Development of Physics-Informed Neural Networks for Solving Partial Differential Equations
- 30 June 2025
- conference paper
- Published by Academy of Cognitive and Natural Sciences
- p. 140-153
Abstract
This study evaluates the effectiveness of physics-informed neural networks (PINNs) for solving both stationary and non-stationary partial differential equations (PDEs), including those with Robin boundary conditions, in rectangular and non-rectangular domains. Although only a small subset of the PINN literature examines mixed boundaries or tackles non-rectangular geometries, and even fewer studies benchmark accuracy against the finite-element method (FEM), the present work provides precisely that comparison. We test feedforward PINNs with tanh activation, whose depth and width were empirically selected for each benchmark to balance accuracy and training cost; training uses Adam optimization with Glorot initialization. These networks are evaluated on three problems: a 2-D Laplace equation on a square (Dirichlet–Neumann– Robin), the same equation on a doubly connected domain with Dirichlet boundary conditions, and a 1-D non- stationary heat equation with Robin boundaries. A weighted mean-squared residual, evaluated via automatic differentiation in TensorFlow, balances equation, boundary, and initial-time terms, thereby handling non- stationary problems without a separate time-stepping scheme. Within the tested class, linear, second-order parabolic and elliptic PDEs in 1-D and 2-D, the network attains ≤ 3 % l ∞ error relative to analytical or FEM solutions after 4–6 min of training on an RTX 3080 Ti Laptop GPU, matching FEM accuracy while eliminating meshing and easing equation and boundary changes. The time to compute a standard PINN solution is longer than for a FEM solution for problems considered in the research, and a broad literature review reveals theoretical convergence limits that constrain standard PINNs to modest-scale, well-conditioned diffusion problems.Keywords
This publication has 23 references indexed in Scilit:
- Tackling the curse of dimensionality with physics-informed neural networksNeural Networks, 2024
- Enforcing continuous symmetries in physics-informed neural network for solving forward and inverse problems of partial differential equationsJournal of Computational Physics, 2023
- FDM-PINN: Physics-informed neural network based on fictitious domain methodInternational Journal of Computer Mathematics, 2022
- Exact imposition of boundary conditions with distance functions in physics-informed deep neural networksComputer Methods in Applied Mechanics and Engineering, 2022
- A physics-informed machine learning approach for solving heat transfer equation in advanced manufacturing and engineering applicationsEngineering Applications of Artificial Intelligence, 2021
- Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equationsJournal of Computational Physics, 2019
- Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential EquationsarXiv, 2017
- Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential EquationsarXiv, 2017
- Adam: A Method for Stochastic OptimizationarXiv, 2017
- Multilayer feedforward networks are universal approximatorsNeural Networks, 1989