DFT-driven multitask deep learning for sulfur gas sensing and removal by transition metal phthalocyanines

Accurate prediction of both gas sensitivity and adsorption capacity is essential for developing high-performance sensing and purification materials for sulfur-based toxic gases. However, these two targets are governed by multiple interrelated physicochemical factors originating from both the sensing material and gas molecules, making it challenging for experimental approaches and traditional machine learning methods to simultaneously capture their shared physicochemical information and task-specific characteristics. Herein, we develop a multi-task deep learning (MTDL) framework that simultaneously learns and shares representations between sensitivity and adsorption tasks to elucidate their cooperative mechanisms in transition-metal phthalocyanines (TM/Pc). Integrated with density functional theory (DFT) calculations, the MTDL model enables joint prediction of adsorption energies (E_ads) and sensing responses (SR) toward H₂S, SO, SO₂, and SO₃, achieving improved predictive performance compared with single-task models. Model interpretability analysis reveals that the predictions are governed by a coherent combination of metal-centered electronic properties, substrate-level electronic structure descriptors, and gas-phase electronic characteristics, providing a mechanistic insight into how adsorption–desorption equilibria and charge redistribution collectively determine sensing performance. Overall, this study demonstrates that multi-task learning effectively bridges adsorption thermodynamics and gas sensing response, offering a generalizable and interpretable strategy for AI-driven design of multifunctional gas-sensing and purification materials.