Open Internet by MindsNet

Training Pathologies in Physics-Informed Neural Networks

Physics-Informed Neural Networks (PINNs) often fail to train due to imbalanced gradients from different loss components and poor convergence on PDE residual terms. This leads to difficulties in learning interior physics and boundary conditions. Several mitigation strategies exist but a full solution remains elusive.

Mathematics & logic, Mathematics, Differential Equations