AI–DFT-DNN-guided dopant–ligand engineering of MOF-derived chemiresistors for ppb-level 2,5-Dimethylpyrazine detection

Abstract

Pollination by bees underpins agriculture and ecosystem health but is increasingly threatened by American foulbrood (AFB) caused by Paenibacillus larvae. Conventional AFB diagnostics are slow and costly, underscoring the need for portable sensors targeting the key biomarker 2,5-dimethylpyrazine (2,5-DMP). Here, natural language processing is applied to mine a large corpus of scholarly literature. A domain-adaptive SensBERT model evaluates 514,088 bimetallic ion–organic ligand combinations for gas sensing, with similarity scores near unity indicating high applicability. Integrated analysis of application context, physicochemical traits, and MOFs performance identifies three optimal systems: Sm(III)/Sn(II) with 3,3′,3″-((1,3,5-triazine-2,4,6-triyl)tris(azanediyl))tribenzoic_acid (H3TATAB); In(III)/Zn(II) with 4-(1H-tetrazol-5-yl)benzoic_acid (H2TZBA); and Tm(III)/Sn(II) with furan-2,5-dicarboxylic_acid (FDCA). Sm/Sn–H3TATAB exhibits superior sensitivity and selectivity toward 2,5-DMP. After plasma conversion, the derived metal oxides retain highly specific 2,5-DMP recognition with excellent response consistency and stability. Density functional theory (DFT) calculations model adsorption of multiple AFB-related VOCs at the atomic scale, revealing a uniquely strong binding preference for 2,5-DMP; coupling DFT-generated data with a deep neural network further clarifies the sensing mechanism. The resulting sensor achieves a detection limit of 20 ppb for 2,5-DMP, with a response of 2.7 at 20 ppb. This work establishes a synergistic paradigm that integrates deep learning, first-principles computation, and experimental validation to deliver a high-performance, non-invasive platform for rapid AFB detection, highlighting the promise of cross-disciplinary strategies for environmental monitoring and agricultural biosecurity.