Efficient Machine Learning Prediction of Solvent-Dependent 1 H $$ {}^1\mathrm{H} $$ NMR Chemical Shifts in Zinc...

Accurate prediction of NMR chemical shifts in transition metal complexes remains challenging due to the wide range of coordination environments and complex electronic structures of these systems. In this work, we present a machine learning approach (ML) for rapid and accurate prediction of ${}^1\text{H}$ NMR shifts in zinc complexes across multiple solvent environments. We systematically selected a diverse set of zinc complexes from the transition metal quantum mechanics (tmQM) database using K-means clustering on SOAP descriptors, and performed DFT NMR calculations across five solvents to generate training data. We combine smooth overlap of atomic positions (SOAP) descriptors with tree-based ensemble methods to predict proton chemical shifts. Among several ML algorithms evaluated, LightGBM achieved the best performance on held-out test complexes (MAE = 0.016 ppm, RMSE = 0.028 ppm, $R^2$  = 0.99), demonstrating excellent generalization to unseen molecular structures. External validation against experimental NMR data across multiple solvents revealed strong predictive performance ( $R^2$  = 0.90, MAE = 0.56 ppm), with exceptional accuracy in methanol ( $R^2$  = 0.96) and acetonitrile ( $R^2$  = 0.91). Notably, the model demonstrated robust transferability to acetonitrile despite this solvent not being included in the training set. This approach provides a computationally efficient alternative to expensive quantum chemical calculations for predicting ${}^1\text{H}$ NMR shifts in transition metal complexes, offering prediction times that are orders of magnitude faster while maintaining accuracy comparable to DFT methods, potentially accelerating the characterization and design of organometallic compounds.