Preprint tests AI generation and repair of molecular structures from spectra

Researchers report in a preprint that an AI framework can propose molecular structures for compounds absent from reference spectral libraries and explain its choices using mass spectral evidence. The system, DiffGCMS, generates candidate structures from measurements of molecular fragments; a language model then validates, repairs and ranks those candidates. On 13,696 spectra from NIST 20, the generative model placed the correct structure first in 6.01% of cases and within its first ten candidates in 15.76%. For molecules with no more than 10 heavy atoms, language-model processing increased candidate validity from 91.04% to 100%, but first-choice accuracy rose only from 21.28% to 21.95%.