Predicting 31P NMR Shifts in Large-Scale, Heterogeneous Databases by Gas Phase DFT: Impact of Conformer and Solvent Effects.

In this study, we extend the Ilm-NMR-P31 data set by systematically enriching experimental 31P nuclear magnetic resonance (NMR) data with quantum-chemically derived information. Using density functional theory (DFT) at the BLYP and B3LYP levels, we calculated 31P chemical shifts and optimized molecular geometries for 10,007 phosphorus-containing compounds. This hybrid data set provides a consistent framework that combines experimental measurements with theoretical descriptors, thereby facilitating the development of improved machine learning (ML) models and advanced structure elucidation tools. To evaluate the reliability of the computational data, we compared the DFT-derived shifts against experimental values. Single-point calculations in vacuum resulted in a root mean squared error (RMSE) of 30.82 ppm, which improved to 29.37 ppm when up to 20 conformers per molecule were included. Implicit solvent models produced minor additional improvements. These findings highlight the importance of conformational sampling. Evaluation of different functionals on the Ilm-NMR-P31 database, including those closely reproducing coupled-cluster results, demonstrated that B3LYP not only achieves sufficient accuracy for reproducing NMR shifts across large experimental databases but also emerges as a broadly applicable choice for routine applications. Overall, this work provides a large-scale, quantum chemically enriched 31P NMR data set that enhances the value of experimental data and supports future ML-based prediction models.