Prediction of thick-target yields for medical radionuclide production based on automated machine learning

Accurate knowledge of thick-target yields (TTY) is critical for the efficient and reliable production of medical radionuclides. In this study, we developed and evaluated a suite of machine learning models to perform prediction of the TTY for four medically relevant (p,n) reactions: ¹⁶⁷Er(p,n)¹⁶⁷Tm, ⁵⁸Fe(p,n)^58mCo, ¹¹⁹Sn(p,n)¹¹⁹Sb, and ¹⁸⁶W(p,n)^186gRe. A unified dataset was curated from the IAEA's evaluated data library. Physics-informed features, including the proton, neutron, and mass numbers of both the target and product nuclides, were engineered to provide a physical context for the models. The performance of fourteen algorithms, including ensemble methods, kernel-based models, and a Automated Machine Learning (AutoML) framework, Autogluon, was systematically evaluated. The Autogluon model demonstrated strong predictive performance, achieving a high coefficient of determination (R² = 0.999995) and a low root mean squared error (RMSE = 0.021 MBq/μA·h) on a held-out test set. It outperformed all other models, particularly simple linear models (R² < 0.5) which failed to capture the non-linear nature of the yield curves. The model closely reproduced the TTY curves for four individual reactions in trained data this study. The observed large relative errors were confined to physically insignificant, near-threshold energy regions where absolute errors were negligible. This work presents a successful application of machine learning for the prediction of thick-target yields. The results establish that data-driven models, particularly those developed through AutoML, show promise as a complementary tool for nuclear data evaluation, supporting the optimization of radionuclide production for medical applications.