Enhancing nuclear cross-section predictions with deep learning: the DINo algorithm

Abstract

Accurate modeling of nuclear reaction cross-sections is crucial for applications such as hadron therapy, radiation protection, and nuclear reactor design. Despite continuous advancements in nuclear physics, significant discrepancies persist between experimental data and theoretical models such as TENDL, and ENDF/B. These deviations introduce uncertainties in Monte Carlo simulations widely used in nuclear physics and medical applications. In this work, DINo (Deep learning Intelligence for Nuclear reactiOns) is introduced as a deep learning-based algorithm designed to improve cross-section predictions by learning correlations between charge-changing and total cross-sections. Trained on the TENDL-2021 dataset and validated against experimental data from the EXFOR database, DINo demonstrates a significant improvement in predictive accuracy over conventional nuclear models. The results show that DINo systematically achieves lower chi2 values compared to TENDL-2021 across multiple isotopes, particularly for proton-induced reactions on a 12C target. Specifically, for 11C production, DINo reduces the discrepancy with experimental data by \sim 28\% compared to TENDL-2021. Additionally, DINo provides improved predictions for other relevant isotopes produced, such as 4He, 6Li, 9Be, and 10B, which play a crucial role in modeling nuclear fragmentation processes. By leveraging neural networks, DINo offers fast cross-section predictions, making it a promising complementary tool for nuclear reaction modeling. However, the algorithm's performance evaluation is sensitive to the availability of experimental data, with increased uncertainty in sparsely measured energy ranges. Future work will focus on refining the model through data augmentation, expanding its applicability to other reaction channels, and integrating it into Monte Carlo transport codes for real-time nuclear data processing.