Deep learning approach for predicting multi-component stress fields in fiber-reinforced composites under different load paths

Abstract

Fiber-reinforced composites are widely used in various fields due to their excellent performance, and in-depth analysis of their stress fields is crucial for improving material properties and optimizing mechanical structures. However, the traditional analytical and numerical analysis approaches are still limited by fixed input loading and limited fiber volume fractions. To address this challenge, this paper presents a deep learning (DL) framework that enables rapid and accurate prediction of multi-component stress fields for representative volume element (RVE) geometries of fiber composites, considering various fiber volume fractions and different input load paths. The framework is developed based on the 3D TransU-Net framework, which incorporates transformer layer and effectively captures both local and global features of samples. By utilizing randomly distributed RVE geometrical microstructures, the stress fields at diverse fiber volume fractions can be accurately predicted. To adapt different load paths, transfer learning is integrated to fine-tune the weights of pre-training model. Several performance metrics, including relative error (RE) and coefficient of determination ($R^2$), are selected to validate the accuracy of stress distribution predictions. Additionally, a series of results demonstrated the superiority of transfer learning using the same training and validation datasets, and further tests confirmed the model's robustness when faced with unseen samples with diverse volume fractions.