Multiscale Analysis of Deep Learning and Machine Learning: New Insights into the Adsorption Mechanism of VOCs Gas-Sensitive Materials

Volatile organic compounds (VOCs) associated with lung cancer are key biomarkers for early noninvasive diagnosis, yet their adsorption behaviors on sensing materials remain highly complex and material-dependent. Efficient screening and accurate prediction of adsorption performance are therefore essential for designing next-generation gas sensors. Nanocomposites, with their superior surface reactivity and tunable properties, show great potential but lack a universal predictive framework that integrates computational simulations with intelligent algorithms. To overcome this limitation, this work constructs a comprehensive dataset of 336 adsorption cases and integrates first-principles calculations with machine learning to systematically predict VOC adsorption energies on nanocomposites. Eight algorithmsincluding SVR, GBR, GPR, XGBoost, MLP, KRR, and a small-sample Transformer modelwere benchmarked to identify the optimal predictive strategy. Among them, the KRR model achieved the best performance with an R2 of 0.8997 on the test set, exhibiting excellent generalization capability. This study provides the first comparative evaluation of deep learning and traditional ML methods for VOC adsorption prediction on nanocomposites based on first-principles data, revealing their respective strengths and limitations in gas-sensing research. The established universal predictive model offers a powerful tool for rapid screening of lung-cancer-related VOC biomarkers and lays a solid theoretical foundation for the rational design of high-performance gas sensors in medical diagnostics and health monitoring.