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

Infrared (IR) spectroscopy is fundamental to molecular characterization, yet automated functional group identification remains challenging because of spectral overlap, weak absorptions, and context-dependent vibrational features. In addition to these spectral difficulties, the high computational cost of model training limits practical adoption. This constraint matters in practice because changing research objectives often require redefining target functional groups, which can trigger repeated model retraining. For real-world use, predictive accuracy must therefore be balanced with computational efficiency on office-grade hardware. To address these needs, we propose a physically motivated dual-head CNN that jointly models local diagnostic bands and global spectral context, enabling efficient and accurate functional group identification from IR spectra. We evaluate the proposed model on multiple NIST-derived data sets and externally published benchmarks, and we compare it with representative mainstream methods. The results show consistent performance gains and strong cross-data set generalization. When the original 21 functional groups are expanded to 35 classes with finer-grained subgroups, the model maintains strong performance, which indicates good scalability to more detailed chemical taxonomies. Feasibility tests on office-grade devices further confirm that the model can be trained within practical time budgets without specialized hardware. Overall, these findings suggest that the proposed approach achieves robust predictive performance, reliable cross-data set generalization, and scalable adaptability to expanded functional group systems, while remaining suitable for everyday laboratory deployment.