Deep learning-assisted chemiresistive sensor arrays for postharvest maturity identification via fruit-emitted mixed gas monitoring

Target gas selection was based on the GC–MS test of mango fruit during storage. • A 4-sensor array consisted of Pd-SnO 2 exhibited good performance for target gases. • The TabPFN model was used for mixed gas detection based on the sensor array. • The sensing mechanism was explained by characterization and DFT calculations. • Mango maturity identification was achieved by the sensor-algorithm integrated system. Plants maintain continuous gaseous exchange with their surroundings, and the volatile organic compounds (VOCs) they emit can act as indicators of plant physiological status. As an important plant organ, fruit maintains active physiological metabolism, including during the postharvest stage. Therefore, the detection of volatile compounds in postharvest fruits can be used to evaluate the quality and monitor their spoilage during storage and transportation. However, current gas detection instruments are mostly benchtop devices that are bulky and costly, making them unsuitable for practical industrial applications. Thus, developing gas-based rapid maturation assessment technologies for fruits is growing increasingly essential. In this study, we developed an intelligent monitoring system for fruit ripening process monitoring, with mangoes chosen as the research subject. Specifically, ethylene, ethanol, ethyl acetate and dipentene were identified as four target gases based on the detection results of VOCs released by mangoes during storage via solid-phase microextraction combined with gas chromatography–mass spectrometry. To detect these signals, a 4-sensor array was constructed using SnO 2 nanospheres functionalized with palladium nanoparticles, leveraging resistance variations to capture the dynamics of gases. Density functional theory calculations and comprehensive characterization were employed to elucidate the sensing mechanism. The Tabular Prior-data Fitted Network, a Transformer-based deep learning model, was used to analyze four target gases and their multicomponent mixtures. This method achieved high-precision classification (94.8%) for the gases and their mixtures, as well as accurate concentration regression (R 2  = 0.839) for the target components. Finally, the system was integrated into a portable platform, enabling the on-site identification of mango maturity across different storage durations. This work establishes a robust framework for sensor design and intelligent fruit quality monitoring. Download: Download high-res image (194KB) Download: Download full-size image