Interpretable Machine Learning for Proton-induced Neutron Reaction Cross-sections Prediction

The accurate prediction of proton-induced neutron \((p, n)\) reaction cross sections is critical for applications in nuclear engineering, medical isotope production, and astrophysics. This study introduces a machine learning framework that combines high-predictive accuracy with interpretability and uncertainty quantification. We integrate the Hiking Optimization Algorithm (HOA) with the eXtreme Gradient Boosting (XGBoost) model and employ Monte Carlo (MC) Dropout for uncertainty estimation. The framework is interpreted using SHapley Additive exPlanations (SHAP). First, HOA is utilized to navigate the high-dimensional hyperparameter space, optimizing the XGBoost model for performance. Subsequently, the trained model’s predictions are benchmarked against experimental data from the EXFOR database and theoretical calculations from TALYS 2.0. Our HOA–XGBoost model demonstrates better predictive accuracy over other machine learning models and provides predictions closer to experimental values than TALYS 2.0. The inclusion of MC Dropout provides uncertainty bounds for the model’s predictions. A detailed SHAP analysis reveals the underlying physical drivers of the model’s decisions: the incident proton energy (\(EN\)) is identified as the most influential feature, with its strong interaction with the reaction \(Q\)-value and product proton number (\(Z_2\)) highlighting the model’s ability to learn fundamental concepts such as reaction thresholds and the Coulomb barrier. The product nuclide’s neutron (\(N_2\)) and proton (\(Z_2\)) numbers also show influence related to nuclear stability, while the product mass number (\(A_2\)) has a lesser impact. This work presents a complementary methodology for nuclear data evaluation, paving the way for more reliable predictions and targeted experimental design. Abstract Published by the Jagiellonian University 2025 authors