High-Fidelity prediction of radioisotope production Cross-Sections using Bayesian neural networks and Auto Machine learning

YanBang Tang,
High-Fidelity prediction of radioisotope production Cross-Sections using Bayesian neural networks and Auto Machine learning,
Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms,
Volume 569,
2025,
165887,
ISSN 0168-583X,
https://doi.org/10.1016/j.nimb.2025.165887.
(https://www.sciencedirect.com/science/article/pii/S0168583X25002770)
Abstract: The precise production of copper radioisotopes (⁶1Cu, ⁶2Cu, ⁶4Cu) is critical for advancing positron emission tomography (PET) and theranostics in nuclear medicine. A persistent challenge is the accurate prediction of reaction cross-sections to optimize production yields and purity. While experimental data from libraries like the IAEA are the gold standard, they are often sparse, and theoretical models such as TENDL-2021 can exhibit significant discrepancies. This study systematically investigates the use of advanced machine learning to create high-fidelity predictive models for six key Ni-target reactions leading to Cu isotopes. Using evaluated data from the IAEA database, we developed and benchmarked a suite of models, focusing on a Bayesian Neural Network (BNN) for robust uncertainty quantification and the AutoGluon framework for state-of-the-art automated machine learning (AutoML). The models were trained on physics-informed features and rigorously validated against an unseen test set, raw experimental data from EXFOR, and the TENDL-2021 theoretical library. Both the BNN and AutoGluon achieved exceptional predictive fidelity on unseen test data (Coefficient of Determination R2 > 0.999; Root Mean Squared Error < 5.5 mbarn), vastly outperforming a range of traditional machine learning algorithms. Crucially, the data-driven models demonstrated superior agreement with the evaluated experimental data compared to the TENDL-2021 theoretical library, which exhibited notable deviations in peak cross-section and energy for several key reactions. The BNN successfully quantified predictive uncertainty, yielding narrow confidence intervals in data-rich regions and wider intervals where data was sparse. This work establishes that for high-fidelity applications, advanced machine learning models trained on evaluated data can serve as a more reliable predictive tool than general-purpose theoretical models. The validated models were used to determine optimal, data-driven energy windows for the production of each radioisotope, offering practical guidance for clinical and research applications. This study provides a robust framework for creating continuous, uncertainty-aware representations of discrete nuclear data, with the potential to guide future experiments and accelerate the development of next-generation radiopharmaceuticals.
Keywords: Nuclear Data; Machine Learning; Medical Radioisotopes; Reaction Cross-Section