Combination of ab initio descriptors and machine learning approach for the prediction of the plasticity mechanisms in β-metastable Ti alloys

Abstract

This study proposes a machine-learning (ML) model combining ab initio calculations and an experimental dataset of 201 alloys (in addition to pure Ti) to predict the activated plasticity mechanisms in β-Ti alloys. This methodology is shown to be more efficient than the so-called Bo¯-Md¯ approach, achieving 82% prediction accuracy while the Bo-Md approach leads to 52% correct predictions on the same dataset. In addition, four new alloy compositions were produced to verify the model validity. Specific cases where the present model disagreed with the Bo-Md predictions were chosen to increase the benefits of the produced results. The plasticity mechanisms of the four alloys experimentally confirmed the validity of the ML model. This approach particularly helps the design of specific Ti alloys exhibiting a high work hardening rate owing to the simultaneous activations of the Transformation-Induced Plasticity (TRIP) and mechanical twinning (TWIP) effects. Indeed, the class corresponding to the combination of TRIP and TWIP effects reach a prediction accuracy of 88%.