NPSolver: Neural Poisson Solver with Iterative Physics Supervision

Researchers introduced NPSolver, a neural Poisson solver trained without ground-truth labels by using a few preconditioned conjugate gradient steps as supervision.

The method replaces raw residual losses with iterative physics-based guidance, aiming for more stable training on PDE problems.

It also adds BA-Transolver, a boundary-aware design for mixed boundary conditions and irregular geometries.

NPSolver outperforms physics-informed and data-driven baselines on 2D and 3D benchmarks and supports a thermal control application.