With the rapid development of industrialization, large amounts of toxic and harmful gases such as NO2, CO, and NH3 are emitted during industrial production, posing serious threats to human health and the ecological environment. Therefore, achieving real-time and accurate detection of gas types and concentrations in the environment is of great significance for industrial safety and public health. Gas sensors, known for their simple structure, rapid response, and high sensitivity, are widely used in industrial safety, environmental monitoring, and intelligent systems. However, the screening of gas-sensing materials still heavily relies on experience-driven trial-and-error methods, which are costly and time-consuming. Although first-principles calculations can reveal gas adsorption behavior at the atomic scale, single descriptors are insufficient to fully capture the complex gas–material interactions, while multi-parameter coupling introduces highly nonlinear problems, severely limiting efficient screening and prediction capabilities. Thus, constructing a high-throughput prediction framework that integrates multi-dimensional descriptors and intelligent algorithms is of significant importance for the rapid screening and rational design of gas-sensing materials.
Recently, a team of material scientists led by Guiwu Liu from Jiangsu University, China combined first-principles calculations with machine learning offers a promising research path for predicting gas-sensing performance. In this study, black phosphorus is used as the model system. A gas-sensing response database is constructed based on multi-source literature data, and a binary classification strategy is adopted to reduce bias caused by experimental inconsistencies. Furthermore, key electronic structure descriptors extracted from first-principles calculations are used to build high-precision machine learning prediction models.
The team published their work in Journal of Advanced Ceramics on January 06, 2026.
The first author, Mingyuan Wang from Jiangsu University, stated "this paper presents four highlights: (1) Integration of Physics and Artificial Intelligence: A novel collaborative framework is proposed that combines first-principles calculations with machine learning, enabling rapid and high-precision prediction of gas-sensing performance. This effectively overcomes the limitations of traditional experimental screening methods, including long time consumption, high labor intensity, and high costs. (2) High Accuracy and Robustness: Using black phosphorus as the model material and based on a dataset of 21 gases, six machine learning models were systematically trained and evaluated. Among them, the Extra Trees model performed particularly well, achieving a prediction accuracy of 96% in five-fold cross-validation and F1 scores of 1.000 and 0.933 in independent test sets, respectively. (3) Mechanistic Analysis Based on SHAP: Feature importance analysis and SHAP interpretation results indicate that adsorption energy, p-band center, vacuum level, valence band maximum, and conduction band minimum are key descriptors determining gas-sensing responses. These factors govern the gas response process by modulating band structures and carrier transport behavior. (4) Visualization System for Gas-Sensing Prediction: A lightweight, Python-based prediction and visualization system was developed. By inputting only five key features obtained from first-principles calculations, this system enables real-time assessment of the response trends of black phosphorus to different gas molecules."
The corresponding author Guiwu Liu stated that: "while the current model is limited to black phosphorus, it is capable of predicting responses to multiple gases. The next step will involve extending this model to a broader range of materials to enable the prediction of gas-sensing responses across different materials and gas types."
Other contributors include Yaqi Zhang, Bowen Xiong, Xiangzhao Zhang, Jian Yang, Lin Xu, Guanjun Qiao from Jiangsu University in Zhenjiang, China; Ke Wang from Chinese Academy of Sciences in Beijing, China.
This work was supported by the Key Research and Development Plan of Jiangsu province (No.BE2019094), and the Innovation/Entrepreneurship Program of Jiangsu Province (No.JSSCTD202146). The first author acknowledges the support from the Scientific Research Start-up Funds of Jiangsu University for this work.
About Author
Mingyuan Wang is a qualified Associate Professor at Jiangsu University. He received his Ph.D. in Electronic Science and Technology from Southeast University in September 2024, where he conducted joint doctoral research at Uppsala University, Sweden, from March 2023 to March 2024. In April 2025, he joined the School of Mechanical Engineering at Jiangsu University. His research focuses on wearable sensors, nanoenergy devices, machine learning, and first-principles calculations. He has published over 30 SCI papers as first/co-first or corresponding author in journals such as Advanced Materials, Advanced Science, ACS Nano, Journal of Advanced Ceramics, Chemical Engineering Journal, Sensors and Actuators B: Chemical, Fuel, Chemistry of Materials, and Journal of Materials Chemistry C. Among these, he is the first author of 11 papers, including 5 highly cited articles. His SCI publications have been cited over 1,700 times, with an H-index of 24.
About Journal of Advanced Ceramics
Journal of Advanced Ceramics (JAC) is an international academic journal that presents the state-of-the-art results of theoretical and experimental studies on the processing, structure, and properties of advanced ceramics and ceramic-based composites. JAC is Fully Open Access, monthly published by Tsinghua University Press, and exclusively available via SciOpen . JAC's 2024 IF is 16.6, ranking in Top 1 (1/33, Q1) among all journals in "Materials Science, Ceramics" category, and its 2024 CiteScore is 25.9 (5/130) in Scopus database. ResearchGate homepage: https://www.researchgate.net/journal/Journal-of-Advanced-Ceramics-2227-8508
