Convergence Analysis of Natural Gradient Descent for Over-parameterized Physics-Informed Neural Networks