Accelerating NMR Shielding Calculations Through Machine Learning Methods: Application to Magnesium Sodium Silicate Glasses

In this work, we have applied the Kernel Ridge Regression (KRR) method using a Least Square Support Vector Regression (LSSVR) approach for the prediction of the NMR isotropic magnetic shielding (σiso) of active nuclei (17O, 23Na, 25Mg, and 29Si) in a series of (Mg, Na)–silicate glasses. The Machine Learning (ML) algorithm has been trained by mapping the local environment of each atom described by the Smooth Overlap of Atomic Position (SOAP) descriptor with isotropic chemical shielding values computed with DFT using the Gauge-Included-Projector-Augmented-Wave (GIPAW) approach. The influence of different training datasets generated through molecular dynamics simulations at various temperatures and with different inter-atomic potentials has been tested and we demonstrate the importance of a wide exploration of the configurational space to enhance the transferability of the ML-regressor. Finally, the trained ML-regressor has been used to simulate the 29Si MAS NMR spectra of systems containing up to 20000 atoms by averaging hundreds of configurations extracted from classical MD simulations to account for thermal vibrations. This ML approach is a powerful tool for the interpretation of NMR spectra using relatively large systems at a fraction of the computational time required by quantum mechanical calculations which are of high computational cost.