A newly developed multitask deep learning framework can simultaneously predict how strongly a material adsorbs sulfur gases and how effectively it can sense them, according to a study published in Artificial Intelligence & Environment. By combining artificial intelligence with density functional theory calculations, the approach accelerates the discovery of materials designed for hazardous gas detection and purification.
Multitask AI Predicts Adsorption and Sensing in Transition-Metal Phthalocyanines
Toxic sulfur-containing gases like hydrogen sulfide and sulfur oxides present major risks to human health, industrial safety, and ecosystems. Finding materials that balance strong adsorption with reliable sensing remains difficult. To tackle this, researchers investigated 28 transition-metal phthalocyanine materials and four sulfur-based gases: H₂S, SO, SO₂, and SO₃. This work generated a dataset of 78 gas-material adsorption systems, according to findings published in the journal.
Transition-metal phthalocyanines are materials whose electronic properties change by swapping the metal atom at their center. Xiliang Yan, a corresponding author of the study, explained the core challenge of modeling these interactions. “Gas sensing and gas adsorption are closely connected, but they are not governed by exactly the same factors,” Yan said. Traditional machine learning typically evaluates one property at a time. In contrast, the new model learns shared physicochemical information while maintaining separate branches for adsorption energy and sensing response.
Did you know? Transition-metal phthalocyanines allow scientists to tune electronic behavior simply by altering the single metal atom sitting at the center of the molecular structure.
Performance Metrics and Candidate Materials Identified by AI
The multitask deep learning framework demonstrated clear improvements over conventional single-task machine-learning models when tested on independent data sets. For classifying high and low sensing responses, the multitask model reached an accuracy and F1 score of 0.83 alongside a recall of 0.88. Older single-task models achieved test accuracies ranging from only 0.44 to 0.63, according to the published data. For predicting adsorption energy, the model yielded a test-set R² of 0.86, an RMSE of 0.46 eV, and an MAE of 0.35 eV.
These computational methods also pointed toward specific material candidates for practical applications. Calculations showed that Fe/Pc displays strong sensitivity toward SO. Additionally, Fe/Pc and Cr/Pc exhibited promising responses to H₂S. Other transition-metal phthalocyanines showed strong adsorption capabilities for SO and SO₂, which suggests they could work effectively as gas-removal materials.
Interpretable Machine Learning for Rational Materials Design
Rather than acting as a black-box prediction tool, the AI model was built to be interpretable. Analysis revealed that the atomic radius and electronic properties of the transition-metal center, the electronic structure of the phthalocyanine, and the characteristics of the gas molecule all dictate sensing and adsorption behavior. “The model does more than identify promising candidates,” Yan said, noting that it uncovers the physical mechanisms controlling these processes to guide future material design.
The research team acknowledges that their current model relies on a relatively small dataset. Expanding the training data to include additional metals, gases, and adsorption configurations will likely enhance its capability to screen previously unexplored materials. The study was authored by Y. Li, Y. He, D. Zhou, and colleagues under the title “DFT-driven multitask deep learning for sulfur gas sensing and removal by transition metal phthalocyanines” in Artificial Intelligence & Environment (2026, Vol. 1, Issue 3).
Frequently Asked Questions
What gases were evaluated in the study?
The researchers investigated four sulfur-based gases: hydrogen sulfide (H₂S), sulfur monoxide (SO), sulfur dioxide (SO₂), and sulfur trioxide (SO₃).

How did the multitask model perform compared to older models?
The multitask model achieved an accuracy and F1 score of 0.83 for sensing response classification, outperforming conventional single-task models that scored between 0.44 and 0.63.
Which materials showed the strongest potential?
Calculations identified Fe/Pc as showing strong sensitivity to SO, while Fe/Pc and Cr/Pc displayed promising responses to H₂S.
Stay Updated on Environmental AI Research
Explore more breakthroughs in computational materials science and artificial intelligence by subscribing to our research updates.
Keep reading