FK-PINNs tame ill-conditioned loss with Feynman-Kac supervision
New method solves tough PDEs where standard PINNs fail completely.
Deep Dive
FK-PINNs augment Physics-Informed Neural Networks with pointwise labels from Feynman-Kac Monte Carlo averages. This supervision acts as an operator-level preconditioner, guaranteeing a substantially smaller condition number than under the standard PINN loss. The authors derive non-asymptotic L² error bounds for tanh networks trained with gradient descent. Experiments on Poisson, Schrödinger, mean exit time, and committor problems show FK-PINNs succeed where standard PINNs exhibit severe failure modes.
Key Points
- FK-PINNs add pointwise Feynman-Kac labels as operator preconditioner, reducing loss condition number.
- Non-asymptotic L² error bounds proven for tanh networks trained with gradient descent.
- Beats standard PINNs on Poisson, Schrödinger, and committor problems where standard methods fail.
Why It Matters
Makes PINNs viable for stiff PDEs, enabling robust neural solvers in science and engineering.