Taming the Loss Landscape of PINNs with Noisy Feynman-Kac Supervision: Operator Preconditioning and Non-Asymptotic Error Bounds
Researchers have developed a new method to improve the training of Physics-Informed Neural Networks (PINNs) for solving complex partial differential equations (PDEs). The technique, termed "FK-PINNs," introduces a data-fidelity term to the standard PINN loss function, acting as an operator-level preconditioner. This approach is shown to significantly reduce the condition number of the loss landscape, enabling convergence where standard PINNs fail. The method leverages Monte Carlo averages of Feynman-Kac functionals to generate labels and provides non-asymptotic error bounds for networks with tanh activations. AI
IMPACT Introduces a novel technique to enhance the stability and performance of neural networks used in scientific simulations.