Stochastic projection based approach for gradient free physics informed learning

Abstract

We propose a stochastic projection-based gradient free physics-informed neural network. The proposed approach, referred to as the stochastic projection based physics informed neural network (SP-PINN), blends upscaled stochastic projection theory with the recently proposed physics-informed neural network. This results in a framework that is robust and can solve problems involving complex solution domain and discontinuities. SP-PINN is a gradient-free approach which addresses the computational bottleneck associated with automatic differentiation in conventional PINN. Efficacy of the proposed approach is illustrated by a number of examples involving regular domain, complex domain, complex response and phase field based fracture mechanics problems. Case studies by varying network architecture (activation function) and number of collocation points have also been presented.