Charge-aware machine learning for the infrared spectra of interstellar polycyclic aromatic hydrocarbons

Aims. Polycyclic aromatic hydrocarbons (PAHs) are among the most abundant molecules in the interstellar medium. Their characteristic infrared (IR) emission acts as a sensitive probe of astrophysical environments, yet detailed spectral analyses have been limited by the high computational cost of density functional theory (DFT) calculations. This constraint has hindered a systematic exploration of how spectral features such as the aromatic IR bands depend on a PAH's charge state and molecular structure. Methods. Our goal is to develop a computationally efficient machine learning model capable of predicting IR spectra for PAHs across charge states, and to critically reassess established interpretations of how ionization influences these spectra. Results. We developed a neural network framework to predict PAH IR spectra across four charge states, utilizing a dataset of 12599 species. Molecular structures were represented by topological fingerprints, with charge states integrated via learnable embeddings. Additionally, a random forest classifier was implemented to infer charge states directly from spectral data. Conclusions. The model achieves near-DFT accuracy in predicting IR spectra while offering orders-of-magnitude acceleration in computation. It reliably handles PAHs containing up to 150 carbon atoms, including anions, neutrals, cations, and di-cations. The predictive capability for larger molecules is currently limited by the available training data. The classifier predicts charge states with over 99% accuracy. Our analysis of the DFT-computed spectra shows that anions exhibit strong emission across multiple bands, often matching or exceeding cation intensities, and the 11.2 micrometer band shows a distinct charge dependence.