SPARC-Net: A Spectral, Causality-Aware, and Hard-Constrained Physics-Informed Architecture for Stiff and Shock-Dominated Partial Differential Equations

Research Question and Approach

Physics-Informed Neural Networks (PINNs) are powerful mesh-free solvers for partial differential equations (PDEs), but they often fail on stiff or shock-dominated problems. These failures occur because small residuals in the PDE do not always correspond to accurate global solutions. The authors identify four distinct, interacting pathologies: spectral bias (difficulty learning sharp features), loss-weight collapse (imbalanced optimization), violation of temporal causality, and under-resolved collocation.

To address these, the authors propose SPARC-Net, a co-designed framework that integrates several architectural and training innovations. Key components include a multi-scale spectral encoder with learnable gating, a gated residual backbone, and a hard-constraint output ansatz that enforces initial and boundary conditions algebraically. By design, this hard-constraint approach eliminates the need for soft-constraint loss terms, structurally preventing the loss-weight collapse common in standard PINNs.

SPARC-Net was validated against exact analytic solutions for four challenging benchmarks: the viscous Burgers' equation, the Allen-Cahn equation, a high-speed convection equation, and a reaction equation. The results demonstrate substantial improvements over vanilla PINNs and other specialized baselines. For instance, the relative L2 error on the convection equation was reduced by 100% (from 5.14e-1 to 9.88e-5), and the Allen-Cahn error was reduced by 94%. The authors also provide ablation studies and Wilcoxon significance tests to confirm that the joint approach is more effective than addressing these pathologies in isolation.

This work is significant because it moves beyond the "one-size-fits-all" approach to PINN training. By systematically identifying and addressing the multi-causal nature of PINN failure, SPARC-Net provides a robust, unified framework for solving complex physical systems that were previously intractable for standard neural solvers. The inclusion of a characteristic-coordinate encoder for hyperbolic transport further highlights the framework's adaptability to specific physical regimes.