arxivcs.LGcs.AI2026-07-15
LIGO-PINN: Learned Initialization via Gated Optimization to Alleviate Convergence Failures in Physics Informed Neural Networks
Nilay Anurag, Shital Adhikari, Taniya Kapoor, Nikhil Muralidhar
Physics-informed neural networks (PINNs) have had a broad research impact in modeling domains governed by partial differential equations (PDE). However, PINNs have been shown to perform poorly, sometimes even converging to trivial solutions, in challenging PDE domains, or when ge…