Numerically-Inspired Novel Loss Functions for Physics-Informed Neural Networks of Hyperbolic Conservation Laws
Abstract
Physics-Informed Neural Networks (PINNs) present a compelling mesh-free, data-light alternativefor solving partial differential equations, however, they encounter significant difficulties whenapplied to hyperbolic conservation laws that develop shocks and discontinuities. Classicalnumerical techniques specifically, Essentially Non-Oscillatory (ENO) schemes, Total VariationDiminishing (TVD) limiters, and entropy-stability conditions have long been employed to addressthese challenges. In this work, we introduce five differentiable loss-function components thattranslate these established numerical ideas directly into the PINN optimization objective:(i) Arc-Length Regularization, drawn from ENO smoothness indicators,(ii) TVD Loss, which penalizes spurious oscillations,(iii) Conservation Loss, which enforces global integral invariants,(iv) Shock-Aware Smoothness, which adaptively weights second-derivative penalties in thevicinity of discontinuities, and(v) Entropy-Stable Loss, which imposes physical admissibility through the entropy inequality.Six PINN variants are evaluated against three canonical benchmarks: linear advection with asmooth initial profile, and the Burgers equation with shock formation. The proposed lossfunctions consistently reduce final PDE residuals and improve solution accuracy, at the costof only a 20–40\% increase in computational overhead. This framework draws a principledconnection between classical numerical analysis and modern deep learning, offering a robustpath toward reliable PINNs for hyperbolic problems.