Computation of CCSD(T)-Quality NMR Chemical Shifts via Δ-Machine Learning from DFT

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Spectroscopy

Background:

• Nuclear Magnetic Resonance (NMR) spectroscopy is crucial for molecular structure determination.
• Accurate computational prediction of NMR parameters, especially chemical shifts, is highly desirable in chemistry.
• Current Density Functional Theory (DFT) methods often require correction for precise NMR chemical shift predictions.

Purpose of the Study:

• To develop and present a novel machine learning (ML) approach for correcting DFT-computed NMR chemical shifts.
• To enhance the accuracy of NMR chemical shift predictions using a delta-machine learning (Δ-ML) model.
• To validate the ML correction method's performance against highly accurate reference data and external benchmarks.

Main Methods:

• A Δ-machine learning approach was employed to correct DFT-computed NMR chemical shifts.
• The ML model utilized input features from DFT calculations and highly accurate reference data (CCSD(T)/pcSseg-2 with basis set extrapolation).
• The model was trained on a dataset of 1000 small organic molecule structures with 7090 1H and 4230 13C NMR chemical shifts.

Main Results:

• The Δ-ML approach reduced the mean absolute deviation (MAD) by 81% for 1H and 92% for 13C NMR chemical shifts when applied to the PBE0/pcSseg-2 method.
• ML-corrected NMR shifts showed a low MAD, ranging from 0.021 to 0.039 ppm for 1H and 0.38 to 1.07 ppm for 13C across 12 DFT functional/basis set combinations.
• The ML correction consistently outperformed traditional linear regression techniques and demonstrated robustness on external benchmark sets.

Conclusions:

• The presented Δ-machine learning approach offers a significant and computationally inexpensive improvement for predicting NMR chemical shifts.
• The method provides accurate and robust NMR chemical shift predictions applicable across various DFT functional and basis set combinations.
• This correction scheme is readily applicable to Density Functional Theory-based spectral simulations, enhancing their reliability.