Spectral-Bias and Kernel-Task Alignment in Physically Informed Neural Networks