[arXiv]score: 0.53
NPSolver Trains Neural Poisson Solver Without Solution Labels Using PCG Supervision
May 26, 2026
NPSolver uses a small number of preconditioned conjugate gradient steps to generate stable training signals for a neural Poisson solver on irregular domains, avoiding both large labeled datasets and unstable raw PDE residual losses.
cs.LGcs.AI
HOW THIS AFFECTS YOU
●
researcherThe iterative physics supervision approach with stop-gradient design offers a practical training recipe for physics-informed neural operators that sidesteps the instability of direct residual minimization.