Towards practical molecular structure elucidation: A lightweight multi-spectral CNN-Transformer framework

<title>Abstract</title> <p>Machine learning (ML)-assisted molecular structure elucidation has recently been more advanced in analytical chemistry. However, most current spectra-to-structure methods heavily rely on complex pre- or post-processing, and external knowledge, while starting from a quite low baseline. Herein, we present a lightweight hybrid CNN-Transformer architecture with unlikelihood fine-tuning that establishes a state-of-the-art baseline directly from infrared (IR), Raman, and NMR pure spectral inputs. Without any other processing or knowledge, our model achieves top-1 (top-5) accuracies of 69.14% (89.42%), 78.09% (93.14%), 38.14% (64.28%), 81.88% (95.23%), and 84.23% (96.27%) for IR-only, Raman-only, NMR-only, dual IR + Raman, and triple IR + Raman + NMR spectroscopic inputs, respectively. The model also yields reliable generation results on small experimental datasets. Mechanism analyses reveal that the hierarchical contribution of different spectra is Raman &gt; IR &gt; NMR, and the token acts as both the initiation of SMILES sequence generation and a global aggregation anchor for distinguishing chain and cyclic molecules. This work provides a high-baseline, plug-and-play, and interpretable framework for spectroscopic molecular elucidation, offering new mechanistic insights into ML-based spectral learning.</p>