Recurrent neural network based on attention mechanism in prediction of glass forming ability by element proportion

Abstract

The traditional trial and error method is time-consuming and laborious to study glass forming ability (GFA), so it is necessary to design a fast and accurate method to predict GFA. In this paper, four deep learning models of BidiRNN, BidiLSTM, Attention-BidiRNN and Attention-BidiLSTM are constructed by using recurrent neural network (RNN) and Attention mechanism, and the GFA of alloys is directly predicted by element proportion. After model selection and model evaluation through cross-validation, the prediction accuracy of the four models on the test set is 0.727, 0.796, 0.822 and 0.843, respectively, which has good generalization performance. Since amorphous alloys usually have high similarity to similar alloy systems, the attention mechanism can dynamically identify alloys with high similarity to the current prediction task. By weight matching, the feature extraction ability of the model is enhanced, and the prediction accuracy of Attention-BidiRNN and Attention-BidiLSTM is 9.05% and 4.7% higher than that of BidiRNN and BidiLSTM, respectively. Finally, two systems of CuZrAl and CuCeGa are predicted in this paper, and the reliability of the model is verified by experimental data. This realization of predicting GFA only by combination of elements is expected to provide important guidance for the development and preparation of novel amorphous alloys.