Physics-Informed Neural Networks with Adaptive Localized Artificial Viscosity

TL;DR

This paper introduces three adaptive methods for artificial viscosity in physics-informed neural networks to better handle shock solutions in hyperbolic PDEs, improving accuracy without prior viscosity tuning.

Contribution

It proposes novel adaptive localized AV approaches for PINNs that automatically learn viscosity values and shock locations, enhancing solution accuracy in stiff PDE problems.

Findings

Adaptive methods accurately locate shocks and learn small AV values.

Proposed approaches outperform nonadaptive global AV methods.

Methods successfully applied to Burgers and Buckley-Leverett equations.

Abstract

Physics-informed Neural Network (PINN) is a promising tool that has been applied in a variety of physical phenomena described by partial differential equations (PDE). However, it has been observed that PINNs are difficult to train in certain "stiff" problems, which include various nonlinear hyperbolic PDEs that display shocks in their solutions. Recent studies added a diffusion term to the PDE, and an artificial viscosity (AV) value was manually tuned to allow PINNs to solve these problems. In this paper, we propose three approaches to address this problem, none of which rely on an a priori definition of the artificial viscosity value. The first method learns a global AV value, whereas the other two learn localized AV values around the shocks, by means of a parametrized AV map or a residual-based AV map. We applied the proposed methods to the inviscid Burgers equation and the Buckley-Leverett equation, the latter being a classical problem in Petroleum Engineering. The results show that the proposed methods are able to learn both a small AV value and the accurate shock location and improve the approximation error over a nonadaptive global AV alternative method.