An approximate Riemann solver approach in Physics-Informed Neural Networks for hyperbolic conservation laws

Abstract

This study enhances the application of Physics-Informed Neural Networks (PINNs) for modeling discontinuous solutions in both hydrodynamics and relativistic hydrodynamics. Conventional PINNs, trained with partial differential equation residuals, frequently face convergence issues and lower accuracy near discontinuities. To address these issues, we build on the recently proposed locally linearized PINNs (LLPINNs), which improve shock detection by modifying the Jacobian matrix resulting from the linearization of the equations, only in regions where the velocity field exhibits strong compression. However, the original LLPINN framework required a priori knowledge of shock velocities, limiting its practical utility. We present a generalized LLPINN method that dynamically computes shock speeds using neighboring states and applies jump conditions through entropy constraints. Additionally, we introduce locally Roe PINNs (LRPINNs), which incorporate an approximate Roe Riemann solver to improve shock resolution and conservation properties across discontinuities. These methods are adapted to two-dimensional Riemann problems by using a divergence-based shock detection combined with dimensional splitting, delivering precise solutions. Compared to a high-order weighted essentially non-oscillatory solver, our method produces sharper shock transitions but smoother solutions in areas with small-scale vortex structures. Future research will aim to improve the resolution of these small-scale features without compromising the model's ability to accurately capture shocks.

20 pages, 14 figures. Accepted version of the paper published in Physics of Fluids

References in corpus (14)

• DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators
• Artificial Neural Networks for Solving Ordinary and Partial Differential Equations
• On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks
• Finite Basis Physics-Informed Neural Networks (FBPINNs): a scalable domain decomposition approach for solving differential equations
• A Comparison of Optimization Algorithms for Deep Learning
• Residual-based attention in physics-informed neural networks
• Physics Informed Neural Networks for Simulating Radiative Transfer
• Discontinuity Computing using Physics-Informed Neural Network
• Multilevel domain decomposition-based architectures for physics-informed neural networks
• Unveiling the optimization process of Physics Informed Neural Networks: How accurate and competitive can PINNs be?
• Least-Squares ReLU Neural Network (LSNN) Method For Linear Advection-Reaction Equation
• Solving the Teukolsky equation with physics-informed neural networks
• An Efficient Implementation of Flux Formulae in Multidimensional Relativistic Hydrodynamical Codes
• Least-Squares Neural Network (LSNN) Method For Linear Advection-Reaction Equation: Discontinuity Interface