A comprehensive review of advances in physics-informed neural networks and their applications in complex fluid dynamics

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(10)

Published: Oct. 1, 2024

Physics-informed neural networks (PINNs) represent an emerging computational paradigm that incorporates observed data patterns and the fundamental physical laws of a given problem domain. This approach provides significant advantages in addressing diverse difficulties field complex fluid dynamics. We thoroughly investigated design model architecture, optimization convergence rate, development modules for PINNs. However, efficiently accurately utilizing PINNs to resolve dynamics problems remain enormous barrier. For instance, rapidly deriving surrogate models turbulence from known characterizing flow details multiphase fields present substantial difficulties. Additionally, prediction parameters multi-physics coupled models, achieving balance across all scales multiscale modeling, developing standardized test sets encompassing dynamic are urgent technical breakthroughs needed. paper discusses latest advancements their potential applications dynamics, including turbulence, flows, multi-field flows. Furthermore, we analyze challenges face these outline future trends growth. Our objective is enhance integration deep learning facilitating resolution more realistic problems.

Language: Английский

---

A novel complex variable hyper-reduction model for efficient determination of convective heat transfer coefficient at the inlet of a steam turbine

International Communications in Heat and Mass Transfer, Journal Year: 2025, Volume and Issue: 162, P. 108587 - 108587

Published: Jan. 8, 2025

Language: Английский

---

The Application of Physics-Informed Machine Learning in Multiphysics Modeling in Chemical Engineering

Industrial & Engineering Chemistry Research, Journal Year: 2023, Volume and Issue: 62(44), P. 18178 - 18204

Published: Oct. 26, 2023

Physics-Informed Machine Learning (PIML) is an emerging computing paradigm that offers a new approach to tackle multiphysics modeling problems prevalent in the field of chemical engineering. These often involve complex transport processes, nonlinear reaction kinetics, and coupling. This Review provides detailed account main contributions PIML with specific emphasis on momentum transfer, heat mass reactions. The progress method development (e.g., algorithm architecture), software libraries, applications coupling surrogate modeling) are detailed. On this basis, future challenges highlight importance developing more practical solutions strategies for PIML, including turbulence models, domain decomposition, training acceleration, modeling, hybrid geometry module creation.

Language: Английский

---

Practical uncertainty quantification for space-dependent inverse heat conduction problem via ensemble physics-informed neural networks

International Communications in Heat and Mass Transfer, Journal Year: 2023, Volume and Issue: 147, P. 106940 - 106940

Published: July 20, 2023

Language: Английский

---

Physics-informed graph neural network based on the finite volume method for steady incompressible laminar convective heat transfer

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(1)

Published: Jan. 1, 2025

The rapid development of deep learning has significantly influenced computational studies in convective heat transfer. To facilitate broader applications models transfer, this paper proposes a physics-informed graph neural network based on the finite volume method (FVGP-Net) for unsupervised training and prediction steady incompressible laminar transfer problems. In model, mesh data generated by (FVM) are converted into data, preserving mesh's topological properties. This conversion allows FVGP-Net to utilize convolutional information aggregation, capturing both local global flow features enhancing model's geometric adaptability predictive performance. model incorporates physical laws directly its loss function, ensuring compliance these without reliance data. Unlike traditional networks (PINNs), replaces automatic differentiation with FVM-based numerical differentiation, balancing efficiency accuracy. Boundary conditions handled accordance FVM, that strictly satisfies constraints. We validated using representative test cases, also examining effects different initialization methods training. results demonstrate achieves high accuracy predicting Compared PINNs, inherits conservation properties velocity problems 70.03%. Furthermore, application markedly accelerates training, achieving approximately 70% faster compared Xavier initialization.

Language: Английский

---

Rapid prediction of indoor airflow field using operator neural network with small dataset

Building and Environment, Journal Year: 2024, Volume and Issue: 251, P. 111175 - 111175

Published: Jan. 20, 2024

Language: Английский

---

A finite element-based physics-informed operator learning framework for spatiotemporal partial differential equations on arbitrary domains

Engineering With Computers, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 2, 2024

Abstract We propose a novel finite element-based physics-informed operator learning framework that allows for predicting spatiotemporal dynamics governed by partial differential equations (PDEs). The Galerkin discretized weak formulation is employed to incorporate physics into the loss function, termed (FOL), along with implicit Euler time integration scheme temporal discretization. A transient thermal conduction problem considered benchmark performance, where FOL takes temperature field at current step as input and predicts next step. Upon training, network successfully evolution over any initial high accuracy compared solution element method (FEM) even heterogeneous conductivity arbitrary geometry. advantages of can be summarized follows: First, training performed in an unsupervised manner, avoiding need large data prepared from costly simulations or experiments. Instead, random patterns generated Gaussian process Fourier series, combined constant fields, are used cover possible cases. Additionally, shape functions backward difference approximation exploited domain discretization, resulting purely algebraic equation. This enhances efficiency, one avoids time-consuming automatic differentiation optimizing weights biases while accepting discretization errors. Finally, thanks interpolation power FEM, geometry microstructure handled FOL, which crucial addressing various engineering application scenarios.

Language: Английский

---

Fusion of theory and data-driven model in hot plate rolling: A case study of rolling force prediction

Expert Systems with Applications, Journal Year: 2023, Volume and Issue: 245, P. 123047 - 123047

Published: Dec. 28, 2023

Language: Английский

---

Deep Learning Method Based on Physics-Informed Neural Network for 3D Anisotropic Steady-State Heat Conduction Problems

Mathematics, Journal Year: 2023, Volume and Issue: 11(19), P. 4049 - 4049

Published: Sept. 24, 2023

This paper uses the physical information neural network (PINN) model to solve a 3D anisotropic steady-state heat conduction problem based on deep learning techniques. The embeds problem’s governing equations and boundary conditions into treats network’s output as numerical solution of partial differential equation. Then, is trained using Adam optimizer training set. progressively converges toward accurate In first example, we demonstrate convergence PINN by discussing effect number layers, each hidden layer’s neurons, initial rate decay rate, size set, mini-batch size, amount points boundary, steps relative error solution, respectively. solutions are presented for three different examples. Thus, effectiveness method verified.

Language: Английский

---

Hybrid Physics-Data-Driven Model for Temperature Field Prediction of Asphalt Pavement Based on Physics-Informed Neural Network

Published: Jan. 1, 2025

Traditional models for pavement temperature prediction are either physics-driven or data-driven, which fail to account deviations between physical law and reality, dependent on data size, thereby limiting accuracy. The physics-informed neural network (PINN) was adopted fuse observed temperatures with the law. Initially, heat transfer differential equations asphalt were derived. Subsequently, a hybrid sampling method, two subdomain decomposition modes, boundary initial condition interpolation method proposed address problem. PINN-based physics-data-driven developed. Finally, effect of PINN parameters predicted analyzed, followed by comparisons model accuracy data-driven models. results indicate that mean absolute percentage error field is below 0.08% when utilizing parallel suitable parameters. By fusing data, can predict without requiring convection coefficient. loss weight should be proportional number points. outperformed using sparsely data.

Language: Английский

---