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

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Spectroscopy

Background:

• Predicting NMR chemical shifts in transition metal complexes is complex due to diverse coordination environments and electronic structures.
• Existing quantum chemical calculations are computationally expensive and time-consuming for these systems.

Purpose of the Study:

• To develop a rapid and accurate machine learning (ML) model for predicting proton (${}^1\text{H}$) NMR chemical shifts in zinc complexes.
• To provide a computationally efficient alternative to density functional theory (DFT) calculations for characterizing organometallic compounds.

Main Methods:

• Systematic selection of diverse zinc complexes from the transition metal quantum mechanics (tmQM) database using K-means clustering on Smooth Overlap of Atomic Positions (SOAP) descriptors.
• Generation of training data through DFT NMR calculations across five different solvent environments.
• Combination of SOAP descriptors with tree-based ensemble methods, specifically LightGBM, for predicting proton chemical shifts.

Main Results:

• The LightGBM model achieved high accuracy on held-out test data (MAE = 0.016 ppm, R² = 0.99).
• External validation against experimental NMR data showed strong predictive performance (R² = 0.90, MAE = 0.56 ppm) across multiple solvents.
• The model demonstrated excellent transferability, maintaining accuracy for solvents not included in the training set, such as acetonitrile.

Conclusions:

• The developed ML approach provides a computationally efficient and accurate method for predicting ${}^1\text{H}$ NMR shifts in zinc complexes.
• Prediction times are orders of magnitude faster than DFT methods, while achieving comparable accuracy.
• This approach can significantly accelerate the characterization and design of organometallic compounds.