Material-Response-Informed DeepONet and Its Application to Polycrystal Stress–Strain Prediction in Crystal Plasticity

Crystal plasticity (CP) model is a vital tool for understanding structure–property relations, but it is computationally expensive. Hence, data-driven models have been used as surrogate. We proposed a Deep Operator Network (DeepONet) to predict polycrystal stress–strain response. It employs a convolutional network to encode microstructure. To account for different material properties and boundary conditions, we proposed using single-crystal responses as inputs to the branch, furnishing a material-response-informed DeepONet. This is the most novel contribution. We demonstrate that our model can be trained on one material and loading and generalized to new conditions via transfer learning. Results show that using single-crystal responses as input outperforms a similar model using material properties and overcomes limitations with changing boundary conditions. The new model achieved a \(R^2\) $R^2$ value of above 0.99, and over 95% of predicted stresses have a relative error of \(\le \) $\le$ 5%, indicating superior accuracy. With as few as 20 new data and under 1 min training time, the DeepONet can be fine-tuned to generate accurate predictions on different materials and loadings. The prediction speed is \(10^{4}\) ${10}^4$ times faster than CP simulations. The efficiency and generalizability of DeepONet render it a powerful data-driven surrogate to bridge scale gaps in multi-scale analyses.