ProPINN: Demystifying Propagation Failures in Physics-Informed Neural Networks