A Two-Stage Fragment-Connection Framework for Infrared-Spectrum-Driven Molecular Structure Elucidation

Structure elucidation of molecules from infrared spectra is difficult because the number of possible solutions is very large, direct end-to-end generation is hard to learn, and training is often unstable. In this work, we present a two-stage approach that encodes molecules as Breaking of Retrosynthetically Interesting Chemical Substructures (BRICS) fragments, learns fragment representations from infrared spectra, and then frames the structure elucidation task as a spectrum-based fragment-connection prediction problem between fragment pairs. A complexity-controlled evaluation protocol is proposed to minimize distortion caused by simple molecules with small numbers of fragment nodes. Oracle-fragment and strict-vocabulary analyses further separate fragment-coverage errors from connection errors. The proposed approach successfully produces valid structures and competitive hit rates. Results under the strict-vocabulary protocol show that fragment coverage remains a significant bottleneck, while results in the oracle-fragment setting show improved hit rates but not reduced reconstruction errors, confirming the existence of a dual bottleneck. Finally, the plateau observed between the 5.8k and 63.0k datasets suggests that the primary source of error shifts from fragment coverage and local prediction to higher-order discrimination and ranking for molecular connection generation as data size and molecular complexity increase. In summary, improving structure elucidation guided by infrared spectra requires not only state-of-the-art neural models but also suitable structural abstractions, controlled decoding, and evaluation measures that better reveal reconstruction bottlenecks.