A Machine Learning Model of Chemical Shifts for Chemically and Structurally Diverse Molecular Solids

Area of Science:

• Solid-state chemistry
• Computational materials science
• Spectroscopy

Background:

• Nuclear magnetic resonance (NMR) chemical shifts reveal local atomic environments crucial for solid material structure determination.
• Predicting accurate chemical shifts is computationally expensive, hindering NMR crystallography.
• Existing ShiftML model predicts chemical shifts for C, H, N, O, S based on minimum-energy geometries.

Purpose of the Study:

• Extend ShiftML capabilities to predict chemical shifts for finite temperature structures.
• Enhance ShiftML to handle more chemically diverse compounds.
• Maintain high prediction accuracy and computational efficiency.

Main Methods:

• Machine learning model (ShiftML) trained on diverse molecular solid structures.
• Incorporation of finite temperature effects in structural data.
• Validation against experimental NMR chemical shift data for 13 molecular solids.

Main Results:

• Achieved a root-mean-squared error of 0.47 ppm for ¹H shift predictions on a benchmark set.
• Maintained accuracy comparable to density functional theory (DFT) calculations (0.35 ppm RMSE).
• Reduced computational cost by over four orders of magnitude compared to traditional DFT methods.

Conclusions:

• The enhanced ShiftML model accurately predicts NMR chemical shifts for diverse molecular solids at finite temperatures.
• ShiftML significantly accelerates NMR crystallography by drastically reducing computational demands.
• This advancement facilitates faster and more efficient structural determination of solid materials.