RW-GPINN: A Data- and Physics-Driven Algorithm for Full-Energy Peak Passive Efficiency Calibration of Gamma-Ray Spectrometer

Full-energy peak efficiency (FEPE) calibration of gamma-ray spectrometers is critical in applications such as non-destructive characterization of nuclear waste and quantitative radionuclide analysis. Conventional Monte Carlo simulations (e.g., Geant4) provide accurate physical modeling but suffer from high computational costs, cumbersome implementation, and limited adaptability across detector types and geometric configurations. While data-driven approaches are computationally efficient, they often lack physical consistency and exhibit poor generalization in low-efficiency regimes. To address these challenges, we propose a Reweighted Geometric Physics-Informed Neural Network (RW-GPINN) that integrates physical knowledge with machine learning. Within the Physics-Informed Neural Network (PINN) framework, gamma-ray attenuation laws are incorporated as physics-guided loss terms, while geometry-aware features, target-variable transformation, and a sample reweighting mechanism enhance model robustness and accuracy, particularly in sparse low-efficiency regions. Within a physics-informed paradigm, our contribution is not a direct re-use of PINN; rather, we integrate (i) geometry-aware features that encode solid-angle and incidence-angle effects, (ii) a target-variable transformation for the highly skewed efficiency scale, and (iii) a sample reweighting mechanism that emphasizes low-efficiency regimes. These design choices make the approach effective for passive calibration beyond a straightforward PINN application. A large-scale multi-parameter simulation dataset generated with Geant4 was used for training and validation. The proposed model achieved an R2of 0.9020 and an MAE of 2.40×10−3, outperforming benchmark methods including ANN, DNN, CF, and VPD. Ablation studies confirm the contribution of each component, and experimental validation demonstrates strong agreement with measured data. This work establishes RW-GPINN as a physically consistent and computationally efficient framework for passive efficiency calibration under complex experimental conditions.