Accuracy of machine learning algorithms for HPGe detector efficiency determination

Abstract

The accurate determination of full-energy peak efficiency (FEPE) in High-Purity Germanium (HPGe) detectors is critical for gamma-ray spectrometry, especially when source-detector geometries vary. In this study, we investigate the application of six supervised machine learning (ML) algorithms—Polynomial Regression, Random Forest, XGBoost, LightGBM, Sparse Gaussian Process, and Multi-Layer Perceptron—for predicting FEPE of a Low Energy HPGe (LEGe) detector across a broad energy range (40–1600 keV) and diverse source types (point and volumetric). Datasets used for training, validation and testing the ML models were generated using Monte Carlo simulations (GESPECOR). Model performance was evaluated using cross-validation and standard error metrics (R², RMSE, MRE). Among the tested models, Polynomial Regression and LightGBM demonstrated superior predictive accuracy and interpretability, achieving R² values above 0.9999. SHAP values were used for explainability, demonstrating that the models successfully capture the key physical mechanisms influencing FEPE. These results position ML models as reliable and generalizable alternative to conventional FEPE calibration methods.