Revving up ¹³C NMR shielding predictions across chemical space: benchmarks for atoms-in-molecules kernel machine learning with new data for 134 kilo molecules

Abstract The requirement for accelerated and quantitatively accurate screening of nuclear magnetic resonance spectra across the small molecules chemical compound space is two-fold: (1) a robust ‘local’ machine learning (ML) strategy capturing effect neighborhood on an atom’s ‘near-sighted’ property—chemical shielding; (2) reference dataset generated with state-of-the-art first-principles method training. Herein we report QM9-NMR comprising isotropic shielding over 0.8 million C atoms in 134k QM9 gas five common solvent phases. Using these data training, present benchmark results prediction transferability kernel-ridge regression models popular local descriptors. Our best model, trained 100k samples, accurately predicts 50k ‘hold-out’ mean error less than 1.9 ppm. For rapid new query molecules, were geometries from inexpensive theory. Furthermore, by using ?-ML strategy, quench below 1.4 Finally, test non-trivial sets that include 10–17 heavy drugs.