Prediction of crystalline structure evolution during solidification of aluminum at different cooling rates using a hybrid neural network model

The solidification behavior significantly influences the mechanical properties of pure metals, where cooling rates play a pivotal role in the evolution of crystal structure, nucleation processes, and the subsequent development of stacking faults and dislocations. This research introduces a novel approach that employs a hybrid deep learning architecture, combining 1-dimensional convolutional neural network (Conv1D) layers with long short-term memory (LSTM) units, aimed at forecasting the crystallization behavior of pure aluminum throughout the solidification process. By combining the temporal pattern descriptors of LSTMs with the feature extraction potential of convolutional neural networks (CNN), the hybrid Conv1D-LSTM model achieves higher accuracy in predicting crystal structural evolution curves, in contrast to the performance of standalone LSTM and CNN models. The model was trained using a dataset generated from 240 molecular dynamics (MD) simulations conducted at 48 different cooling rates, with cooling rates, timesteps, and temperature as inputs. The output consists of the percentage distribution of body-centered cubic (BCC), face-centered cubic (FCC), hexagonal close-packed (HCP), icosahedral (ICO), and liquid or amorphous solid structures. The performance of the hybrid model is assessed using mean squared error (MSE), mean absolute error (MAE), and the coefficient of determination (R²). The results demonstrate high accuracy, robust handling of complex patterns, and strong generalization of the hybrid model to unseen data, both within and beyond the training range, confirmed by R² value exceeding 0.99 within the training range.