Transformer + Mixture-of-Experts predicts molecular structures from IR spectra

Automated molecular structure elucidation from infrared (IR) spectroscopy has long been limited by the need for a pre‑determined chemical formula as input, restricting predictions to isomer identification. A team led by Ethan J. Mick at the University of Missouri introduces key modifications to the standard encoder‑decoder transformer to tackle the unconstrained problem. They replace the conventional decoder with a Mixture‑of‑Experts (MoE) module that uses non‑additive aggregation via linear‑order statistics and the Choquet integral, better handling the vast chemical space. The encoder also incorporates these non‑additive operators when aggregating spectral representations, and an auxiliary contrastive alignment loss improves separation between correct and incorrect structures.

Evaluated on IR spectra datasets, the enhanced transformer achieves a Top‑K accuracy improvement of over 10 percentage points compared to baseline IR‑only models. Substructure fragment analysis confirms that infrared spectra encode the majority of relevant chemical information, meaning previous lower performance was due to overlapping absorption bands rather than missing data. This work demonstrates that full molecular structure predictions can be made reliably without prior formula knowledge, significantly expanding the utility of AI in analytical chemistry for applications such as drug discovery, materials design, and forensic analysis.