Data Fusion and Contrastive Alignment for Unconstrained IR Molecular Structure Elucidation

Short summary

This paper improves automated molecular structure elucidation from IR spectroscopy by modifying encoder-decoder transformers with a Mixture-of-Experts decoder using non-additive aggregation via the Choquet integral. A contrastive alignment loss term further enhances performance. Together these modifications improve Top-K prediction accuracy by over 10 percentage points versus baseline IR-only models, demonstrating that IR spectra encode most relevant chemical information for unconstrained structure prediction.

• •MoE decoder with Choquet integral aggregation improves IR-based molecular structure prediction
• •Contrastive alignment loss adds over 10 percentage points to Top-K accuracy vs baseline
• •Findings confirm IR spectra encode most relevant chemical information for unconstrained elucidation

Generated with AI, which can make mistakes.