Speaker
Description
Achieving high-purity, detector-grade germanium (HPGe) crystals is essential for rare-event physics experiments such as dark matter detection and neutrinoless double-beta decay searches. We present a predictive machine learning framework that leverages Long Short-Term Memory (LSTM) networks to forecast the detector-grade yield fraction in HPGe crystal growth, based on experimentally obtained process data. The LSTM model is trained to predict detector-grade portions and assess the relative impact of input parameters, with interpretability enabled via SHAP (SHapley Additive exPlanations) analysis. Initial results demonstrate moderate-to-high predictive accuracy, revealing nontrivial correlations between impurity-related parameters and growth outcomes. To enhance the physical consistency of the input data, a physics-based impurity segregation model is incorporated to refine feature representations. The trained model is also employed in an inverse mode to identify optimal input parameter ranges that maximize detector-grade yield. This study provides a data-driven foundation for increasing the yield of detector-grade HPGe in each growth cycle by utilizing LSTM model–predicted input parameters for process control and optimization.
