Prediction of the Ground-State Electronic Structure from Core-Loss Spectra of Organic Molecules by Machine Learning

The core-loss spectrumreflects the partial density of states (PDOS)of the unoccupied states at the excited state and is a powerful analyticaltechnique to investigate local atomic and electronic structures ofmaterials. However, various molecular properties governed by the ground-stateelectronic structure of the occupied orbital cannot be directly obtainedfrom the core-loss spectra. Here, we constructed a machine learningmodel to predict the ground-state carbon s- and p-orbital PDOS inboth occupied and unoccupied states from the C K-edge spectra. Wealso attempted an extrapolation prediction of the PDOS of larger moleculesusing a model trained by smaller molecules and found that the extrapolationprediction performance can be improved by excluding tiny molecules.Besides, we found that using smoothing preprocess and training byspecific noise data can improve the PDOS prediction for noise-containedspectra, which pave a way for the application of the prediction modelto the experimental data.

Automated summary

A machine learning model was constructed to predict the partial density of states (PDOS) of the ground-state carbon s- and p-orbitals from C K-edge spectra. The model was also tested for extrapolation prediction of PDOS for larger molecules, and excluding tiny molecules improved the extrapolation performance. Additionally, the PDOS prediction for spectra with noise can be enhanced by using smoothing preprocess and training with specific noise data, enabling the application of the prediction model to experimental data.