A neural network with uncertainty quantification for predicting deuteron-induced reaction cross sections using theoretical priors

The prediction of deuteron-induced neutron emission ( d , x n ) reaction cross sections is important for applications in fusion energy and medical isotope production. Theoretical models, such as TALYS, provide baseline predictions but can show discrepancies with experimental data. This study develops a framework that integrates machine learning with theoretical calculations to improve predictive performance. We utilize a Multi-Layer Perceptron (MLP) model with Monte Carlo (MC) Dropout to predict ( d , x n ) cross sections and to quantify the associated model uncertainty. A key aspect of this work is the inclusion of TALYS-calculated cross sections as an input feature, serving as a theoretical prior. The model was trained and validated on a curated dataset of 1702 experimental data points from the EXFOR database. On an unseen test set, the optimized model achieved a coefficient of determination (R 2 ) of 0.90 and a Root Mean Squared Error (RMSE) of 77.85 mb. To interpret the model's behavior, SHapley Additive exPlanations (SHAP) was employed. The analysis indicates that the TALYS prior is the most influential feature, while properties such as target and product proton numbers (Z), which relate to the Coulomb barrier, are also identified as significant inputs. This work demonstrates a methodology for combining theoretical calculations with a data-driven model to produce predictions with quantified uncertainty, offering a useful tool for nuclear data evaluation.