NMR shifts in aluminosilicate glasses via machine learning

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• Nuclear Magnetic Resonance (NMR) spectroscopy is a powerful tool for characterizing aluminosilicate glasses.
• Density Functional Theory (DFT) calculations complement NMR but are computationally intensive and limited in scale.
• Existing methods for predicting NMR parameters in glasses face significant computational challenges.

Purpose of the Study:

• To investigate machine learning (ML) approaches for predicting NMR parameters in aluminosilicate glasses.
• To develop a computationally efficient method for obtaining accurate NMR parameters compared to DFT.
• To explore ML models for predicting isotropic magnetic shielding (σiso) in various glass compositions.

Main Methods:

• Employed machine learning (ML) algorithms, including linear ridge regression (LRR), to predict NMR parameters.
• Utilized atom-centered representations and descriptors like the smooth overlap of atomic positions (SOAP) to encode local atomic environments.
• Compared ML predictions with DFT-GIPAW calculations on relaxed and room-temperature glass structures.

Main Results:

• ML predictions for isotropic magnetic shielding (σiso) showed high accuracy compared to DFT-GIPAW calculations.
• Deviations from DFT were 0.7 ppm for 29Si and 1.5 ppm for 17O in SiO2 glasses.
• Accurate predictions were also achieved for 23Na in Na2O-SiO2 and 27Al in Al2O3-Na2O-SiO2 systems.

Conclusions:

• Machine learning provides a rapid and accurate method for predicting NMR parameters in aluminosilicate glasses.
• ML approaches overcome the computational limitations of DFT, enabling faster analysis of complex glass systems.
• The developed ML procedures offer a valuable tool for materials scientists studying glass structures and properties.