Yield Prediction Model for Ingot Samples Based on Machine Learning and Data Augmentation

The preparation of high-performance radiation detector materials such as cadmium zinc telluride (CZT) relies on rigorous and efficient quality control to ensure the consistency of device performance. Traditional manual evaluation based on wafer-by-wafer inspection is time-consuming and makes it difficult to assess the downstream product yield at the ingot level in advance. This paper proposes a machine-learning-based prediction framework for CZT ingots, in which the product-level yield of test wafers from the same ingot is predicted using the double-sided electrical performance and spectral characterization data of a limited number of evaluation wafers. To address the limited number of ingot samples and the significant internal variability among wafers, statistical aggregate features, A/B-side difference features, threshold-ratio features, and intra-ingot Bootstrap resampling were combined, and multiple regression methods, including linear models, Random Forest, XGBoost, and neural networks, were systematically evaluated. The results show that the XGBoost model achieved the best overall performance, with the lowest mean squared error of 0.0352, a mean absolute error of 0.1448, and a Pearson correlation coefficient of 0.3187 on the test set. Furthermore, after combining model prediction with empirical rules, the true yield of test wafers for the top 22% candidate ingots increased from 61.50% to 63.59%. These results indicate that the proposed method can effectively support early ingot screening and processing-priority decisions. This study demonstrates the application potential of data-driven methods in early-stage quality evaluation of CZT crystals and provides a reference framework for yield prediction in similar multi-wafer crystalline materials.