Prediction and interpretation of (n,2n) reaction cross sections using a KOA-XGBoost-SHAP machine learning framework

Accurate prediction of (n,2n) reaction cross sections is crucial for nuclear applications but faces challenges due to experimental difficulties and data scarcity. This study proposes a novel triple-hybrid machine learning framework, KOA-XGBoost-SHAP, to address these challenges. The framework integrates the Kepler Optimization Algorithm (KOA) for hyperparameter tuning, the XGBoost model for prediction, and SHapley Additive exPlanations (SHAP) for interpretability. Using 1,282 (n,2n) cross-section data points from the EXFOR database, covering incident neutron energies (EN) from 1 to 375 MeV and target nuclides with Z from 1 to 94 and A from 2 to 239, six ML models were optimized using KOA. The KOA-XGBoost model demonstrated superior predictive accuracy (R² = 0.94, RMSE = 152.99 mb) when benchmarked against the EXFOR test set. Notably, its direct agreement with these raw EXFOR experimental data was also better than that achieved by ENDF/B-VII.1 evaluations (R² = 0.84, RMSE = 247.90 mb for available points) and default TALYS 2.0 calculations (R² = 0.75, RMSE = 312.59 mb) on the same test set data points. SHAP analysis quantitatively identified the contributions of input features (incident neutron energy, target/product nuclide properties, Q-value), revealing that EN and target characteristics are dominant. Furthermore, SHAP elucidated synergistic feature interactions, offering data-driven insights into the physical mechanisms influencing cross sections. This research demonstrates a robust, accurate, and interpretable ML framework for (n,2n) cross-section prediction, showcasing its capacity to further refine predictions based directly on experimental data, complementing established methods like ENDF and TALYS.