A data-driven multimodal deep learning framework for predicting gas-sensing performance of two-dimensional materials

Two-dimensional (2D) materials show promise for gas sensing, yet efficient prediction remains challenging due to trial-and-error experiments. In this study, we integrated multimodal data with deep learning to predict the sensing responses of 2D material-based gas sensors. By compiling data from the literature and relevant databases, we constructed a curated dataset comprising 254 samples covering 10 distinct 2D materials and 14 target gases. Three orthogonal structural views of materials and gases were used as structural modalities, coupled with physical properties and dynamic response features to develop multimodal deep learning models via early and intermediate fusion. Five-fold cross-validation showed that the intermediate fusion strategy outperformed early fusion, achieving a mean R 2 of 0.88 and a mean MAE of 0.76. Furthermore, SHAP analysis indicated that dynamic response features contributed the most to the predictions, with three geometry-derived structural features ranking among the five most important features. Commercial MoS₂ sensors experimentally confirmed the model's predicted gas selectivity, with triethylamine (TEA) response about four times higher than for HCHO and CH₃OH, consistent with observations, and with minimum prediction errors of about 3% for HCHO and CH₃OH at 380 K and about 8% for TEA at 407 K. DFT calculations confirmed physisorption of HCHO, CH₃OH, and TEA on MoS₂, with TEA exhibiting the strongest adsorption. TEA also induced the largest band-gap reduction, providing a microscopic explanation for its higher sensing response. This study provides a data-driven framework for accelerating the evaluation and performance prediction of 2D gas-sensing materials, supported by experimental and theoretical validation.