AstroSpec-LLM: A Large Language Model Framework for High-throughput Infrared Spectral Prediction of Interstellar PAHs

Abstract Polycyclic aromatic hydrocarbons (PAHs) are vital probes of the interstellar medium, but the quartic scaling of traditional quantum chemical calculations hinders the characterization of large species observed by the James Webb Space Telescope (JWST). We present AstroSpec-LLM, a deep-learning framework that adapts the large language model architecture to treat molecular SMILES strings as “chemical sentences” for high-throughput, charge-sensitive spectral prediction. By leveraging a transformer-based encoder enhanced with rotary position embeddings and fine-tuned on a dataset of 24,146 PAH spectra, our model achieves exceptional structural generalization and data efficiency. By bypassing the bottleneck of density functional theory calculations, AstroSpec-LLM enables the rapid synthesis of massive, charge-aware spectral libraries, serving as a useful diagnostic tool for decoding the complex infrared universe.