Photoemission spectroscopy of organic molecules using plane wave/pseudopotential density functional theory and machine learning: A comprehensive and predictive computational protocol for isolated molecules, molecular aggregates, and organic thin films

Abstract

Photoemission measurements in the gas phase at low pressure have enabled the exploration of the intricate relationship between electronic and structural properties at the single-molecule level. Experimental data collected from isolated molecules, free from interactions with other species, have provided an ideal testing ground for developing ab initio simulations capable of interpreting and predicting photoemission spectra. In particular, accurate computational methods for determining atom- and site-specific core ionization binding energies (BEs) facilitate experimental data interpretation, enabling the assignment of contributions from non-equivalent atoms of the same species, even when spectral features remain unresolved due to molecular structure. In this context, we have developed, extensively tested, and made widely available a computational protocol based on plane wave/pseudopotential density functional theory (PW-DFT) within a ΔSCF framework to predict x-ray photoemission spectra (XPS) of isolated molecules. Moreover, we have preliminarily tested and demonstrated the applicability of the same method to large molecular aggregates and thin molecular films deposited on inorganic substrates. The protocol has been assessed using a representative set of semilocal and hybrid density functionals with increasing fractions of Hartree–Fock exact exchange (EXX), including PBE, B3LYP (20% EXX), HSE (range-separated with 25% EXX at short range), and BH&HLYP (50% EXX). As a benchmark, we have also employed the equation-of-motion coupled-cluster method with single and double excitations. Our protocol has been validated across a diverse range of molecular classes—including aromatic, heteroaromatic, and aliphatic compounds; drugs; and biomolecules—demonstrating high accuracy and robustness, even when using semilocal DFT. In addition, valence photoemission measurements complement core photoemission by providing insights into delocalized and π-conjugated molecular orbitals. These measurements are particularly useful for studying chemical modifications in large molecules mediated by non-covalent interactions. Using the same set of density functionals, we have evaluated their capability to predict valence-shell ionization spectra, employing Kohn–Sham eigenvalues as estimators. Finally, our PW-DFT dataset of C1s, N1s, and O1s BEs has been used to train machine learning (ML) models for predicting XPS spectra of isolated organic molecules based on their structure. To ensure reproducibility and encourage the adoption of our protocol, we have made available a public repository containing pseudopotentials, input files for ab initio calculations, and datasets used for ML model training.