Functional Group Identification from Infrared Spectra via a Dual-Head Convolutional Network

Infrared (IR) spectroscopy is widely used for molecular characterization, yet automated functional group identification remains challenging due to spectral overlap, weak absorptions, and context-dependent vibrational features. In this work, we present a dual-head convolutional neural network for multi-label functional group identification directly from single infrared spectra. The proposed architecture captures complementary local and global spectral information through parallel convolutional pathways with cross-scale alignment. The model is evaluated on NIST-derived gas-phase IR datasets covering 21 functional groups using five-fold cross-validation. It achieves a mean macro-averaged F1 score of 0.9169, indicating improved performance across functional groups. In addition to the overall results, the proposed architecture shows improved recognition for several challenging functional groups, as reflected by per-class evaluation and ablation analysis. Overall, this work provides a scalable baseline for IR-based functional group identification and clarifies both the capabilities and current limitations of single-modality infrared analysis, offering a foundation for future extensions toward multi-modal spectroscopic learning approaches.