Machine learning prediction of NMR shifts for rare and transition metal complexes (45Sc, 49Ti, 89Y, 91Zr, 139La)

This study presents a machine learning (ML) approach for predicting nuclear magnetic resonance (NMR) chemical shifts of rare and transition metal complexes, specifically focusing on ⁴⁵Sc, ⁴⁹Ti, ⁸⁹Y, ⁹¹Zr and ¹³⁹La nuclei. A dataset comprising 499 experimental NMR measurements, spanning chemical shifts from −1389 to 1325 ppm, was compiled. Five ML models utilizing 2D molecular descriptors were evaluated, with the CatBoost model employing RDKit descriptors demonstrating superior performance. The model achieved root-mean-square errors (RMSE) of 123.7 ppm (∼7 %) for the combined ⁴⁵Sc, ⁸⁹Y, and ¹³⁹La set, 240.4 ppm (∼9 %) for Ti-containing compounds, and 165.2 ppm (∼13 %) for Zr-containing compounds. SHAP analysis identified key structural features influencing chemical shifts, such as cyclic moieties and electrostatic interactions, while solvent effects were found to be negligible. Notably, a unified model for ⁴⁵Sc, ⁸⁹Y, and ¹³⁹La maintained predictive accuracy, whereas combining ⁴⁹Ti and ⁹¹Zr data degraded performance due to divergent electronic environments. This work establishes a resource-efficient framework for NMR shift prediction in metal complexes, facilitating rapid screening for applications in catalysis, materials science, and molecular diagnostics