Machine learning approach for the prediction of nuclear quadrupole resonance frequencies

Nuclear quadrupole resonance (NQR) is an indispensable experimental technique, which can elucidate microscopic electronic state of materials. NQR frequency, ν_Q, of material is important information which reflects the properties of the material. In the present study, a supervised machine learning (ML) technique is applied to predict ν_Q of ³⁵Cl NQR in organic molecules. The input data of the ML is feature vectors encoded by molecular SMILES. To find the optimum labeling of experimental ν_Q to Cl in SMILES, an Metropolis-Hastings-like algorithm combined with ML is developed. ν_Q is also estimated by calculating electric-field gradient at Cl site based on the density functional theory (DFT) considering all electrons (full-potential local-orbital method). The accuracy of the prediction with only ML using SMILES is comparable to that of DFT calculation. The ML prediction scheme has great advantage of not requiring atomic coordinates and many computer resources and is potentially applicable to the prediction of the other spectroscopic data. (author)