Predicting Density Functional Theory-Quality Nuclear Magnetic Resonance Chemical Shifts via Δ-Machine Learning.

摘要: First-principles prediction of nuclear magnetic resonance chemical shifts plays an increasingly important role in the interpretation experimental spectra, but required density functional theory (DFT) calculations can be computationally expensive. Promising machine learning models for predicting shieldings general organic molecules have been developed previously, though accuracy those remains below that DFT. The present study demonstrates how much higher obtained via Δ-machine approach, with result errors introduced by model are only one-half to one-third expected DFT relative experiment. Specifically, ensemble neural networks is trained correct PBE0/6-31G up target level PBE0/6-311+G(2d,p). It predict 1H, 13C, 15N, and 17O root-mean-square 0.11, 0.70, 1.69, 2.47 ppm, respectively. At same time, approach 1-2 orders magnitude faster than large-basis calculations. also demonstrated predicts solution-phase NMR drug modestly worse model. Finally, ability estimate uncertainty predicted based on variations within network assessed.