Predicting the catastrophic failure of bulk metallic glasses based on time-series prediction models

Abstract

The plastic deformation of bulk metallic glasses (BMGs) is characterized by serrated plastic flows, leading to catastrophic failures, while the prediction of such catastrophic failures is still challenging. In this work, three basic deep learning neural network models, including the Long Short-Term Memory (LSTM), Transformer and Gate Recurrent Unit (GRU), as well as two improved models, the LSTM+Transformer (LSTM+T) and GRU+Convolutional Neural Networks (CNN), were used to predict the plastic flow information of BMGs with different sample aspect ratios and compression rates. The Pearson correlation coefficients among nine parameters, showing the correlations between the load drop and elastic energy accumulation rate, were calculated by heatmap. This also aids in the selection of features and prediction targets in subsequent model training, helping to reduce overfitting. Considering factors such as training set size, model applicability, and prediction accuracy, the failure of BMGs was monitored and predicted in real time from two perspectives: multiple and single data sets. The predictability of load drop and the strain of load drop initiation were observed from the results of multiple sets of data, and the LSTM model can predict their development effectively. Subsequently, the LSTM model was trained specifically using a single set of data, where the improvement on the predictions for peak stress, and the strain of the load drops initiation was achieved. This work provides a new method that predicts the catastrophic failure of BMGs and the damage characteristics for similar solids.