Enhancing Detector-Grade Yield in High-Purity Germanium Crystal Growth for Next-Generation Dark Matter and Rare Event Detectors

High-purity germanium (HPGe) detectors play a critical role in low-background and low-energy experiments, particularly in the search for dark matter and the study of neutrino properties. A major bottleneck in scaling up such experiments is the limited yield of large-mass, detector-grade crystals per growth run. This challenge stems from the inherent difficulty in growing ultra-pure, defect-free germanium crystals suitable for detector fabrication.

To address this, we have developed a machine learning (ML)-driven framework to optimize HPGe crystal growth and improve the detector-grade yield per growth. Using data from over 50 crystal growth runs conducted at the University of South Dakota, we trained Long Short-Term Memory (LSTM) neural networks to model impurity evolution during growth and predict the detector-grade fraction based on process parameters.

The model demonstrates high predictive accuracy, and SHapley Additive exPlanations (SHAP) analysis reveals that net impurity concentration—particularly residual impurities from previous growths and the quality of input materials—is the dominant limiting factor. These insights provide actionable guidance for refining growth protocols and improving material purification strategies.

Ongoing work involves integrating zone-refining data, developing real-time ML-based feedback for dynamic process control, and employing atomistic simulations to study impurity segregation at the solid–liquid interface. This integrated approach aims to enable scalable production of ton-scale HPGe crystals for next-generation dark matter and rare-event detection experiments.