This study presents a rigorous machine-learning framework for modeling the cross-sections ( \(\sigma \) ) of ( \(n, d\) ) nuclear reactions at a standard fast-neutron energy of approximately 14.5 MeV derived from a Deuterium–Tritium (D–T) fusion source. Given that the incident energy remains constant across the curated experimental dataset of 28 nuclides ( \(Z = 2-73\) , \(A = 3-181\) ), the modeling is executed on a base-10 logarithmic scale to effectively accommodate cross-sections spanning nearly three orders of magnitude ( \(0.15-66\) mb). To overcome the inherent constraints of the small-data regime, the regression-adapted Synthetic Minority Oversampling Technique (SMOTER) was implemented. Methodological integrity was preserved via a robust, leakage-free Leave-One-Out cross-validation (LOOCV) protocol, wherein the held-out validation target was restricted to real experimental nuclides, and SMOTER was applied exclusively within each training fold. Furthermore, the synthetic instances are physically constrained to integer-valued atomic numbers ( \(Z\) ) and mass numbers ( \(A\) ) localized within the valley of \(\beta \) -stability. Employing a comprehensive suite of physically motivated nuclear structure descriptors, we systematically evaluated seven regressors (Lasso, Ridge, Elastic Net, Random Forest, Gradient Boosting, XGBoost, and Support Vector Regression) along with two meta-ensembles (Voting and Stacking). Under strict leakage-free evaluation, the Elastic Net model exhibited superior generalization capacity (R2 = 0.816, 95% CI [0.537, 0.894], RMSE = 0.259, and a within-factor-of-two agreement of 78.6%), closely followed by Lasso (R2 = 0.791), whereas tree- and kernel-based architectures performed poorly (e.g., XGBoost R2 = 0.608). Notably, our optimized Elastic Net model outperformed the TALYS-based TENDL-2019 evaluation (R2 = 0.686) for the same nuclide set. Finally, split conformal prediction intervals provide rigorously calibrated uncertainty quantification, yielding 89.3% empirical coverage against a 90% nominal confidence level. These findings underscore that in data-scarce nuclear modeling regimes, explicit regularization and physically informed feature engineering are far more decisive for model generalization than the nominal algorithmic complexity.
