CWAT: Chemical Warfare Agent Transformer : A Machine Learning Approach to GC-MS-EI

Sammanfattning

Identifying unknown chemical compounds from mass spectra is a fundamental challenge in forensic chemistry, particularly when the target compound is absent from standard reference libraries. This thesis presents CWAT (Chemical Warfare Agent Transformer), a transformer-based sequence-to-sequence model that translates GC-MS-EI spectra directly into SMILES molecular strings. CWAT is trained jointly on the NIST Mass Spectral Library (∼266 000 compounds) and a restricted CWA-related dataset (∼6 000 compounds) provided by the Swedish Defence Research Agency (FOI). On the CWA test set (n=978), CWAT achieves a Top-1 exact match rate of 15.3%, a Top-5 rate of 29.3%, and a mean MACCS Tanimoto similarity of 0.86. Functional group detection is particularly strong for forensically critical elements such as phosphorus, chlorine, sulphur and fluorine. On the broader NIST test set (n=978), the model achieves a mean MACCS Tanimoto of 0.53 with 100% chemical validity. To the best of my knowledge, these are the first reported de~novo structure prediction results for chemical warfare related agent spectra.