Multimodal deep learning-driven exploration of lanthanide-based perovskite oxide semiconductors for ultra-sensitive detection of 2-butanone

Abstract

The detection of volatile organic compounds (VOCs), particularly 2-butanone, in exhaled breath has been regarded as a potential non-invasive method for the early identification of gastric cancer. Lanthanide-doped perovskite oxide semiconductor (POS) gas sensing materials have garnered significant research interest in the field of respiratory gas detection. In order to address the challenges encountered in traditional studies, this research adopted a deep learning architecture that integrates natural language processing technology (Word2-Vec) with crystal graph convolutional neural networks (CGCNN), proposing a comprehensive dataset prediction strategy that encompasses one million literature abstracts and 110,000 crystal structure data entries. The objective is to develop sensing materials with high sensitivity, selectivity, and stability for the precise detection of 2-butanone. The study meticulously designed and screened the Tm(III)/Rh(III)/5-(1H-imidazole-1-yl)iso-phthalic acid complex, revealing the regulatory effect of ligand concentration on the n-type to p-type conversion of the thulium-rhodium oxide sensing performance, thoroughly investigating the mechanism of the sensing performance transition, and enhancing the precision of controlling the properties of the sensing materials. Even in highly interfering and humid environments (relative humidity RH = 90 %), the TmRhO3 material with a functionalized perovskite structure demonstrated exceptional sensing performance for 2-butanone gas, with a detection limit reaching 50 parts per billion (ppb). This comprehensive research approach has facilitated the construction of a multimodal-enhanced deep learning framework, which holds significant advantages in predicting the performance of perovskite oxide semiconductors and their precise control. Its effectiveness has been confirmed through extensive experimental data.