Predicting grain growth in polycrystalline materials using deep learning time series models

Grain growth strongly influences the mechanical behavior of materials, making its prediction a key objective in microstructural engineering. In this study, several deep learning approaches are evaluated, including recurrent neural networks (RNN), long short-term memory (LSTM), temporal convolutional networks (TCN), and Transformers, to forecast grain size distributions during grain growth. Unlike computationally demanding full-field simulations, the present work relies on normalized mean-field statistical descriptors extracted from validated high-fidelity simulations. A dataset of 112 grain growth sequences is processed into multivariate time series and used to train autoregressive models capable of predicting microstructural evolution from a short temporal history. Across all evaluation scenarios, including long-horizon forecasting, unseen initial distributions, and spatial domain transfer, the LSTM architecture consistently demonstrates the highest accuracy and stability, achieving prediction accuracies around 90% while preserving physically consistent grain growth trends. Once trained, the models generate full grain growth trajectories in only a few seconds, compared to about 20 min for an optimized and highly efficient partial differential equations-based simulation, achieving speedups exceeding two orders of magnitude. These results highlight the potential of low-dimensional descriptors combined with LSTM-based forecasting for efficient and physically consistent microstructure prediction, with direct implications for accelerated materials modeling, digital twin development, and process optimization.