Integrated PredRNN++ and phase-field for microstructure prediction and simulation acceleration

Deep learning methods are integrated with phase-field (PF) simulations to predict microstructural evolution and accelerate PF simulations, thereby overcoming the high computational cost associated with multi-parameter, multi-component parameter optimization. The predictive recurrent neural network++ (PredRNN++) model is developed to predict the precipitation morphologies with circular, cubic and mixed-shape accompanying the growth and coarsening process in superalloys. Compared with convolutional long short-term memory, predictive recurrent neural network, and convolutional neural network models, the PredRNN++ model shows superior prediction performance in precipitate microstructural evolution. Moreover, the hybrid model of PredRNN++ and PF is developed to accelerate the prediction of precipitate evolution, the computational efficiency of the hybrid model exceeds 2.9 times that of the single PF model. In the hybrid model, the calculation speed of PredRNN++ is 40 times that of the PF, and the mean square error remains within 0.004. In particular, it enables the analysis of precipitation kinetics in mixed-shaped precipitates. The advantages of the PredRNN++ model are discovered in predicting the precipitate microstructures, accompanied by Ostwald ripening and coalescence coarsening mechanisms. This study advances the applicability of deep learning to complex microstructural evolution and the feasibility of accelerating PF simulations.