Predicting (n,3n) nuclear reaction cross-sections using XGBoost and Leave-One-Out Cross-Validation

Abstract

Accurately predicting nuclear reaction cross-sections is crucial for advancing various fields, including nuclear medicine, energy production, and materials science. This study aims to address the challenges associated with predicting (n $,3n$) nuclear reaction cross-sections by developing a robust machine learning (ML) model based on the XGBoost (eXtreme Gradient Boosting) algorithm. By leveraging a comprehensive dataset of experimental cross-sectional values, the study demonstrates the potential of ML to overcome limitations in existing theoretical and empirical approaches. LOOCV (Leave-One-Out Cross-Validation) was employed for feature selection and hyperparameter optimization to ensure the reliability of the model. The dataset was meticulously prepared by normalizing values and addressing missing data, which contributed to robust model training. XGBoost's ability to handle complex, non-linear relationships enabled it to provide accurate predictions that closely align with experimental data, as evaluated through key metrics such as Mean Squared Error (MSE) and Mean Absolute Error (MAE), and reduced Chi-Square. To validate the model's accuracy, its predictions were compared with calculations from the TALYS 1.95 nuclear reaction code, TENDL and phenological model. The results highlight the efficacy of XGBoost in improving prediction accuracy, offering a novel approach to solving complex challenges in nuclear data analysis.