Single-Atom Palladium Functionalized Reduced Graphene Oxide Coupled with a Physics-Data Dual-Driven Deep Learning Framework for Interpretable Discrimination of Volatile Organic Compound Homologues.

Abstract

Precise discrimination of structural homologues, namely molecules with highly similar physicochemical properties, remains a critical bottleneck in artificial olfaction and limits its application in precision medicine and environmental monitoring. While deep learning has enhanced pattern recognition, current purely data-driven models often lack physical interpretability and may show limited transferability to homologous analytes not included during training. Here, we present a physics-data dual-driven deep learning framework that integrates time-resolved sensor responses with DFT-derived molecular descriptors, thereby bridging atomic-level electronic information and macroscopic sensing responses for VOC classification. The model achieved direction-averaged accuracy and macro-F1 scores of 83.88 and 83.53%, respectively, in bidirectional validation involving unseen homologues from five chemical categories. Interpretability and ablation analyses further confirmed that the complete DFT-assisted framework contributed substantially to classification. These results demonstrate the potential of integrating microscopic physicochemical descriptors with dynamic sensing signals to recognize unseen VOCs according to their chemical categories, providing a promising route toward interpretable VOC sensing within defined chemical categories.