Design of continuous structures using physics informed neural networks

Abstract

Since their inception, physics informed neural networks (PINNs) have shown advantages over traditional data-driven only models by reducing the need for large training dataset sizes via the embedment of physical information in the training process targeting predictions consistent with the laws of physics. This work investigates the viability and advantages of PINN models used for designing continuous structures. To achieve this, local stiffness relations based on Timoshenko–Ehrenfest beam theory are used to constrain network output spaces for predicting variable member sizes in continuous beams. The results herein indicate that PINNs have improved performance relative to data-driven only design models, but the present correlation between data-losses and physics-losses is tenuous, limiting the verifiability of PINN models during inference. Future avenues of research to develop a extensively verifiable architectures are suggested.