Predicting Melt-Crystal Interface Position and Shape during the Manufacturing Process of Single Crystal via Explainable Machine Learning Models

Abstract

In recent years, employing machine learning models to predict the process parameters during the manufacturing process of single crystals has gained wide attention as it’s reliable and much faster than traditional numerical simulation approaches. However, most machine learning models used in previous studies are black box models, which don’t provide explainable results. In this paper, we present a feasibility study of applying explainable machine learning models to predict steady-state melt-crystal interface position and deflection with the set-point temperature of 5 heaters in a vertical Bridgman furnace. The dataset used to train and evaluate the machine learning models was generated by 2-D numerical simulation. We experimented with linear regression and random forest algorithms, and then used linear regression coefficient and SHAP value to quantify the impact of each input on the output, from which we inferred a heater control strategy that could potentially improve the crystal growth process. Our encouraging results show that explainable machine learning models can be applied to predict crystal growth process parameters in real-time and generate actionable insights to guide crystal manufacturing practice.