M Tamkas
This study investigates the prediction of density-dependent full-energy peak efficiency (FEPE) behavior in Marinelli geometry using ensemble-based machine-learning techniques trained on experimentally derived NaI(Tl) detector calibration data. Experimental FEPE curves corresponding to 0.7, 1.0, and 1.3 g cm-3 sample densities were obtained using mixed gamma-ray standards in the 60-1837 keV energy range. Experimentally derived FEPE datasets were represented using logarithmic polynomial efficiency functions in order to preserve the energy-dependent detector-response structure prior to machine-learning-based interpolation analysis. Gradient Boosting regression and Random Forest regression models were trained using the 0.7 and 1.3 g cm-3 datasets, while the independently measured 1.0 g cm-3 dataset was withheld from training and used exclusively for prediction and validation. Model performance was evaluated using MAE, RMSE, MAPE, and the coefficient of determination (R2). The results demonstrated that both ensemble-learning approaches successfully reproduced the density-dependent detector response with high consistency. The results demonstrated that both ensemble-learning approaches successfully reproduced the density-dependent detector response with high consistency, with Gradient Boosting yielding lower prediction errors than Random Forest across the investigated calibration energies. Although conventional linear and log-linear interpolation yielded lower numerical errors for the present smooth, symmetrically bounded density configuration, both ensemble-learning models successfully reconstructed the independently measured intermediate-density FEPE response. The study demonstrates that ensemble-learning approaches can provide a reliable data-driven framework for reconstructing unseen intermediate-density FEPE responses from experimentally measured boundary-density calibration data within an experimentally characterized density range.