Characterizing and Mitigating the Difficulty in Training Physics-informed Artificial Neural Networks under Pointwise Constraints

Neural networks can be used to learn the solution of partial differential equations (PDEs) on arbitrary domains without requiring a computational mesh. Common approaches integrate differential operators in training neural networks using a structured loss function. The most common training algorithm for neural networks is backpropagation which relies on the gradient of the loss function with respect to the parameters of the network. In this work, we characterize the difficulty of training neural networks on physics by investigating the impact of differential operators in corrupting the back propagated gradients. Particularly, we show that perturbations present in the output of a neural network model during early stages of training lead to higher levels of noise in a structured loss function that is composed of high-order differential operators. These perturbations consequently corrupt the back-propagated gradients and impede convergence. We mitigate this issue by introducing auxiliary flux parameters to obtain a system of first-order differential equations. We formulate a non-linear unconstrained optimization problem using the augmented Lagrangian method that properly constrains the boundary conditions and adaptively focus on regions of higher gradients that are difficult to learn. We apply our approach to learn the solution of various benchmark PDE problems and demonstrate orders of magnitude improvement over existing approaches.