Integrating machine learning and numerical analysis to characterise stress fields around structural defects in steel alloys

Abstract

This study investigates the application of machine learning models – specifically convolutional neural networks (CNNs), hybrid CNN–recurrent neural network (RNN) and CNN–graph neural network (GNN) architectures – for predicting stress fields in steel plates containing arbitrarily shaped pores under tensile loading. Among the models tested, CNN–GNN achieved the highest accuracy, with an R² of 0.95 and a root mean square error (RMSE) of 4.93%, outperforming CNNs (R² = 0.88, RMSE = 10.93%) and CNN–RNN (R² = 0.92, RMSE = 7.41%). This advantage stems from CNN–GNN's graph-based architecture, which effectively models complex, non-linear stress distributions around irregular pores. Nevertheless, prediction accuracy decreases with larger pore sizes, increased pore count or when pores are near plate edges. Additionally, the CNN–GNN model performs most effectively with medium-strength alloys such as stainless steel 304, where stress distributions are more predictable. In contrast, its accuracy diminishes when applied to high-strength alloys like AISI D2 or low-strength alloys like ASTM A36. These variations reflect the unique mechanical properties of each alloy, which introduce greater complexity into the stress distribution and challenge the model's ability to generalise across different materials.