Testing the Limits of Transformer-Based Sequence Inference from Copolymer Spectra

<jats:p>Sequence-defined synthetic copolymers promise biological-like precision in material design, but their utility depends on the ability to read sequences directly from accessible analytical measurements. We introduce a multimodal transformer that decodes copolymer sequences from simulated UV-Vis, 1H NMR, and mass spectra, using three spectral transformers feeding a deduction layer that collaboratively predicts each monomer along the chain. Applied to donor-acceptor conjugated copolymers, we first establish a theoretical ceiling on what each spectral technique can resolve by enumerating sequence-spectrum degeneracies, and demonstrate that our trained models saturate these limits for every modality. The combined model achieves &gt;90% top-1 sequence reconstruction and remains robust to additional noise while faltering only upon substantial information loss, with 1H NMR proving essential for preserving local k-mer structure when the UV-Vis and mass channels degrade. These results quantify how much sequence information is recoverable from standard spectral measurements and motivate ML-assisted multi-spectral analysis as a promising route toward general-purpose synthetic copolymer sequencing.</jats:p>