An operator preconditioning perspective on training in physics-informed machine learning