Rapid estimation of residual stress in composite laminates using a deep operator network

Abstract

A deep operator network (DeepONet) is designed and developed for rapid estimation of residual stress in composite laminates, which traditionally requires intensive finite element method (FEM) calculations to calibrate the incremental hole-drilling (IHD) method used in measuring residual stresses. The proposed DeepONet model incorporates graph convolution, trigonometric series expansion, and Monte Carlo dropout to effectively learn the relationship between residual stress distribution and the corresponding deformation observed in the IHD procedure. This learning is based on FEM data from various symmetric composite laminate configurations, which are composed of eight layers of fiber-reinforced plates with possible ply orientations at $-45°$, $0°$, $45°$, and $90°$. Trained on 30 configurations, the proposed model exhibits strong generalization capabilities over an additional 40 unseen configurations, achieving a forward strain prediction error of 1.59% and an inverse stress calculation error of 3.92%. These errors are within the range of experimental noise and corresponding stress uncertainty levels commonly encountered in real experiments. The performance of the model suggests the potential for establishing a comprehensive database for the IHD method as applied to composite materials, filling a significant gap in resources when compared to those available for metallic materials.