分数阶导数上采样建模)与拉格朗日/欧拉视角的对照,把记忆与寿命、停留时间、相关衰减时间直接挂钩;准地转流亚格子参数化与柱体串联/横流耦合预测中的记忆与干扰效应则说明记忆在复杂工况下的表现。它们为"核衰减时间—涡脱落周期—采样间隔"之间能否建立可验证联系提供物理依据。
- The Influence of Turbulence Memory on Idealized Tornado Simulations(Aaron Wang, Ying-Kang Pan, P. Markowski, 2020, Monthly Weather Review)
- Lagrangian and Eulerian perspectives of turbulent transport mechanisms in a lateral cavity(M. Barros, C. Escauriaza, 2024, Journal of Fluid Mechanics)
- Dynamical Subgrid-scale Parameterizations for Quasigeostrophic Flows using Direct Numerical Simulations(M. Zidikheri, 2007, No journal)
- Prediction of Coupled In-Line and Cross-Flow Vortex-Induced Vibration of a Near-Wall Cylinder with Physical Impact(Meng-Meng Tao, Xu Sun, Peiyi Han, 2026, China Ocean engineering)
神经算子与序列模型的自主多步预测:路径签名与神经粗糙微分方程、通道辅助 FNO、物理引导网络及相位/耗散失真诊断
这组属"一般历史编码"而非严格投影闭合,关注连续时间历史表达与长时自回归稳定性。神经 RDE 以 log-signature 紧凑编码路径历史逼近路径依赖解;通道辅助 FNO 通过通道注入与残差增强傅里叶层稳定长时自回归尾流预测;物理引导神经网络用于时空场演化;PhyLSNN(卷积 LSTM+最小二乘 NS 残差)指出常规 MLP-PINN 会造成相位漂移与过度平滑,是长期推进失真的诊断证据;振荡状态空间 PINN 以线性振子作模态时间演化先验,显著降低对序列长度与分辨率的记忆开销,提供"结构化动力先验 vs. 通用序列模型"的比较视角;DMD+BiLSTM 浮式风机尾流预测与欧拉视角 LSTM 风湍流预报给出工程场景(含固定点稀疏观测)的可预测时效参照。
- A Neural RDE-based model for solving path-dependent PDEs(Bowen Fang, Hao Ni, Yue Wu, 2023, arXiv.org)
- Channel-assisted Fourier neural operator for high-fidelity far-wake prediction of a wind turbine under yawed conditions(Eunchan Lee, Sang Lee, 2025, The Physics of Fluids)
- Physics guided neural networks for spatio-temporal super-resolution of turbulent flows(Tian-Shu Bao, Sheng-Yu Chen, Taylor T. Johnson, P. Givi, S. Sammak, Xiao-Wei Jia, 2022, Conference on Uncertainty in Artificial Intelligence)
- Physics-informed spatiotemporal neural network for unsteady propeller wake prediction via least squares finite-difference framework(Zheming Tong, Ruizhe Chen, 2026, The Physics of Fluids)
- Oscillatory State-Space Models as Inductive Biases for Physics-Informed Neural PDE Solvers(A. Chandra, Taniya Kapoor, 2026, arXiv.org)
- Wake prediction of a fully coupled floating wind turbine using dynamic mode decomposition and bidirectional long short-term memory method(Xiaodi Wu, Jiaqi Li, Wenhao Lu, Jia-Hao Chen, 2025, The Physics of Fluids)
- Eulerian-DL Forecasting of Wind Turbulence for Wind Turbines: Enabling Smart Sustainable Energy(R. Hassanian, M. Riedel, 2024, 2024 International Conference on Engineering and Emerging Technologies (ICEET))
低阶模态与振幅方程模型:模态系数序列预测、回声状态网络超参数扫描、相幅模型失效机制与 SVD 降阶框架
这组用有限模态或低阶振幅方程承载历史而不追求投影严格性。圆柱阵列尾流以 POD 时间系数驱动回声状态网络,并系统扫描输入尺度、谱半径、泄漏率与流动间歇性/速度的关系,可直接支撑"记忆时间尺度随流动条件变化"的讨论;矩形柱三维尾流以 POD–LSTM 在同一算例中比较不同模态对 A/B 不稳定性的预测能力;tanh 型振幅—频率方程说明低阶模型在相位/频率随时间变化的转变过程中的失效机制;降阶建模与 SVD 方法综述提供整体框架与基线。它们与 Hankel 路线、MZ 路线共同构成"一般历史编码 vs. 坐标增广 vs. 严格投影"的三方对照。
- On the prediction of the turbulent flow behind cylinder arrays via echo state networks(M. Sharifi Ghazijahani, C. Cierpka, 2024, Machine Learning: Science and Technology)
- Three-dimensional wake transition of rectangular cylinders and temporal prediction of flow patterns based on a machine learning algorithm(A. Mashhadi, A. Sohankar, M. Moradmand, 2024, The Physics of Fluids)
- Tanh-like models for analysis and prediction of time-dependent flow around a circular cylinder at low Reynolds numbers(Li-Bao Wang, Min Xu, Bo-Da Zheng, Xiao-Min An, 2023, The Physics of Fluids)
- Reduced order modeling of turbulent flows using statistical coarse-graining(E. Parish, K. Duraisamy, 2016, No journal)
- DATA-DRIVEN METHODS BASED ON SVD AND DMD AND DEEP LEARNING IN FLUID DYNAMICS(A. Mathematics, 2020, No journal)
长时自主推进的验证实践:瞬时预测窗口、能谱与统计保真、"稳定但过度耗散"的识别及跨雷诺数/跨工况外推
这组专门审查初始化后仅用模型自身输出多步推进(而非持续接收新测量或同化)的可信度,并给出正反两面的诊断指标。生成式神经算子以单条轨迹少样本训练,在均匀各向同性湍流上报告五个涡翻转时间的可信预报,用能谱误差与加速比量化长期统计保真;Hankel-DMD 闭环推进研究给出一步精度极佳但自主多步误差放大上千倍的量化反例(Rollout Amplification Factor),是"时延嵌入并不保证长时可信"的关键证据;稀疏特征映射 Markov 模型在同一流动内以相同特征维度比较 POD 与稀疏高斯核映射的长期预测,显示基失相关时 POD 失效、核映射缺乏简约性时重构失真,并跨越极限环、瞬态到持续演化浮力混合流与不同雷诺数检验适用边界,可用于识别"看似稳定实则过度平滑/耗散"的失败模式。
- Learning turbulent flows with generative models for super resolution and sparse flow reconstruction(V. Oommen, Siavash Khodakarami, Aniruddha Bora, Zhi-Cheng Wang, G. Karniadakis, 2026, Nature Communications)
- Beyond One-Step Accuracy: Closed-Loop Rollout Reveals a Validation Gap in Hankel–DMD/Koopman Models for Gravity-Gradient-Stabilized CubeSat Attitude Dynamics(Aswin Karkadakattil, 2026, INCAS Bulletin)
- Sparse feature map-based Markov models for nonlinear fluid flows(Balaji Jayaraman, Chen Lu, Joshua Whitman, G. Chowdhary, 2019, Computers & Fluids)
可验证的时间尺度与基准数据:相关函数衰减形式、Lyapunov 时间与积分时间尺度、涡脱落相干标定及连续时间尾流数据
这组为"有效记忆持续多久"提供独立可测标尺与受控比较的数据基础。相关函数衰减形式给出以自相关衰减时间定义记忆核的经典依据;湍流 Couette 最小流单元的扰动演化给出短时指数增长(主导 Lyapunov 指数)→ 长达数十个积分时间尺度的线性慢增长→ 饱和的完整图景,明确指出失相关所需时间跨度,是评价长时推进与统计可信度的自然标尺;圆柱尾流壁面压力与近尾流速度的相干条件统计实验,用最少压力传感器把低频大尺度涡脱落与表面压力脉动、升阻力估计联系起来,提供把模态/记忆时间锚定 Strouhal 周期的实验标定范式;亚临界圆柱尾流 PIV 数据集(Re=240–2520,约三个涡脱落周期的连续时间分块记录)观测条件明确,可直接构造完整参考态与稀疏/粗化状态的对照,用于同时检验隐藏状态估计、记忆闭合与长期统计三类结论。
- Decay Forms of the Time Correlation Functions for Turbulence and Chaos(H. Mori, M. Okamura, 2012, Progress of Theoretical Physics)
- Perturbation growth over self-sustaining process in wall turbulence beyond the Lyapunov time(Pablo Egerique-de-la-Concha, Yongyun Hwang, 2026, Journal of Fluid Mechanics)
- Experimental study of coherent wall-pressure fluctuations associated with the turbulent near wake of a cylinder(R. Hernández, 2022, The Physics of Fluids)
- Experimental dataset of sub-critical cylinder wake velocity fields(P. Renn, Emily H. Palmer, M. Gharib, 2026, Scientific Data)
合并后共 13 个并列分组,覆盖全部 81 篇唯一文献,按"理论判据 → 记忆的建模表达 → 隐藏状态估计 → 物理闭合与时间尺度 → 长时推进验证 → 基准数据"的逻辑排列。第 1–2 组给出记忆项的精确定义、可测衰减时间与粗粒化构造路径(GLE/GME 精确核、Schur/广义多尺度坐标、辅助变量、随机未解析通量),是区分"严格投影闭合"与"受其启发模型"的判据来源;第 3 组是受 MZ 启发、以数据驱动逼近记忆积分与残差闭合的模态/残差模型,天然给出"仅当前状态 vs. 加历史"的受控比较;第 4 组以时延嵌入/Hankel–Koopman 增广坐标显式编码历史,并把延迟深度与可观秩、最小实现和 Strouhal 周期挂钩,同时提供闭环多步误差放大的量化反例;第 5–7 组构成"历史用于估计隐藏的当前状态"这一支:可辨识性与归因框架、信息状态理论与递归滤波/同化、稀疏重构与传感器—子空间可观测性设计,它们是与"记忆闭合"做归因对照的必需对照组;第 8–9 组构成"历史用于闭合未解析自由度"这一支:滤波尺度 SFS 通量记忆参数化与神经亚格子模型(并含涡黏/动态模型/隐式 LES 等无记忆下界基线),以及湍流寿命、拉格朗日输运、准地转与近壁参数化所给出的物理记忆时间尺度;第 10–11 组是"一般历史编码":神经 RDE/路径签名、通道辅助 FNO、物理引导与振荡状态空间模型、DMD+BiLSTM/LSTM 工程预报,以及 POD/LSTM、回声状态网络与 tanh 振幅—频率方程等低阶模态模型,其中包含相位漂移、过度平滑、频率转变失效等长时诊断证据;第 12 组专门审查初始化后仅用模型自身输出多步推进的验证实践(瞬时窗口、能谱保真、稳定但过度耗散、Rollout 放大因子、跨雷诺数与跨工况外推);第 13 组提供可独立测量的时间尺度(相关衰减、Lyapunov 时间、积分时间尺度、涡脱落相干标定)与观测条件明确、具连续时间记录的基准数据(亚临界圆柱尾流 PIV 等)。流动覆盖圆柱/矩形柱/圆柱阵列/双柱 flip-flop 尾流、钝尾体近尾流、浮式风机尾流、均匀各向同性与二维衰减湍流、槽道与 Couette 壁湍流、后台阶与扫描射流分离剪切流、准地转与浮力混合流及 Lorenz-63 等低维范例。整体看,文献足以支撑"隐藏状态估计 vs. 记忆闭合"两种作用的归因对照,但同一流动内同时做完整参考态/部分观测态受控比较、并把有效记忆时间与涡脱落周期或相关衰减时间做成可验证定量标定的研究仍明显稀缺。
这一组提供记忆项的精确定义与判据,是回答"历史补充了什么信息"的前提。电荷输运两篇在精确 GLE/GME 框架下给出记忆核与马尔可夫近似的定量对照,可作为"有效记忆时长"的可测参照;MZ 随机建模与图示算例说明未解析自由度如何被分裂为记忆核、噪声与正交项,并给出慢流形近似良好、记忆项可略去而噪声主导的条件;快振荡随机整流一文给出慢流形近似失效时记忆项的生成机制。它们是区分"严格投影闭合"与"受其启发的模型"的判据来源。相关文献:Yaming Yan et. al, 2019 等 5 篇文献
这组把抽象记忆核落成可计算的粗粒化流程,回答"记忆项能否被构造而非仅被假设"。广义多尺度坐标/Schur 补给出细尺度消元产生记忆核的显式途径与算子稀疏判据,并厘清其与 Mori–Zwanzig 记忆核的对应;粗网格方程中的记忆效应讨论其数值实现;聚合物隐式溶剂模型用辅助变量把非马尔可夫动力学等价改写为扩展态空间中的马尔可夫动力学,是"向闭合模型注入记忆"的标准技术路径;MZ 降阶建模把框架用于流体 ROM,涡动力学粗粒化模拟把路线带入流体语境;未解析涡通量的随机建模提供统计层面的闭合替代。相关文献:Hee Sun Lee et. al, 2015 等 7 篇文献
这组把 Mori–Zwanzig 记忆积分当作建模目标而非严格投影结果,构成"受理论启发"一类,并在一批工作中提供"仅当前状态 vs. 增加历史"的受控比较。RNN/LSTM 闭合以若干时间步的模态系数为输入回归记忆积分并检验超出训练区间的推进;autoencoder–Galerkin、分区 ROM 与"突破 Kolmogorov 屏障"讨论子空间外截断误差、闭合引起的能量谱过稳定化;MZ 记忆核被并入 Koopman 线性模型与部分可观测下的信念动力学;随机模态消减在同一低维系统中比较线性反馈、随机参数化与噪声膨胀的短期/长期效果并预测隐藏极端事件;Boussinesq→Lorenz 截断分析给出闭合假设与可预测时限 T~1/λ1 的分析范例,把模态截断与预测时限直接联系起来。相关文献:Qian Wang et. al, 2020 等 7 篇文献
组内共享平移嵌入内核,以坐标增广而非显式记忆核表达历史。时延/锁相 DMD 处理仅有局部测点与含差分快照的情形;输入感知 HAVOK、最小实现时延 Koopman 与 Subspace DMD 把延迟长度与可观秩、最小实现联系起来并可接 Kalman 滤波做稀疏含噪实时估计;时空增广 STDMD 同时提取频率与波数;神经振荡模态恢复一文用 Hankel 矩阵分解给出延迟嵌入提升可观秩、扰动提升轨迹秩的互补约束与"有效记忆深度对应可估计模态数"的明确证据;矩形柱 Hankel-ROM 显式扫描嵌入维数;钝尾体 HAVOK 把线性算子特征值与极低频振荡及不同 Strouhal 数挂钩,是"记忆时间能否锚定涡脱落周期"最直接的可检验案例;扫描射流作动器与航天器姿态的 Hankel-DMD 研究给出多周期预测增益与闭环多步误差放大上千倍的量化反例。相关文献:Xuhui Luo et. al, 2025 等 9 篇文献
这组界定"历史信息是否必要"的问题边界与比较设计。压缩感知稀疏动力学识别从稀疏测量反演控制方程;代数重构明确指出部分观测须先补全相空间再辨识方程,把"用时间序列恢复未观测变量"作为独立步骤;Lorenz-63 仅观测两分量的工作证明学习到的隐变量与观测分量各阶导数相关,说明历史补充的正是缺失分量的动力学信息;部分观测生物系统的数据驱动建模与机器学习降阶建模综述提供总体框架。它们共同提醒:单凭预测误差下降无法区分"隐藏当前状态估计"与"未解析自由度闭合"。相关文献:William L. Hamilton et. al, 2013 等 5 篇文献
这组对应第一种作用机制,且预测误差下降几乎全部来自观测信息的在线注入而非模型自身闭合,因此是与"记忆闭合"做归因对照的必需对照组。信息状态两篇严格证明:由过去若干观测与控制输入构成的信息状态可把部分观测非线性问题等价转换为全观测问题,并给出求解确定最优解的条件,是"需要多长历史才足够"的理论边界;ENKF 流场状态估计器、流场特征估计、稀疏流场状态估计、集成滤波方案与 Belief Flow Filter(极端单传感器、生成式后验跟踪)代表贝叶斯同化路线;非线性最优平滑与最小阶控制导向流场估计提供算法基线与压缩瓶颈维度讨论。相关文献:Raman Goyal et. al, 2025 等 10 篇文献
在不做递归同化的前提下,用少量传感器读数一次性重构全场,并回答"传感器放在哪、用哪个子空间"才能最好地观测被隐藏的自由度。SHRED/UQ-SHRED 用超稀疏传感器历史重建全时空场并给出不确定性量化与设置消融;自编码器+交叉注意力、Voronoi/Senseiver 对比与 AE+ROM 给出稀疏含噪重构的精度与效率基准;壁面压力恢复涡结构与聚变诊断代理模型提供跨领域验证;SPOD vs POD 的 LSTM 估计器在同一双柱 flip-flop 流动中比较模态子空间,是子空间选择的受控实验;POD 多时延 LSE 把多时刻历史直接写入观测模型;风电场与贝叶斯传感器选择把可观测性作为优化目标。相关文献:Mars Liyao Gao et. al, 2026 等 9 篇文献
这组对应"历史描述未解析自由度对保留变量后续演化影响"这一作用在物理闭合层面的实现,同时给出"不引入历史、仅调整当前状态的通量闭合能做到多好"的下界基准。iGLASS 以积分完整的 SFS 湍动能与位能方程显式保留记忆,并与涡扩散、动态重构在同一算例对比,是"有记忆 vs. 无记忆闭合"最直接的受控比较;神经亚格子模型在槽道流中检验局部映射闭合的 a priori/a posteriori 差异与数值不稳定;a priori/a posteriori 分析把亚滤波应力分解为 Leonard/Cross/Reynolds 并逐项给出能量与拟涡能收支,是识别"稳定但过度耗散"的可操作框架;动态 SGS 模型、二维涡黏理论与隐式 LES 代表以数值耗散承担闭合的经典路线。相关文献:Xiao-Ming Shi et. al, 2019 等 6 篇文献
这组把"有效记忆多长"锚定到可独立测量的结构时间尺度。湍流有限生成—耗散速率导致的边界应力记忆、横向腔体中由相干结构控制的非局域输运记忆(以分数阶导数上采样建模)与拉格朗日/欧拉视角的对照,把记忆与寿命、停留时间、相关衰减时间直接挂钩;准地转流亚格子参数化与柱体串联/横流耦合预测中的记忆与干扰效应则说明记忆在复杂工况下的表现。它们为"核衰减时间—涡脱落周期—采样间隔"之间能否建立可验证联系提供物理依据。相关文献:Aaron Wang et. al, 2020 等 4 篇文献
这组属"一般历史编码"而非严格投影闭合,关注连续时间历史表达与长时自回归稳定性。神经 RDE 以 log-signature 紧凑编码路径历史逼近路径依赖解;通道辅助 FNO 通过通道注入与残差增强傅里叶层稳定长时自回归尾流预测;物理引导神经网络用于时空场演化;PhyLSNN(卷积 LSTM+最小二乘 NS 残差)指出常规 MLP-PINN 会造成相位漂移与过度平滑,是长期推进失真的诊断证据;振荡状态空间 PINN 以线性振子作模态时间演化先验,显著降低对序列长度与分辨率的记忆开销,提供"结构化动力先验 vs. 通用序列模型"的比较视角;DMD+BiLSTM 浮式风机尾流预测与欧拉视角 LSTM 风湍流预报给出工程场景(含固定点稀疏观测)的可预测时效参照。相关文献:Bowen Fang et. al, 2023 等 7 篇文献
这组用有限模态或低阶振幅方程承载历史而不追求投影严格性。圆柱阵列尾流以 POD 时间系数驱动回声状态网络,并系统扫描输入尺度、谱半径、泄漏率与流动间歇性/速度的关系,可直接支撑"记忆时间尺度随流动条件变化"的讨论;矩形柱三维尾流以 POD–LSTM 在同一算例中比较不同模态对 A/B 不稳定性的预测能力;tanh 型振幅—频率方程说明低阶模型在相位/频率随时间变化的转变过程中的失效机制;降阶建模与 SVD 方法综述提供整体框架与基线。它们与 Hankel 路线、MZ 路线共同构成"一般历史编码 vs. 坐标增广 vs. 严格投影"的三方对照。相关文献:M. Sharifi Ghazijahani et. al, 2024 等 5 篇文献
这组专门审查初始化后仅用模型自身输出多步推进(而非持续接收新测量或同化)的可信度,并给出正反两面的诊断指标。生成式神经算子以单条轨迹少样本训练,在均匀各向同性湍流上报告五个涡翻转时间的可信预报,用能谱误差与加速比量化长期统计保真;Hankel-DMD 闭环推进研究给出一步精度极佳但自主多步误差放大上千倍的量化反例(Rollout Amplification Factor),是"时延嵌入并不保证长时可信"的关键证据;稀疏特征映射 Markov 模型在同一流动内以相同特征维度比较 POD 与稀疏高斯核映射的长期预测,显示基失相关时 POD 失效、核映射缺乏简约性时重构失真,并跨越极限环、瞬态到持续演化浮力混合流与不同雷诺数检验适用边界,可用于识别"看似稳定实则过度平滑/耗散"的失败模式。相关文献:V. Oommen et. al, 2026 等 3 篇文献
这组为"有效记忆持续多久"提供独立可测标尺与受控比较的数据基础。相关函数衰减形式给出以自相关衰减时间定义记忆核的经典依据;湍流 Couette 最小流单元的扰动演化给出短时指数增长(主导 Lyapunov 指数)→ 长达数十个积分时间尺度的线性慢增长→ 饱和的完整图景,明确指出失相关所需时间跨度,是评价长时推进与统计可信度的自然标尺;圆柱尾流壁面压力与近尾流速度的相干条件统计实验,用最少压力传感器把低频大尺度涡脱落与表面压力脉动、升阻力估计联系起来,提供把模态/记忆时间锚定 Strouhal 周期的实验标定范式;亚临界圆柱尾流 PIV 数据集(Re=240–2520,约三个涡脱落周期的连续时间分块记录)观测条件明确,可直接构造完整参考态与稀疏/粗化状态的对照,用于同时检验隐藏状态估计、记忆闭合与长期统计三类结论。相关文献:H. Mori et. al, 2012 等 4 篇文献
总计 81 篇相关文献
Abstract Closure modeling based on the Mori-Zwanzig formalism has proven effective to improve the stability and accuracy of projection-based model order reduction. However, closure models are often expensive and infeasible for complex nonlinear systems. Towards efficient model reduction of general problems, this paper presents a recurrent neural network (RNN) closure of parametric POD-Galerkin reduced-order model. Based on the short time history of the reduced-order solutions, the RNN predicts the memory integral which represents the impact of the unresolved scales on the resolved scales. A conditioned long short term memory (LSTM) network is utilized as the regression model of the memory integral, in which the POD coefficients at a number of time steps are fed into the LSTM units, and the physical/geometrical parameters are fed into the initial hidden state of the LSTM. The reduced-order model is integrated in time using an implicit-explicit (IMEX) Runge-Kutta scheme, in which the memory term is integrated explicitly and the remaining right-hand-side term is integrated implicitly to improve the computational efficiency. Numerical results demonstrate that the RNN closure can significantly improve the accuracy and efficiency of the POD-Galerkin reduced-order model of nonlinear problems. The POD-Galerkin reduced-order model with the RNN closure is also shown to be capable of making accurate predictions, well beyond the time interval of the training data.
We propose a learning-based method to extract symbolic representations of the belief state and its dynamics in order to solve planning problems in a continuous-state partially observable Markov decision processes (POMDP) problem. While existing approaches typically parameterize the continuous-state POMDP into a finite-dimensional Markovian model, they are unable to preserve fidelity of the abstracted model. To improve accuracy of the abstracted representation, we introduce a memory-dependent abstraction approach to mitigate the modeling error. The first major contribution of this paper is we propose a Neural Network based method to learn the non-Markovian transition model based on the Mori-Zwanzig (M-Z) formalism. Different from existing work in applying M-Z formalism to autonomous time-invariant systems, our approach is the first work generalizing the M-Z formalism to robotics, by addressing the non-Markovian modeling of the belief dynamics that is dependent on historical observations and actions. The second major contribution is we theoretically show that modeling the non-Markovian memory effect in the abstracted belief dynamics improves the modeling accuracy, which is the key benefit of the proposed algorithm. Simulation experiment of a belief space planning problem is provided to validate the performance of the proposed belief abstraction algorithms.
The nonlinear and memory-dependent behaviors of soft robots impose significant challenges to accurate dynamic modeling. This paper introduces a novel Mori-Zwanzig (MZ) Koopman operator framework that incorporates memory effects into the dynamic modeling of a cable-driven soft robotic arm. By integrating memory kernels from the MZ modal decomposition into the Koopman operator, the proposed approach renders a data-driven approximation of nonlinear memory-dependent dynamics by a linear dynamic model in the observable space. An experimental platform is employed to generate data for system identification with two scenarios examined: unidirectional bending and multidirectional deformation. The identified MZ-Koopman model is evaluated against the traditional Koopman model, linear state-space model, and LSTM neural-network model by comparing their predictions of the robotic arm’s motion against the truth data. The experimental results demonstrate that the MZ-Koopman model significantly improves modeling accuracy over the other three methods, particularly achieving more than 20% improvement over the traditional Koopman operator model. Additionally, this study explores the effects of actuation speed on memory-dependent behaviors and evaluates the impact of memory kernels on modeling accuracy.
This paper presents a novel nonlinear Reduced Order Model (ROM) that combines Proper Orthogonal Decomposition (POD) with deep learning residual error correction. Our model employs deep learning for error correction in both the projection and time integration phases of the ROM. In contrast to existing hybrid ROM methods, which typically focus on correcting dynamics only within the POD subspace using Mori-Zwanzig-inspired strategies, our approach also explicitly models and corrects truncation errors outside the subspace using an autoencoder. Our hybrid ROM is trained using an end-to-end neural Ordinary Differential Equation (ODE) framework. We assess its performance using a one-dimensional viscous Burgers’ equation and the cylinder flow test case at Reynolds 100, two well-known test cases that challenge complementary aspects of the model. The paper shows that this novel approach is more accurate than all other tested models: POD Galerkin ROMs, a purely data-driven approach using only autoencoders and some state-of-the-art hybrid methods.
No abstract available
No abstract available
No abstract available
No abstract available
We present a bottom-up coarse-graining (CG) method to establish implicit-solvent CG modeling for polymers in solution, which conserves the dynamic properties of the reference microscopic system. In particular, tens to hundreds of bonded polymer atoms (or Lennard-Jones beads) are coarse-grained as one CG particle, and the solvent degrees of freedom are eliminated. The dynamics of the CG system is governed by the generalized Langevin equation (GLE) derived via the Mori-Zwanzig formalism, by which the CG variables can be directly and rigorously linked to the microscopic dynamics generated by molecular dynamics (MD) simulations. The solvent-mediated dynamics of polymers is modeled by the non-Markovian stochastic dynamics in GLE, where the memory kernel can be computed from the MD trajectories. To circumvent the difficulty in direct evaluation of the memory term and generation of colored noise, we exploit the equivalence between the non-Markovian dynamics and Markovian dynamics in an extended space. To this end, the CG system is supplemented with auxiliary variables that are coupled linearly to the momentum and among themselves, subject to uncorrelated Gaussian white noise. A high-order time-integration scheme is used to solve the extended dynamics to further accelerate the CG simulations. To assess, validate, and demonstrate the established implicit-solvent CG modeling, we have applied it to study four different types of polymers in solution. The dynamic properties of polymers characterized by the velocity autocorrelation function, diffusion coefficient, and mean square displacement as functions of time are evaluated in both CG and MD simulations. Results show that the extended dynamics with auxiliary variables can construct arbitrarily high-order CG models to reproduce dynamic properties of the reference microscopic system and to characterize long-time dynamics of polymers in solution.
We derive a set of generalized master equations (GMEs) to study charge transport dynamics in molecular junctions using the Nakajima-Zwanzig-Mori projection operator approach. In the new GME, time derivatives of population on each quantum state of the molecule, as well as the tunneling current, are calculated as the convolution of time non-local memory kernels with populations on all system states. The non-Markovian memory kernels are obtained by combining the hierarchical equations of motion (HEOM) method and a previous derived Dyson relation for the exact kernel. A perturbative expansion of these memory kernels is then calculated using the extended HEOM developed in our previous work [M. Xu et al., J. Chem. Phys. 146, 064102 (2017)]. By using the resonant level model and the Anderson impurity model, we study properties of the exact memory kernels and analyze convergence properties of their perturbative expansions with respect to the system-bath coupling strength and the electron-electron repulsive energy. It is found that exact memory kernels calculated from HEOM exhibit short memory times and decay faster than the population and current dynamics. The high order perturbation expansion of the memory kernels can give converged results in certain parameter regimes. The Padé and Landau-Zener resummation schemes are also found to give improved results over low order perturbation theory.
No abstract available
There has been a long history of applying the generalized master equation (GME) to study charge carrier and exciton transport in molecular systems. Yet exact memory kernels in the GME are generally difficult to obtain. In this work, exact memory kernels of the Nakajima-Zwanzig-Mori GME for a one dimensional Holstein type of model are calculated by employing the Dyson relation for the exact memory kernel, combined with the hierarchical equations of motion method. Characteristics of the exact memory kernels, as well as the transition rate constants within the Markovian approximation, are then analyzed for different sets of parameters ranging from the hopping to bandlike transport regimes. It is shown that, despite the memory effect of the exact kernels, the Markovian approximation to the exact GME can reproduce the diffusion constants accurately. We also investigate the validity of the second and fourth order perturbation theories with respect to the electronic coupling constant in calculating the rate constants and the diffusion constant. It is found that, due to the cancellation of errors, the second order diffusion constant gives a reasonable estimate of the exact one within a wide range of electronic coupling constants.
Data-driven Koopman operator methods, particularly Extended Dynamic Mode Decomposition (EDMD) and Hankel Dynamic Mode Decomposition (Hankel-DMD), have emerged as promising reduced-order models for nonlinear spacecraft attitude dynamics. However, their performance is typically evaluated using one-step prediction accuracy, leaving their reliability for autonomous long-horizon propagation largely unverified. This study investigates the closed-loop rollout performance of a Hankel-DMD surrogate for a gravity-gradient-stabilized 3U CubeSat with reaction-wheel momentum bias and compares it with a conventional equilibrium linearization. The Hankel-DMD model achieves excellent one-step prediction accuracy (RMSE = 0.049°), but its autonomous rollout error increases from 7.49° after 0.05 orbital periods to 90.58° after 3 orbital periods, despite incorporating delay embedding, feature standardization, ridge regularization, and spectral-radius stabilization. In contrast, the equilibrium linearization maintains prediction errors below 3.5° throughout the evaluated horizons and consistently outperforms the data-driven model. To quantify this discrepancy, this work introduces the Rollout Amplification Factor (RAF), which measures the growth of autonomous rollout error relative to one-step prediction and reaches 1,849× after three orbital periods. The results demonstrate that excellent local prediction accuracy does not necessarily imply reliable autonomous long-horizon performance and identify a previously overlooked closed-loop rollout validation gap in Koopman-based spacecraft attitude modelling. These findings establish that autonomous rollout evaluation against classical analytical baselines should complement conventional one-step validation when assessing Koopman-based reduced-order models for spacecraft applications.
Characterising the oscillatory behaviour generated by neural interactions is central to systems neuroscience because these dynamical properties are linked to brain function. However, inferring latent neural dynamics is challenging because recordings sample only part of the system. Time-delay embedding allows dynamical modes to be estimated from partial recordings, using methods such as Hankel dynamic mode decomposition (Hankel DMD). Despite its promise, two questions remain: what dynamical information is recoverable through time-delay embedding, and how can recovery be improved? We address both questions using control theory. Our framework centres on factorising the Hankel matrix of delay-embedded observations into an observability matrix and a trajectory matrix with the form of a controllability matrix. We show that their ranks impose complementary constraints on what any delay-embedding method can recover: delay embedding can increase observability rank, while perturbation can increase trajectory rank by exciting otherwise inactive modes. We validate this framework using Hankel DMD in simulations and human electrocorticography (ECoG). In both, perturbation increased the number of fast-decaying mode estimates, whereas greater delay depth increased the number of reliably estimated modes and, in ECoG, extended their frequency range. These results clarify what partial recordings can reveal about neural dynamics and how to reveal more.
This study proposes a hybrid framework combining a Hankel-inspired reduced-order model (HROM) and a Temporal Convolutional Network (TCN)–Transformer network for predicting the wind field and surface pressure around a ground-mounted rectangular cylinder. Unlike purely data-driven approaches, the proposed method explicitly incorporates time-delay embedding to reconstruct an augmented dynamical phase space, enabling the reduced-order representation to better preserve the intrinsic temporal evolution of coherent flow structures. The HROM can effectively extract dominant large-scale modes associated with flow separation, shear layer, and wake dynamics, thereby providing a physically meaningful low-dimensional representation of the highly nonlinear wind field. On this basis, the TCN–Transformer network is employed to model the temporal evolution of modal coefficients, where the TCN captures local unsteady fluctuations induced by vortex dynamics, while the Transformer component accounts for long-range temporal dependencies associated with global flow organization. The results demonstrate that the proposed framework achieves accurate prediction of both velocity and surface pressure fields. In addition, increasing the delay embedding dimension in the HROM strengthens the energetic coherence of dominant flow modes, improves the physical interpretability of the reduced-order dynamics, and enhances overall prediction accuracy.
Dynamic mode decomposition (DMD) is widely used in extracting the main features of high-dimensional nonlinear systems in computational fluid dynamics. However, it cannot be directly applied to real compressors due to the limitations of data acquisition over the entire flow field. In order to overcome this limitation, our paper proposes a time-delay-locked DMD algorithm. It combines the idea of phase locking, Hankel time-delay method, and DMD method. Based on it, the flow patterns are successfully extracted and the unsteady flow phenomena including rotating stall and rotating instability are explained as modal waves. Multi-peak spectrum nearly 1/2 BPF observed under off-design RI conditions is explained. In addition, multi-scale of rotating stall is analyzed and shock wave is captured.
Numerical simulations and real-world monitoring data are increasingly prevalent in fields such as solid mechanics, fluid dynamics, and structural engineering. However, developing accurate models that capture the underlying system dynamics from sparse and noisy measurements remains a significant challenge. To address this, we propose a novel methodology called Minimal Realization Time-Delay Koopman (MRTK) analysis. This method combines time-delay embedding with Koopman operator theory to transform nonlinear dynamics into a linearized form. Additionally, it employs Singular Value Decomposition (SVD) to reduce model order, enhancing computational efficiency and accurately identifying system dynamics from sparse and noisy measurements. By explicitly modeling noise in the data, we demonstrate that the MRTK method serves as a generalized extension of Dynamic Mode Decomposition (DMD), encompassing variants such as Extended DMD and Total Least Squares DMD, while also establishing theoretical connections with both Hankel Alternative View of Koopman (HAVOK) analysis and Subspace DMD. We validate the proposed approach using simulated data from transitional channel flow and the Lorenz system, as well as real-world wind speed and main girder acceleration measurements from the Hangzhou Bay Bridge. The results show that integrating the identified reduced-order model with a Kalman filter enables real-time, accurate estimation and prediction from sparse data. The method achieves high predictive accuracy across all scenarios, with the maximum Normalized Mean Squared Error (NMSE) prediction error for the wind speed field being 1.911%, underscoring its potential to advance the prediction and control of complex systems.
Dynamic mode decomposition (DMD) is extensively used in fluid dynamics to identify dominant flow structures and facilitate flow reconstruction and prediction. Conventional DMD, however, builds temporal snapshot matrices based exclusively on state variables at individual time instants, which often inadequately captures the behavior of localized or complex flows. To overcome this limitation, this study proposes an enhanced DMD technique that incorporates time-delay embedding by using difference values between state variables at consecutive time steps as the building blocks of the snapshot matrix. This approach not only retains instantaneous flow field information but also emphasizes temporal change characteristics, thereby improving the analytical capability of DMD. Moreover, a modal amplitude optimization procedure is integrated to increase the precision of reconstruction and prediction. The proposed method is validated on three test cases with increasing complexity and diverse backgrounds. Results show that the modified method significantly enhances the robustness and reconstruction accuracy of DMD, indicating its improved reliability for practical fluid dynamics applications.
This study investigates the low-dimensional dynamics of the wake flow around a blunt-based axisymmetric model under low-speed conditions. To capture the underlying flow structures, we apply various data-driven techniques, including proper orthogonal decomposition (POD), Hilbert-Huang transform (HHT), Hankel alternative view of Koopman (HAVOK), and cluster-based reduced-order modeling (CROM). Our findings reveal that vortex shedding, bubble pumping, and the stochastic rotation of vortex shedding are the three dominant modes of the wake flow, each associated with distinct Strouhal numbers. Further analysis of the first five POD mode coefficients using the Hilbert-Huang transform uncovers intermittent behavior in the near-wake dynamics, demonstrating the superiority of HHT over Fourier-based methods for characterizing such unsteady phenomena. The HAVOK analysis decomposes the delayed-embedding representation of the first POD mode into linear and nonlinear components, where the eigenvalues of the linear component capture very low-frequency oscillations and the bubble pumping mechanism. Meanwhile, CROM offers an unsupervised alternative to POD, effectively identifying transition dynamics and predictive information of cluster-based flow structures.
No abstract available
Dynamic mode decomposition (DMD) is a data-driven modeling technique suitable to analyze flow structures in numerical and experimental data. It is widely used to extract temporal information about coherent structures from data. In its standard form, however, it cannot simultaneously extract information on temporal and spatial development, such as wave number and spatial growth rate, which are essential in fully developed flows. Spatio-temporal dynamic mode decomposition (STDMD) is an extension of DMD designed to handle spatio-temporal datasets. The framework is extended to analyze data that varies both spatially and temporally. In this paper we introduce an augmented modification of the STDMD method. It consists of applying delay-embedded coordinates to STDMD sequentially, in time and space. Using this method, dominant frequencies, wavenumbers, and their harmonics can be accurately calculated for complex flows. With different illustrative examples, we demonstrate the applicability of the introduced technique.
Sweeping jet actuators (SJAs) are promising for active flow control in aerospace systems, but integrating actuator-resolved unsteady CFD into full-configuration simulations is often impractical due to small geometric scales and O(102) Hz oscillations that demand fine grids and small time steps. This work develops a reduced-order modeling (ROM) framework to generate time-resolved boundary conditions at the actuator exit from SJA flow data. Dynamic mode decomposition (DMD) is particularly attractive for this purpose because it provides a linear, data-driven input–output representation of the actuator effect, even though it does not explicitly model the underlying nonlinear switching mechanism. We introduce an eigenvalue-sorted dynamic mode decomposition (ES-DMD) method that performs stability-aware mode ranking based on the discrete-time DMD eigenvalues, prioritizing modes with ℜ(λ) closest to unity to retain near-neutrally stable oscillatory dynamics, improving robustness relative to conventional amplitude-based selections for high-frequency oscillatory flows. The method is evaluated across multiple operating conditions, with detailed analysis performed for the highest mass-flow case (m˙=0.01 lb/s), representing the most dynamically demanding condition considered. Across multiple operating conditions, ES-DMD yields consistent reconstructions of the dominant switching dynamics. For one-dimensional exit-plane profiles, combining ES-DMD with time-delay embedding enables accurate reconstruction and multi-period prediction using only 20 modes (7.6% of the full system rank). The proposed approach provides a practical pathway to incorporate unsteady SJA effects into large-scale aerospace CFD through compact, predictive boundary-condition models.
No abstract available
No abstract available
The reconstruction of accurate and robust unsteady flow fields from sparse and noisy data in real-life engineering tasks is challenging, particularly when sensors are randomly placed. To address this challenge, a novel Autoencoder State Estimation (AE-SE) framework is introduced in this paper. The framework integrates sensor measurements into a machine learning-based reduced-order model (ROM) by leveraging the low-dimensional representation of flow fields. The proposed approach is tested on two direct numerical simulation benchmark examples, namely, circular and square cylinders and wake flow fields at Re = 100. The results demonstrate satisfactory performance in terms of accuracy and reconstruction efficiency. It achieves the same accuracy as traditional methods while improving reconstruction efficiency by 70%. Moreover, it preserves essential physical properties and flow characteristics even in the noisy data, indicating its practical applicability and robustness. Experimental data validation confirms a relative error below 5% even at a noise level of 12%. The flexibility of the model is further evaluated by testing it with a trained ROM under varying Reynolds numbers and benchmark cases, demonstrating its ability to accurately estimate and recognize previously unseen flow fields with appropriate training datasets. Overall, the proposed AE-SE flow reconstruction method efficiently and flexibly leverages ROM for the low-dimensional representation of complex flow fields from sparse measurements. This approach contributes significantly to the development of downstream applications such as design optimization and optimal control.
No abstract available
The objective of this paper is to incorporate sparse sensor data to improve flow-field estimates in a wind farm, which can then be used to perform better online wind farm optimization and control. A sparse-sensing algorithm is used to determine the optimal locations of sensors to improve the overall estimation precision of the flow field within the wind farm. This algorithm takes advantage of the dominant atmospheric structures in a wind farm to reconstruct the flow field from point measurements in the field. These measurements, in their optimal locations, have the ability to improve the observability of a wind farm and thus provide faster, more accurate, state estimation.
Deep learning has been widely utilized to accurately estimate the flow state from the sparse sensor measurements. Yet there is still a lack of understanding of the actual dimension of this type of regression problem. In this study, we propose an Autoencoder (AE) based estimation method to tackle the control-oriented, sensor-based flow estimation problem. This method encodes the input information into a few-nodes bottleneck, then decodes the compressed information to the estimated the flow state. This network stands for the least order representation of the input-output relationship. We choose the fluidic pinball as the benchmark problem which contains multiple-input and multiple-output. The rotations of three cylinders in the flow generates rich flow physics. Consequently, strong non-linearity arises between input measurements and output flow state. For fair comparison the Deep Neural Network (DNN) is also employed to highlight the influence of the encoding-decoding process in AE. The results show that when the flow is periodic, the input vector can be effectively encoded into a five-dimensional subspace without a significant loss of estimation accuracy. However, the five-dimensional encoding will eliminate partial information from the input vector when the forced flow becomes chaotic, resulting in a lower estimation accuracy than DNN.
Getting access to the state of turbulent flow from limited sensor measurements in engineering systems is a major challenge. Development of technologies to accurately estimate the state of the flow is now possible with the use of machine learning. We present a supervised machine learning technique to reconstruct turbulent vortical structures in a pump sump from sparse surface pressure measurements. For the current flow reconstruction technique, a combination of multi-layer perceptron and three-dimensional convolutional neural network is utilized. This technique provides accurate flow estimation from only a few sensor measurements, identifying the presence of adverse vortices. The dependence of the model performance on the amount of training data, the number of input sensors, and the noise levels are investigated. The present machine learning-based flow estimator supports safe operations of pumps and can be extended to a broad range of applications for industrial fluid-based systems.
No abstract available
Currently, wind power is the fastest-growing sus-tainable energy resource, and accurately predicting wind velocity for individual turbines is essential. Wind behavior is turbulent, with variable patterns, characteristics, and unpredictable extreme events. Moreover, the wind reaching each turbine is impacted by wake loss effects from surrounding turbines, which underscores the need for a precise local prediction model for each turbine to enhance accuracy in smart wind power production. This study proposes a novel approach using local measurements at fixed points, adopting the Eulerian perspective in fluid dynamics to forecast wind turbulence for the coming hour using a deep learning model. A long short-term memory (LSTM) forecasting model is tuned and applied. Laboratory data from wind tunnel tests are utilized to account for velocity fluctuations and wake loss effects, aiming to establish similarity to an affected single wind turbine. Model training and inference run on the high-performance computing (HPC) DEEP-DAM systems at the Julich Supercomputing Centre. Three train-test datasets are employed to evaluate the model. The measured metrics indicate that the mean absolute error is 0.048, the root mean square error is approximately 0.065, and the Pinball loss for the 90th percentile is 0.0042. The model demonstrates a robust predictive capability, supporting the development of advanced instruments for individ-ual turbines and facilitating smart, clean, and sustainable energy production.
No abstract available
We derive a five–dimensional Modified Lorenz System (MLS) from the incompressible Boussinesq Navier–Stokes equations with coupled potential temperature and moisture fields via systematic Galerkin projection onto a small set of physically motivated spatial modes. Beginning from ∂_t u+(u⋅∇)u=-∇p+ν∇^2 u+gα Θ z ̂, ∂_t Θ+(u⋅∇)Θ=κ_Θ ∇^2 Θ+S_Θ, and a passive moisture equation, we nondimensionalize and project onto a divergence–free modal basis, leading to q ̇_m=∑_n▒ L_mn q_n+∑_(i,j)▒ C_mij q_i q_j+F_m, with closure hypotheses for unresolved modes. A judicious choice of five dominant amplitudes (T,H,P,W,R)^⊤ yields the MLS, which generalizes the classical Lorenz–63 model by incorporating moisture coupling and bounded nonlinear closures. We provide a detailed analytical characterization of the MLS: (i) fixed points are identified and classified, (ii) linear stability is determined via the Jacobian eigenvalue spectrum, and (iii) Lyapunov exponents are defined through the tangent–linear system ξ ̇=J(X(t)) ξ, with the largest exponent λ_1 setting a predictability time scale T_L∼1/λ_1. Using energy estimates, we prove the existence of an absorbing set under admissible damping and closure conditions, ensuring global boundedness of solutions. The MLS thus serves as a mathematically well–posed reduced–order model capturing essential nonlinear dynamics of moist convective systems and providing a reproducible analytical testbed for future studies in predictability, ensemble design, and reduced modeling of complex atmospheric processes.
No abstract available
Turbulence parameterization plays a critical role in the simulation of many weather regimes. For challenging cases such as the stratocumulus-capped boundary layer (SCBL), traditional schemes can produce unrealistic results even when a fine large-eddy-simulation (LES) resolution is used. Here we present an implicit generalized linear algebraic subfilter-scale model (iGLASS) to better represent unresolved turbulence in the simulation of the atmospheric boundary layer, at both standard LES and so-called terra incognita (TI) resolutions. The latter refers to a range of model resolutions where turbulent eddies are only partially resolved, and therefore the simulated processes are sensitive to the representation of unresolved turbulence. iGLASS is based on the truncated conservation equations of subfilter-scale (SFS) fluxes, but it integrates the full equations of the SFS turbulence kinetic energy and potential energy to retain “memory” of the SFS turbulence. Our evaluations suggest iGLASS can perform significantly better than traditional eddy-diffusivity models and exhibit skills comparable to the dynamic reconstruction model (DRM). For a neutral boundary layer case run at LES resolution, the simulation using iGLASS exhibits a wind profile that reasonably matches the similarity-theory solution. For an SCBL case with 5-m vertical resolution, iGLASS maintains more realistic cloud water profiles and boundary layer structure than traditional schemes. The SCBL case is also tested at TI resolution, and iGLASS also exhibits superior performance. iGLASS permits significant backscatter, whereas traditional models allow forward scatter (diffusion) only. As a physics-based approach, iGLASS appears to be a viable alternative for turbulence parameterization.
Turbulence modeling has been always a challenge, given the degree of underlying spatial and temporal complexity. In this paper, we propose the use of a partitioned reduced order modeling (ROM) approach for efficient and effective approximation of turbulent flows. A piecewise linear subspace is tailored to capture the fine flow details in addition to the larger scales. We test the partitioned ROM for a decaying two-dimensional (2D) turbulent flow, known as 2D Kraichnan turbulence. The flow is initiated using an array of random vortices, corresponding to an arbitrary energy spectrum. We show that partitioning produces more accurate and stable results than standard ROM based on a global application of modal decomposition techniques. We also demonstrate the predictive capability of partitioned ROM through an energy spectrum analysis, where the recovered energy spectrum significantly converges to the full order model’s statistics with increased partitioning. Although the proposed approach incurs increased memory requirements to store the local basis functions for each partition, we emphasize that it permits the construction of more compact ROMs (i.e., of smaller dimension) with comparable accuracy, which in turn significantly reduces the online computational burden. Therefore, we consider that partitioning acts as a converter which reduces the cost of online deployment at the expense of offline and memory costs. Finally, we investigate the application of closure modeling to account for the effects of modal truncation on ROM dynamics. We illustrate that closure techniques can help to stabilize the results in the inertial range, but over-stabilization might take place in the dissipative range.
No abstract available
No abstract available
No abstract available
The channel-assisted Fourier neural operator (CFNO) for real-time, high-resolution, long-horizon prediction of yawed wind-turbine far wakes is introduced. CFNO injects yaw through a channel-assisted layer and combines a post-channel-mixing depthwise-separable spectral convolution (DSSC) with real-valued weights and residual-enhanced Fourier layers. These choices stabilize long autoregressive predictions and reduce parameters to about 23% of the Fourier neural operator (FNO) and about 50% of DSSC-FNO. CFNO is trained and validated on large-eddy simulation (LES) produced with an actuator-disk model with rotation (ADM-R), which closely reproduces far-wake behavior. It achieves 0.89% error on a 180 × 70 high-fidelity far-wake grid (up to x/D=12), exceeds an FNO baseline, and runs at millisecond-to-sub-second latency on a single GPU. These capabilities enable digital-twin-assisted wake steering and differentiable, yaw-based model-predictive control with potential to reduce wake losses, moderate cyclic loads, and stabilize aggregate power.
Deep learning is of growing interest to the fluids community due its potential applications for real-time prediction and control. Indeed, whereas computational fluid dynamics solvers are prohibitively time-intensive for real-time implementation, deep learning methods can be evaluated in milliseconds once trained. However, the majority of efforts to apply deep learning methods to flows have been performed with computational data rather than experimental data, in part due to the lack of freely available large experimental datasets. In this paper we present a 64.7 GB dataset of two-dimensional velocity fields measured with particle image velocimetry. These fields show the temporal evolution of cylinder wakes in the sub-critical regime for Reynolds numbers ranging from 240 to 2520. The data is broken into discrete time series made up of 100 time-steps each, representing approximately three vortex shedding cycles. The dataset, available for immediate and free use, is poised to facilitate research on deep learning methods in fluids and other endeavors where such a large, detailed experimental dataset is of utility.
This study aims at the prediction of the turbulent flow behind cylinder arrays by the application of Echo State Networks (ESN). Three different arrangements of arrays of seven cylinders are chosen for the current study. These represent different flow regimes: single bluff body flow, transient flow, and co-shedding flow. This allows the investigation of turbulent flows that fundamentally originate from wake flows yet exhibit highly diverse dynamics. The data is reduced by Proper Orthogonal Decomposition (POD) which is optimal in terms of kinetic energy. The Time Coefficients of the POD Modes (TCPM) are predicted by the ESN. The network architecture is optimized with respect to its three main hyperparameters, Input Scaling (INS), Spectral Radius (SR), and Leaking Rate (LR), in order to produce the best predictions in terms of Weighted Prediction Score (WPS), a metric leveling statistic and deterministic prediction. In general, the ESN is capable of imitating the complex dynamics of turbulent flows even for longer periods of several vortex shedding cycles. Furthermore, the mutual interdependencies of the TCPM are well preserved. However, optimal hyperparameters depend strongly on the flow characteristics. Generally, as flow dynamics become faster and more intermittent, larger LR and INS values result in better predictions, whereas less clear trends for SR are observable.
No abstract available
When employing traditional low-order approximation equations to forecast the Hopf bifurcation phenomenon in the wake of a circular cylinder at low Reynolds numbers, inaccuracies may arise in estimating the phase. This is due to the fact that, in this transition process, the frequency varies with time. In this paper, we propose a method for analyzing and predicting the vortex shedding behind a cylinder at low Reynolds numbers. The proposed method is based on coordinate transformation and description function and is demonstrated using data from computational fluid dynamics simulation of flow around a cylinder at Reynolds number 100. The resulting governing equations explicitly contain the flow amplitude and implicitly contain the flow frequency. The proposed method is found to have higher accuracy compared to other methods for nonlinear identification and order reduction. Finally, the method is extended to predict nonlinear vortex shedding in the Reynolds number range of 80–200.
Floating offshore wind farms require rapid, accurate wake predictions to support control and layout optimization, yet fully coupled large-eddy simulations (LES) remain prohibitively expensive. This study proposes an integrated combining LES, dynamic mode decomposition (DMD), and bidirectional long short-term memory (BiLSTM) workflow for short-term wake forecasting around floating offshore wind turbines. First, LES coupled with the MoorDyn dynamic-mooring model resolves six degrees of freedom platform motions for the offshore code comparison collaboration continuation-DeepCwind semi-submersible floating wind turbine under combined wind-wave-current loading. DMD compresses the resulting flow fields into a set of spatial modes associated with distinct frequencies, and a BiLSTM network learns the temporal evolution of the leading coefficients. The framework is benchmarked on canonical cylinder flow and then applied to the floating wind turbine. Retaining the first 27 DMD modes captures >98% of wake energy, while the BiLSTM achieves mean absolute and root-mean-square errors below 1% and 1.5%, respectively, over a 5 s prediction horizon (one rotor revolution), with R2 > 0.96. Compared with full LES cases, the surrogate achieves an order-of-magnitude reduction in computational cost without compromising key wake features such as velocity deficit recovery and vortex shedding dynamics. The proposed DMD-BiLSTM model offers a physics-informed, real-time capable tool for wake prediction, providing a physics-informed surrogate that supports advanced control strategies and provides high-fidelity input for dense farm layout optimization.
Abstract The evolution of small perturbations applied to turbulent Couette flow is examined over a long time horizon, until the perturbed flow field becomes completely decorrelated from the original state. To elucidate the fundamental physical processes involved, we focus on the minimal flow unit, where the dynamics of coherent structures is well understood in terms of the self-sustaining process (Hamilton et al. J. Fluid Mech. vol. 287, 1995, pp. 317–348). As expected, in the short term, perturbations exhibit exponential growth governed by the leading Lyapunov exponent. This mechanism is driven by the streamwise-dependent flow, which is known to involve intense turbulent dissipation events within the self-sustaining process, consistent with previous findings. Beyond the initial exponential phase, we observe a slow, sustained growth of perturbations over a long period – spanning tens of integral time scales – before eventual saturation. During this stage, the perturbation energy increases approximately linearly with time. While this behaviour resembles observations and predictions in other turbulent flows, the underlying physical process here is fundamentally different. Specifically, the perturbation field during this period is dominated by streaky structures and the growth mechanism is linked to the saturation of the wall-normal streak length scale at the largest dimension permitted by the flow geometry (i.e. the channel height). Finally, an evaluation of the dominant production term components reveals that the well-known lift-up effect is primarily responsible for the growth of these streaky perturbations.
No abstract available
No abstract available
<p><span>Large-eddy simulation (LES) is an often used technique in the geosciences to simulate turbulent oceanic and atmospheric flows. In LES, the effects of the unresolved turbulence scales on the resolved scales (via the Reynolds stress tensor) have to be parameterized with subgrid models. These subgrid models usually require strong assumptions about the relationship between the resolved flow fields and the Reynolds stress tensor, which are often violated in reality and potentially hamper their accuracy.</span></p><p><span>In this study, using the finite-difference computational fluid dynamics code MicroHH (v2.0) and turbulent channel flow as a test case (friction Reynolds number Re<sub>τ</sub> 590), we incorporated and tested a newly emerging subgrid modelling approach that does not require those assumptions. Instead, it relies on neural networks that are highly non-linear and flexible. Similar to currently used subgrid models, we designed our neural networks such that they can be applied locally in the grid domain: at each grid point the neural networks receive as an input the locally resolved flow fields (u,v,w), rather than the full flow fields. As an output, the neural networks give the Reynolds stress tensor at the considered grid point. This local application integrates well with our simulation code, and is necessary to run our code in parallel within distributed memory systems.</span></p><p><span>To allow our neural networks to learn the relationship between the specified input and output, we created a training dataset that contains ~10.000.000 samples of corresponding inputs and outputs. We derived those samples directly from high-resolution 3D direct numerical simulation (DNS) snapshots of turbulent flow fields. Since the DNS explicitly resolves all the relevant turbulence scales, by downsampling the DNS we were able to derive both the Reynolds stress tensor and the corresponding lower-resolution flow fields typical for LES. In this calculation, we took into account both the discretization and interpolation errors introduced by the finite staggered LES grid. Subsequently, using these samples we optimized the parameters of the neural networks to minimize the difference between the predicted and the ‘true’ output derived from DNS.</span></p><p><span>After that, we tested the performance of our neural networks in two different ways:</span></p><ol><li><span>A priori or offline testing, where we used a withheld part of the training dataset (10%) to test the capability of the neural networks to correctly predict the Reynolds stress tensor for data not used to optimize its parameters. We found that the neural networks were, in general, well able to predict the correct values. </span></li> <li><span>A posteriori or online testing, where we incorporated our neural networks directly into our LES. To keep the total involved computational effort feasible, we strongly enhanced the prediction speed of the neural network by relying on highly optimized matrix-vector libraries. The full successful integration of the neural networks within LES remains challenging though, mainly because the neural networks tend to introduce numerical instability into the LES. We are currently investigating ways to minimize this instability, while maintaining the high accuracy in the a priori test and the high prediction speed.</span></li> </ol>
Surface friction contributes to tornado formation and maintenance by enhancing the convergence of angular momentum. The traditional lower boundary condition in atmospheric models typically assumes an instant equilibrium between the unresolved stress and the resolved shear. This assumption ignores the physics that turbulent motions are generated and dissipated at finite rates—in effect, turbulence has a memory through its lifetime. In this work, a modified lower boundary condition is proposed to account for the effect of turbulence memory. Specifically, when an air parcel moves along a curved trajectory, a normal surface-shear-stress component arises owing to turbulence memory. In the accompanying large-eddy simulation (LES) of idealized tornadoes, the normal surface-shear-stress component is a source of additional dynamic instability, which provides an extra pathway for the development of turbulent motions. The influence of turbulence memory on the intensity of quasi-steady-state tornadoes remains negligible as long as assumptions employed by the modified lower boundary condition hold over a relatively large fraction of the flow region of interest. However, tornadoes in a transient state may be especially sensitive to turbulence memory.Friction between the wind and the ground can influence atmospheric phenomena in important ways. For example, surface friction can be a significant source of rotation in some thunderstorms, and it can also help to intensify rotation when rotation is already present. Unfortunately, the representation of friction’s effects in atmospheric simulations is especially error-prone in phenomena characterized by rapid temporal evolution or strong spatial variations. Our work explores a new framework for representing friction to include the effect of the so-called turbulence memory. The approach is tested in idealized tornado simulations, but it may be applied to a wide range of atmospheric vortices.
The accurate prediction of unsteady propeller wake dynamics is pivotal for advancing marine propulsion analysis and hydrodynamic design. These wakes are characterized by complex three-dimensional interactions involving blade loading, rotation, and turbulence, resulting in highly transient velocity and pressure fields. High-fidelity methods such as direct numerical simulation and large-eddy simulation (LES) can resolve these dynamics but are computationally prohibitive for iterative design and real-time analysis. Physics-informed neural networks offer a promising alternative by embedding physical laws into learning, yet conventional multilayer perceptron-based formulations often fail to capture instantaneous unsteady features, leading to phase shifts and over-smoothed predictions. To address these challenges, we propose the physics-informed Least Squares Neural Network (PhyLSNN), a novel spatiotemporal framework for propeller wake prediction. The architecture integrates a convolutional encoder-convolutional long short-term memory-decoder backbone with a least squares finite-difference scheme to robustly compute Navier–Stokes residuals. This hybridization enables the network to preserve coherent temporal evolution and physically accurate flow representations. Trained on LES-derived propeller flow data, PhyLSNN achieves high-fidelity and temporally stable reconstruction of velocity and pressure fields with minimal computational demand. The findings highlight PhyLSNN as a robust and physics-consistent framework for unsteady propeller wake prediction, offering an efficient pathway toward performance assessment and hydrodynamic design optimization.
Abstract The dynamics of turbulent flows past lateral cavities is relevant for multiple environmental applications. In rivers and coastal environments, these lateral recirculating regions constitute surface storage zones, where large-scale turbulent coherent structures control the transport and fate of contaminants. Mass transport in these flows is typically represented by one-dimensional first-order equations that predict the evolution of the spatially integrated concentration between the cavity and the main channel. These models, however, cannot represent the long-term evolution of the concentration or incorporate memory effects induced by turbulence. In this investigation, we carry out large-eddy simulations (LES) of the open-channel flow with a lateral square cavity of Mignot et al. (Phys. Fluids, vol. 28, issue 4, 2016, 045104). The model is coupled with an advection–diffusion equation and a Lagrangian particle model to investigate the transport mechanisms in the cavity and across the interface. From the simulations we provide quantitative comparisons of the physical processes from both perspectives, and investigate the effects of turbulent coherent structures on residence times and trajectories from finite-time Lyapunov exponents. From the Lagrangian results, we identify general spatial distributions of time scales in the cavity associated with the dynamics of coherent structures, providing new insights into the mechanisms that drive the global transport. We also show that an upscaled model informed by LES and based on a fractional derivative captures the evolution of concentration, and the exchange between the cavity and the main channel, providing accurate predictions of mass transport and reproducing the temporal dependence observed at larger scales.
No abstract available
Abstract Transport closures lose accuracy when coarse models are built in coordinates that do not match the dominant geometry of transport: smoothing, sharp fronts, dispersion, irregular connectivity, or phase-space scattering. This article answers the question How do we choose the right generalized multiscale coordinates for transport closure? by proposing an operator-first selection principle: choose dictionaries that render the transport generator sparse or compressible, then close by coarse–fine elimination in coefficient space under admissibility constraints. The principle is made concrete and verifiable through three interlocking contributions not present in standard multiscale reduction methods (POD, DMD, homogenization, NLMC): (i) a quantitative operator sparsity criterion (ρ p ( A , k) ≤ ρ*) that measures dictionary quality directly from the generator matrix; (ii) a concrete, regime-agnostic selection algorithm with explicit computational cost and verifiable stopping criteria; and (iii) systematic numerical tests on five baseline regimes – each with a head-to-head sparsity comparison between the proposed dictionary and at least one competitor – demonstrating 3–10× compression gains and quantified constraint preservation. We develop a unified constitutive operator framework allowing anisotropy, spatial/temporal nonlocality, and memory, and reformulate transport generators in regime-adapted multiscale coordinates – wavelets (diffusive smoothing), directional frames (sharp fronts), wave packets (dispersive transport), diffusion wavelets (irregular media), and hybrid MCA decompositions (mixed morphology). In these coordinates, linear problems close via Schur-complement elimination and nonlinear problems via constraint-preserving backreaction maps. We address A FF invertibility conditions, clarify the Schur/Mori–Zwanzig correspondence, and show how memory kernels emerge from fine-scale elimination. Constraint-preserving admissibility criteria (conservation, dissipativity, symmetry, realizability) are stated as propositions with proofs, and all nine figures are tied to specific quantitative claims.
No abstract available
No abstract available
Significance Slow–fast systems arise in many scientific applications, in particular in atmospheric and oceanic flows with fast inertia–gravity waves and slow geostrophic motions. When the slow and fast variables are strongly coupled—symptomatic of breakdown of slow-to-fast scales deterministic parameterizations—it remains a challenge to derive reduced systems able to capture the dynamics. Here, generic ingredients for successful reduction of such systems are identified and illustrated for the paradigmatic atmospheric Lorenz 80 model. The approach relies on a filtering operated through a nonlinear parameterization that separates the full dynamics into its slow motion and fast residual dynamics. The latter is mainly orthogonal to the former and is modeled via networks of stochastic nonlinear oscillators, independent of the slow dynamics. The problems of identifying the slow component (e.g., for weather forecast initialization) and of characterizing slow–fast interactions are central to geophysical fluid dynamics. In this study, the related rectification problem of slow manifold closures is addressed when breakdown of slow-to-fast scales deterministic parameterizations occurs due to explosive emergence of fast oscillations on the slow, geostrophic motion. For such regimes, it is shown on the Lorenz 80 model that if 1) the underlying manifold provides a good approximation of the optimal nonlinear parameterization that averages out the fast variables and 2) the residual dynamics off this manifold is mainly orthogonal to it, then no memory terms are required in the Mori–Zwanzig full closure. Instead, the noise term is key to resolve, and is shown to be, in this case, well modeled by a state-independent noise, obtained by means of networks of stochastic nonlinear oscillators. This stochastic parameterization allows, in turn, for rectifying the momentum-balanced slow manifold, and for accurate recovery of the multiscale dynamics. The approach is promising to be further applied to the closure of other more complex slow–fast systems, in strongly coupled regimes.
The identification of dynamical systems from data is essential in control theory, enabling the creation of mathematical models that accurately represent the behavior of complex systems. However, real-world applications often present challenges such as the unknown dimensionality of the system and limited access to measurements, particularly in partially observed systems. The Hankel alternative view of Koopman (HAVOK) method offers a data-driven approach to identify linear representations of nonlinear systems, but it often overlooks the influence of external control signals (inputs) and disturbances. This paper introduces a novel input-aware modeling method for unstable linear systems using data-driven Koopman analysis. By explicitly incorporating the impact of inputs and disturbances, our method enhances the accuracy and robustness of system identification, even in the face of incomplete observations. The proposed approach leverages Koopman operator theory on augmented state-input data to capture both the intrinsic dynamics and the system’s sensitivity to external control. Through extensive numerical examples, we demonstrate the effectiveness of our method in accurately identifying and predicting the behavior of various dynamical systems, including real-world nonlinear systems and simulated unstable linear systems with and without disturbances. The results highlight the potential of our approach to advance the field of system identification and control, offering a powerful tool for modeling and analyzing complex systems in diverse applications.
Abstract. The state of the atmosphere, or of the ocean, cannot be exhaustively observed. Crucial parts might remain out of reach of proper monitoring. Also, defining the exact set of equations driving the atmosphere and ocean is virtually impossible because of their complexity. The goal of this paper is to obtain predictions of a partially observed dynamical system without knowing the model equations. In this data-driven context, the article focuses on the Lorenz-63 system, where only the second and third components are observed and access to the equations is not allowed. To account for those strong constraints, a combination of machine learning and data assimilation techniques is proposed. The key aspects are the following: the introduction of latent variables, a linear approximation of the dynamics and a database that is updated iteratively, maximizing the likelihood. We find that the latent variables inferred by the procedure are related to the successive derivatives of the observed components of the dynamical system. The method is also able to reconstruct accurately the local dynamics of the partially observed system. Overall, the proposed methodology is simple, is easy to code and gives promising results, even in the case of small numbers of observations.
This article develops a model-based reinforcement learning (RL) approach to the closed-loop control of nonlinear dynamical systems with a partial nonlinear observation model. We propose an “information-state”-based approach to rigorously transform the partially observed problem into a fully observed problem where the information state consists of the past several observations and control inputs. We further show the equivalence of the transformed and the initial partially observed optimal control problems and provide the conditions to solve for the deterministic optimal solution. We develop a data-based generalization of the iterative linear quadratic regulator (ILQR) for the RL of partially observed systems using a local linear time-varying model of the information-state dynamics approximated by an autoregressive-moving-average (ARMA) model that is generated using only the input–output data. This approach allows us to design a local perturbation feedback control law that provides an optimum solution to the partially observed feedback design problem locally. The efficacy of the developed method is shown by controlling complex high-dimensional nonlinear dynamical systems in the presence of model and sensing uncertainty.
The identification of partially observed continuous nonlinear systems from noisy and incomplete data series is an actual problem in many branches of science, for example, biology, chemistry, physics, and others. Two stages are needed to reconstruct a partially observed dynamical system. First, one should reconstruct the entire phase space to restore unobserved state variables. For this purpose, the integration or differentiation of the observed data series can be performed. Then, a fast-algebraic method can be used to obtain a nonlinear system in the form of a polynomial dynamical system. In this paper, we extend the algebraic method proposed by Kera and Hasegawa to Laurent polynomials which contain negative powers of variables, unlike ordinary polynomials. We provide a theoretical basis and experimental evidence that the integration of a data series can give more accurate results than the widely used differentiation. With this technique, we reconstruct Lorenz attractor from a one-dimensional data series and B. Muthuswamy’s circuit equations from a three-dimensional data series.
Extreme events appear in many complex nonlinear dynamical systems. Predicting extreme events has important scientific significance and large societal impacts. In this paper, a new mathematical framework of building suitable nonlinear approximate models is developed, which aims at predicting both the observed and hidden extreme events in complex nonlinear dynamical systems for short-, medium-, and long-range forecasting using only short and partially observed training time series. Different from many ad hoc data-driven regression models, these new nonlinear models take into account physically motivated processes and physics constraints. They also allow efficient and accurate algorithms for parameter estimation, data assimilation, and prediction. Cheap stochastic parameterizations, judicious linear feedback control, and suitable noise inflation strategies are incorporated into the new nonlinear modeling framework, which provide accurate predictions of both the observed and hidden extreme events as well as the strongly non-Gaussian statistics in various highly intermittent nonlinear dyad and triad models, including the Lorenz 63 model. Then, a stochastic mode reduction strategy is applied to a 21-dimensional nonlinear paradigm model for topographic mean flow interaction. The resulting five-dimensional physics-constrained nonlinear approximate model is able to accurately predict extreme events and the regime switching between zonally blocked and unblocked flow patterns. Finally, incorporating judicious linear stochastic processes into a simple nonlinear approximate model succeeds in learning certain complicated nonlinear effects of a six-dimensional low-order Charney-DeVore model with strong chaotic and regime switching behavior. The simple nonlinear approximate model then allows accurate online state estimation and the short- and medium-range forecasting of extreme events.
This paper develops a data-based approach to the closed-loop control of nonlinear dynamical systems with a partial nonlinear observation model. We propose an "information-state" based approach to rigorously transform the partially observed problem into a fully observed problem where the information-state consists of the past several observations and control inputs. We further show the equivalence of the transformed and the initial partially observed optimal control problems and provide the conditions to solve for the deterministic optimal solution. We develop a data-based generalization of the iterative Linear Quadratic Regulator (ILQR) for partially-observed systems using a local linear time-varying model of the information-state dynamics approximated by an Autoregressive–moving-average (ARMA) model that is generated using only the input-output data. This open-loop trajectory optimization solution is then used to design a local feedback control law, and the composite law then provides an optimum solution to the partially observed feedback design problem. The efficacy of the developed method is shown by controlling complex high dimensional nonlinear dynamical systems in the presence of model and sensing uncertainty.
No abstract available
Abstract A nonlinear optimal smoother and an associated optimal strategy of sampling hidden model trajectories are developed for a rich class of complex nonlinear turbulent dynamical systems with partial and noisy observations. Despite the strong nonlinearity and the significant non-Gaussian characteristics in the underlying systems, both the optimal smoother estimates and the sampled trajectories can be solved via closed analytic formulae. Thus, they are computationally efficient and the methods are applicable to high-dimensional systems. The nonlinear optimal smoother is able to estimate the hidden model states associated with various non-Gaussian phenomena and is particularly skillful in capturing the onset, demise and amplitude of the observed and hidden extreme events. On the other hand, the sampled hidden trajectories succeed in recovering both the dynamical and statistical features of the underlying nonlinear systems, including the fat-tailed non-Gaussian probability density function and the temporal autocorrelation function. In the situations with only a short period of partially observed training time series, the optimal sampling strategy can be used to efficiently create a sufficient number of samples in an unbiased fashion that facilitates an accurate prediction of important non-Gaussian features in both the observed and hidden variables. In addition, the information provided by the sampled trajectories based on imperfect models allows an effective way of quantifying the model error. It also offers a systematic approach to improve approximate models and stochastic parameterizations in highly non-Gaussian systems and thus advances the real-time forecasts.
<p>In this work, we consider turbulence closures of LES (Large Eddy Simulation) type for classical decaying 2D turbulence in a priori and a posteriori experiments using explicit filtering approach. According to Germano 1986 decomposition, full subfilter stress for Gaussian filter is decomposed into three Galilean-invariant parts: Leonard, Cross and Reynolds stresses. By analysing spectral transfer of energy and enstrophy, we show that Leonard stress redistributes resolved energy toward large scales and dissipates substantial part of enstrophy, while Cross stress provides an additional enstrophy dissipation at subfilter scales and Reynolds stress predominantly injects energy into middle scales (i.e., Kinetic Energy Backscatter). Substantial part of enstrophy dissipation is located on subfilter scales, and it should be accounted for by choosing base filter wide enough compared to mesh step of LES model. Otherwise, significant fraction of enstrophy dissipation will correspond to subgrid scale stress, which is less universal and harder to approximate. As a result of a priori analysis, we propose LES closure consisting of three parts: SSM (Scale Similarity Model), which is equivalent to Leonard stress, biharmonic Smagorinsky damping as a Cross stress counterpart and ADM (Approximate Deconvolution Model) approximation for Reynolds stress ("backscatter"). The proposed model have two free parameters: Smagorinsky constant and amplitude of the backscatter. These parameters are estimated in a posteriori experiments utilizing dynamic approach and energy-enstrophy balance equation, correspondingly. The proposed model have the following distinctive features: it reproduces energy and enstrophy transfer spectra in accordance to the individual components of the subfilter forces, "reproduces" base filter and reproduces energy growth in accordance to the filtered DNS (Direct Numerical Simulation) solution.</p><p>The work was supported by the Russian Foundation for Basic Research (projects 19-35-90023, 18-05-60184) and Moscow Center for Fundamental and Applied Mathematics (agreement with the Ministry of Education and Science of the Russian Federation No. 075-15-2019-1624).</p>
Accurate reconstruction of flow fields from sparse sensor measurements remains challenging due to inherent limitations in sensor sparsity, irregular positioning, and computational inefficiency of existing methods. This study proposes a deep learning framework that integrates autoencoders with attention mechanisms through a two-stage training strategy. In the first stage, a convolutional autoencoder compresses high-resolution flow fields into low-dimensional latent representations. In the second stage, sensor data (including positions and measurements) are encoded via positional embeddings and processed by a self-attention module. These features are mapped to the latent space through a cross-attention regressor and subsequently decoded to reconstruct the complete flow field. The innovations of the proposed framework include the adaptability to arbitrary sensor numbers/positions and a significant improvement of computational efficiency. Case studies on the datasets of cylinder-wake flow and global sea surface temperature demonstrate that this framework achieves superior reconstruction accuracy, outperforming Voronoi convolutional neural network, and is comparable to the state-of-the-art Senseiver in terms of accuracy while reducing training time by 16 times. The framework also exhibits robustness against sensor failures and data noise, enabling real-time, high-precision flow reconstruction for engineering applications such as aerodynamic monitoring and climate modeling.
Neural operators are promising surrogates for dynamical systems but when trained with standard L2 losses they tend to oversmooth fine-scale turbulent structures. Here, we show that combining operator learning with generative modeling overcomes this limitation. We consider three practical turbulent-flow challenges where conventional neural operators fail: spatio-temporal super-resolution, forecasting, and sparse flow reconstruction. For Schlieren jet super-resolution, an adversarially trained neural operator (adv-NO) reduces the energy-spectrum error by 15 × while preserving sharp gradients at neural operator-like inference cost. For 3D homogeneous isotropic turbulence, adv-NO trained on only 160 timesteps from a single trajectory forecasts accurately for five eddy-turnover times and offers 114 × wall-clock speed-up at inference than the baseline diffusion-based forecasters, enabling near-real-time rollouts. For reconstructing cylinder wake flows from highly sparse Particle Tracking Velocimetry-like inputs, a conditional generative model infers full 3D velocity and pressure fields with correct phase alignment and statistics. These advances enable accurate reconstruction and forecasting at low compute cost, bringing near-real-time analysis and control within reach in experimental and computational fluid mechanics.
The application of Artificial Neural Networks (ANNs) in developing sensor-based estimators for unsteady flows has become an active area of research over the last decade. One of the challenges in this area is the selection of a low-dimensional subspace that enables the ANN to reconstruct relevant spatiotemporal dynamics in the flow, as both sparsity and interpretability are simultaneously desired. The present study demonstrates the use of a flow-estimation framework based on Long Short-Term Memory (LSTM) neural networks and the time-domain Spectral Proper Orthogonal Decomposition (SPOD) [Sieber et al., “Spectral proper orthogonal decomposition,” J. Fluid Mech. 792, 798–828 (2016)], which was proposed as an extension to the traditional POD. The two-cylinder flow selected for analysis in this study is referred to as the “flip-flop” regime for exhibiting intermittent changes in the phase alignment of the vortices shed by the two cylinders. This flow has dynamics occurring over a wide range of frequencies, and it was selected to demonstrate the usefulness of SPOD in characterizing the wake dynamics of periodic wake flows and to determine if it can provide a better subspace for training and estimation when compared to POD. It was found that a particular SPOD basis obtained with an empirically determined filter length completely separated the frequency centered phenomena present in the spectrum of the most energetic POD modes into different modes. These new SPOD modes were observed to have a direct relationship with the vortex dynamics in the flow, providing direct access to the antiphase and in-phase flow states. The LSTM neural networks estimation capacities were very similar across all the modal spaces investigated, performing well regardless of whether the frequency content of the modal space used for training and estimation was found superimposed in the spectrum of the most energetic modes (POD) or separated into different modes (SPOD). Further investigation is required to determine if this result holds for turbulent flows.
This study investigates the three-dimensional (3D) wake transition in unconfined flows over rectangular cylinders using direct numerical simulation (DNS). Two different cross-sectional aspect ratios (AR) and Reynolds numbers (Re) are scrutinized: AR = 0.5 at Re = 200 and AR = 3 at Re = 600. The investigation focuses on characterizing the flow patterns and forecasting their temporal evolution utilizing the proper orthogonal decomposition (POD) technique coupled with a long short-term memory (LSTM) network. The DNS results reveal the emergence of an ordered mode A for AR = 3, attributed to the stabilizing effect of the elongated AR. On the other hand, the case with a smaller AR (= 0.5) exhibits a mode-swapping regime characterized by modes A and B's distinct and simultaneous manifestation. The spanwise wavelengths of mode A and mode B are approximately 4.7 and 1.2 D for AR = 0.5, while the spanwise wavelength of mode A is 3.5 D for AR = 3. The POD serves as a dimensionality reduction technique, and LSTM facilitates temporal prediction. This algorithm demonstrates satisfactory performance in predicting the flow patterns, including the instabilities of modes A and B, across both transverse and spanwise directions. The employed algorithm adeptly predicts the pressure time series surrounding the cylinders. The duration for training the algorithm is only about 0.5% of the time required for DNS computations. This research, for the first time, demonstrates the effectiveness of the POD–LSTM algorithm in predicting complex 3D instantaneous wake transition patterns for flow past rectangular cylinders.
The objective of this work is to present a coherent detection method to determine relationships between pressure fluctuations and wake velocity in the case of the flow past a cylinder in a turbulent regime. We report experimental results and pressure–velocity statistics measured simultaneously on the surface boundary layer and the near wake in the range of Reynolds numbers ([Formula: see text]). The novelty of the analysis lies in the application of a coherent mean method to determine conditioned wall pressure statistics to velocity events observed in the turbulent cylinder wake. The method allows one to associate wake turbulent events with a corresponding wall-pressure signal profile on the cylinder surface performing discrete angular measurements with a minimum number of pressure sensors. Simultaneous measurements of both time series of wake velocity and wall pressure allow us to characterize the mutual influence of hydrodynamic perturbations from the boundary layer surface pressure with the near wake velocity fluctuations. Large scale turbulent vortex shedding is well correlated with surface pressure fluctuations as expected in the low frequency range. Near wake velocity signals were used as timing signals to determine the coherent influence of a representative velocity pattern on the statistical properties of wall pressure fluctuations. The dynamics of low-frequency large-scale structures is consistent with the presence of counter-rotating vortices and is well correlated with a great part of the pressure fluctuations of the boundary layer. Conditional statistics using the velocity pattern obtained from the coherent averaging method provided the instantaneous pressure fluctuating profile associated with the large-scale vortex shedding cycle, thus, enabling a coherent estimation of the fluctuating lift and drag forces. Conditional statistics on wake velocity are well influenced by the phase of the vortex life-cycle.
Abstract Data-driven Markov linear models of nonlinear fluid flows using maps of the state into a sparse feature space are explored in this article. The underlying principle of low-order models for fluid systems is identifying maps to a feature space where the system evolution (a) is simpler and efficient to model accurately and (b) the state can be recovered accurately from the features through inverse mapping. Such methods are useful when real-time models are needed for online decision making from sensor data. The Markov linear approximation is popular as it allows us to leverage the well established linear systems machinery. Examples include the Koopman operator approximation techniques and evolutionary kernel methods in machine learning. The success of these models in approximating nonlinear dynamical systems is tied to the effectiveness of the feature map in accomplishing both (a) and (b) above as long as the system provides a feasible prediction horizon using data. We assess this by performing an in-depth study of two different classes of sparse linear feature transformations of the state: (i) a pure data-driven POD-based projection that uses left singular vectors of the data snapshots – a staple of common Koopman approximation methods such as Dynamic Mode Decomposition (DMD) and its variants such as extended DMD; and (ii) a partially data-driven sparse Gaussian kernel (sGK) regression (a mean sparse Gaussian Process (sGP) predictor). The sGK/sGP regression equivalently represents a projection onto an infinite-dimensional basis characterized by a kernel in the inner product reproducing kernel Hilbert space (RKHS). We are particularly interested in the effectiveness of these linear feature maps for long-term prediction using limited data for three classes of fluid flows with escalating complexity (and decreasing prediction horizons) starting from a limit-cycle attractor in a cylinder wake followed by a transient wake evolution with a shift in the base flow and finally, a continuously evolving buoyant Boussinesq mixing flow with no well-defined base state. The results indicate that a purely data-driven POD map is good for full state reconstruction as long as the basis remains relevant to the predictions whereas the more generic sparse Gaussian Kernel (sGK) basis is less sensitive to the evolution of the dynamics but prone to reconstruction errors from lack of parsimony. Contrastingly, the sGK-maps outperform POD-based maps in learning the transient nonlinear evolution of the state for the same feature dimension in systems that contain a well-defined attractor(s). Consequently, both POD and sGK-maps require additional layer(s) to help mitigate these shortcomings. For example, POD-maps require nonlinear functional extensions for improved feature space predictions whereas sGK-maps require dimensionality reduction to balance the large feature dimension needed for accurate full state reconstruction. However, both classes of multilayer feature maps fail to predict the highly evolving buoyant mixing flow for very different reasons.
No abstract available
No abstract available
With a view to “nowcasting” of depth-scale flow fields in rivers, estuaries, and the near-coast, this work develops empirical “measurement models” that correlate velocity at a water surface to subsurface flow fields. Free-surface flow over a backstep, as simulated by LES, is chosen as a test case. Proper Orthogonal Decomposition (POD) is first applied to both subsurface and surface velocity fields from a “training data set”, whence we develop regression models between the vectors of surface and subsurface POD coefficients. These models are then applied to estimate the subsurface flow field outside the training data based on virtual “measurements” of streamwise velocity along a streamwise line at the water surface. Good predictive capability is obtained in a region some 20 step heights downstream of the step that is characterized by strong upwellings. INTRODUCTION This paper applies measurement models developed by Mokhasi et al. [1], who were motivated by “micronowcasting” and back-casting of chemical release in a city, and the “multi-time” extensions reported by Nguyen et al. [2]. “Nowcasting” referred originally to detailed, short-term weather forecasting, e.g. to predict the development of thunderstorms. The mathematics are the same as a “traditional” forecasting system, and typically include a nonlinear Kalman Filter (KF) that continually updates the “current estimate” for the vector of variables that specify the state of the system (atmosphere, ocean, etc.) in the chosen dynamical model. Since such physical systems are chaotic, monitoring data must be continually fed back to keep the current estimate close to reality; in the KF this data is compared with the “current state” via a “measurement model”. In the scenario of [1], data from sensors distributed through a cityscape would feed back on a KF designed to track the turbulent flow; then, based on the estimated flow field, chemical concentration measurements would be “back-tracked” to estimate release location(s). As a prototype problem, they considered unidirectional flow past a surface-mounted cube. The measurement models proposed in [1] develop an idea from the fluids engineering literature known as “stochastic estimation”, in which one estimates a flow field, via its POD coefficients, based on a limited number of measurements, typically pressure at one or more surface points. The dominant example has been Linear Stochastic Estimation (LSE). For example, Taylor and Glauser [3] built a low dimensional pressure sensing and control system via POD and LSE for a flow between a backward facing ramp and an adjustable flap; Durgesh and Naughton [4] applied a multitime delay LSE-POD that used pressure to estimate POD coefficients of an unsteady near wake flow. We are motivated by nowcasting of environmental flows with a free surface, notably estuaries and rivers. Existing systems (e.g. the NYHOPS network, [5]) are typically based on sparse point measurements, and cannot resolve horizontal variations much finer than the river width. For the future, we expect that High Frequency radar and/or imaging methods will be able to monitor surface velocity over a certain area. Indeed, highly-resolved data on surface velocity were obtained as early as 1967 by large-format aerial photography when rising flood waters generated surface bubbles as “PIV tracers”, and the patterns of surface divergence even allowed Kinoshita [6] to infer the instantaneous spatial structure of subsurface bedforms. With this justification we investigate here the use of data measured over a portion of free surface as input for stochastic estimation. In the context of stochastic estimation, the introduction of a free surface was the key novelty in our recent communication [7]. That work considered simulated flow over a sinusoidal bed, with a free surface modeled as a stress-free lid. To simplify analysis, we confined attention to the flow velocity in three vertical-streamwise planes, and built POD-based correlations for the subsurface flow based on “measured” surface velocity. In this paper, we also use information on free-surface velocity as input for estimation models. However, we instead study the flow past a backstep as a prototype for a subaqueous dune. Such “bedforms” are important because of their form drag, and because of their key role in sediment dynamics. In addition, the current work applies the correct dynamical conditions at the free surface. The LES of the backward-facing step flow with a free-surface is followed by a brief mathematical introduction of POD. Next, single-time and multi-time POD-based estimations are described and the performances of the POD-based estimations are discussed.
Bayesian filtering provides a principled framework for online state estimation under uncertainty, yet its application to systems with high-dimensional states and complicated posterior distributions remains challenging. Recent generative models, such as flow matching, have shown potential in Bayesian filtering. However, they still rely on particle-based representations of the posterior, which lose the rich information of the full distribution, or tackle a trajectory-level inverse problem that conflicts with the recursive structure of Bayesian filtering. To address this, we propose a new perspective of directly encoding the evolving distribution into flow matching model weights, namely, the Belief Flow Filter (BFF). It is a generative filtering framework that updates model weights via gradient descent at test time to track the posterior evolution. Thereby, BFF bypasses the scalability issue of particle representations or the flexibility limitation of Gaussian assumptions in conventional filters. We theoretically justify that the BFF design is structurally aligned with Bayesian filtering, and its training objective targets the recursive filtering operator. BFF is empirically verified across 5 different physical systems, including ones with chaotic dynamics and highly sparse, non-linear observations. The results show that BFF attains the best score in 8 of 9 metric-benchmark cells across the three standard 1D and 2D PDE benchmarks, and similarly leads on the extreme single-moving-sensor setting and on a real-world-grounded tokamak plasma estimation task, demonstrating its potential to accurately approximate the Bayesian filtering operator in high-dimensional probability space.
Reconstructing high-dimensional spatiotemporal fields from sparse sensor measurements is critical in a wide range of scientific applications. The SHallow REcurrent Decoder (SHRED) architecture is a recent state-of-the-art architecture that reconstructs high-quality spatial domain from hyper-sparse sensor measurement streams. An important limitation of SHRED is that in complex, data-scarce, high-frequency, or stochastic systems, portions of the spatiotemporal field must be modeled with valid uncertainty estimation. We introduce UQ-SHRED, a distributional learning framework for sparse sensing problems that provides uncertainty quantification through a neural network-based distributional regression called engression. UQ-SHRED models the uncertainty by learning the predictive distribution of the spatial state conditioned on the sensor history. By injecting stochastic noise into sensor inputs and training with an energy score loss, UQ-SHRED produces predictive distributions with minimal computational overhead, requiring only noise injection at the input and resampling through a single architecture without retraining or additional network structures. On complicated synthetic and real-life datasets including turbulent flow, atmospheric dynamics, neuroscience and astrophysics, UQ-SHRED provides a distributional approximation with well-calibrated confidence intervals. We further conduct ablation studies to understand how each model setting affects the quality of the UQ-SHRED performance, and its validity on uncertainty quantification over a set of different experimental setups.
Magnetohydrodynamic (MHD) effects play a key role in the design and operation of nuclear fusion systems, where electrically conducting fluids (such as liquid metals or molten salts in reactor blankets) interact with magnetic fields of varying intensity and orientation, which affect the resulting flow. The numerical resolution of MHD models involves highly nonlinear multiphysics systems of equations and can become computationally expensive, particularly in multi-query, parametric, or real-time contexts. This work investigates a fully data-driven framework for MHD state reconstruction that combines dimensionality reduction via Singular Value Decomposition (SVD) with the SHallow REcurrent Decoder (SHRED), a neural network architecture designed to recover the full spatio-temporal state from sparse time-series measurements of a limited number of observables. The methodology is applied to a parametric MHD test case involving compressible lead-lithium flow in a stepped channel subjected to thermal gradients and magnetic fields spanning a broad range of intensities. To improve efficiency, the full-order dataset is first compressed using SVD, yielding a reduced representation used as reference truth for training. Only temperature measurements from three sensors are provided as input, while the network reconstructs the full fields of velocity, pressure, and temperature. To assess robustness with respect to sensor placement, thirty randomly generated sensor configurations are tested in ensemble mode. Results show that SHRED accurately reconstructs the full MHD state even for magnetic field intensities not included in the training set. These findings demonstrate the potential of SHRED as a computationally efficient surrogate modeling strategy for fusion-relevant multiphysics problems, enabling low-cost state estimation with possible applications in real-time monitoring and control.
State estimation is required whenever we deal with high-dimensional dynamical systems, as the complete measurement is often unavailable. It is key to gaining insight, performing control or optimizing design tasks. Most deep learning-based approaches require high-resolution labels and work with fixed sensor locations, thus being restrictive in their scope. Also, doing Proper orthogonal decomposition (POD) on sparse data is nontrivial. To tackle these problems, we propose a technique with an implicit optimization layer and a physics-based loss function that can learn from sparse labels. It works by minimizing the energy of the neural network prediction, enabling it to work with a varying number of sensors at different locations. Based on this technique we present two models for discrete and continuous prediction in space. We demonstrate the performance using two high-dimensional fluid problems of Burgers’ equation and Flow Past Cylinder for discrete model and using Allen–Cahn equation and Convection-diffusion equations for continuous model. We show the models are also robust to noise in measurements.
Solving time-dependent partial differential equations (PDEs) is an important problem in computational science and engineering. Physics-informed neural networks (PINNs) learn PDE solutions from governing equations. However, accurately capturing temporal evolution remains challenging. Recent sequence-model-based approaches parameterize time evolution using general-purpose sequence models, which capture temporal dependencies but do not explicitly encode the structured dynamics of PDE solutions. In addition, their memory requirements can scale unfavorably with sequence length and resolution, limiting applicability in large-scale or high-dimensional settings. This work introduces a PINN approach that incorporates oscillatory state-space dynamics to represent the modal structure of PDE solutions. The proposed method leverages a linear-oscillator-based temporal evolution, together with a PDE-aware spectral basis in space. This design enables closed-form spatial differentiation and consistent enforcement of boundary conditions. The method is evaluated on forward, inverse, and high-dimensional PDE problems, including cases up to 100 spatial dimensions. The results show improved accuracy and reduced memory usage compared to recent sequence-model-based PINN approaches. Overall, this work highlights the benefits of incorporating structured dynamical priors into the temporal evolution of neural PDE solvers and suggests designing more physics-aligned and computationally efficient PINN architectures.
The concept of the path-dependent partial differential equation (PPDE) was first introduced in the context of path-dependent derivatives in financial markets. Its semilinear form was later identified as a non-Markovian backward stochastic differential equation (BSDE). Compared to the classical PDE, the solution of a PPDE involves an infinite-dimensional spatial variable, making it challenging to approximate, if not impossible. In this paper, we propose a neural rough differential equation (NRDE)-based model to learn PPDEs, which effectively encodes the path information through the log-signature feature while capturing the fundamental dynamics. The proposed continuous-time model for the PPDE solution offers the benefits of efficient memory usage and the ability to scale with dimensionality. Several numerical experiments, provided to validate the performance of the proposed model in comparison to the strong baseline in the literature, are used to demonstrate its effectiveness.