Perovskite materials have emerged as a focal point in materials science due to their compositional diversity and structural uniqueness. However, the vast combinatorial possibilities of these materials pose challenges for traditional experimental and computational methods in identifying the optimal compounds. This study aims to systematically explore their application potential through a transferable multimodal deep learning approach. By formatting and standardizing one million abstracts from materials science literature using predefined rules and functions, the processed data becomes suitable for further analysis or machine learning tasks such as text classification and entity recognition. Compound vector representations were constructed and designed, followed by training lanthanide metal ions, transition metal ions, and ligand word vectors using the Word2Vec model. The correlations between compounds and keywords such as gas sensors, perovskites, heterojunctions, and crystal structures were computed. These correlation rankings provided foundational predictions for the crystal graph convolutional neural network (CGCNN) model. A dataset of 110,000 crystallographic information files (CIF) was loaded using predefined rules and functions to train and evaluate the CGCNN model. The processed data was used to predict the properties and characteristics of materials identified by the Word2Vec model. By loading the CIF files of the predicted materials, structural objects were created, surfaces generated, and adsorption model diagrams plotted to assess the optimal adsorption surfaces and sites, thereby achieving transferable multimodal deep learning. Experimental validation of the predicted results demonstrated excellent sensing performance. Our findings indicate a high correlation between the predicted properties and experimental results, validating the robustness of our approach.