Traditional experimental methods for evaluating gas adsorption performance of metal–organic frameworks (MOFs) are costly and time-consuming, while simulation techniques such as molecular dynamics and grand canonical Monte Carlo have high computational demands that limit large-scale screening. Existing deep learning models, including graph neural networks (GNNs) like graph convolutional networks and crystal graph convolutional neural networks, face two key limitations: One is difficulty adapting to the inherent periodicity of crystal structures, which is a defining feature of MOFs, and the other is a tendency to overfit training data that compromises generalization to new materials.​
To overcome these issues, researchers proposed two core innovations: a multi-scale crystal graph and a multi-scale multi-head attention crystal graph network. The multi-scale crystal graph captures MOF structural features across three spatial scales, each tied to gas interaction. The 0–2 Å scale focuses on local bonding such as open metal sites and metal–oxygen bonds, the 2–3 Å scale on functional groups and surface interactions, and the 3–5 Å scale on pore structure and topology. This graph uses self-connecting edges to encode crystal periodicity, ensuring the model accounts for infinite lattice repetition. This design is an improvement over traditional GNNs that ignore periodicity. ​
Beyond the precise structural characterization enabled by the multi-scale crystal graph, the accompanying network leverages a multi-head self-attention mechanism to prioritize meaningful atomic interactions and filter redundant noise. It assigns attention scores to crystal graph edges, focusing on relevant features like metal–ligand bonds that drive gas adsorption. It also includes a graph pooling layer optimized at a ratio of 0.6. This layer refines feature extraction by retaining high-impact edges, which helps mitigate overfitting.​
Despite these structural and mechanistic optimizations targeting traditional method limitations, the model’s practical utility and superiority required rigorous experimental validation. To carry out this validation, the CoRE MOF database and adsorption data simulated via grand canonical Monte Carlo are employed. The CoRE database contains experimentally synthesized materials, and the simulated data is generated via RASPA 2.0 with UFF and TraPPE force field. This validation demonstrates the model’s superiority. For single-component adsorption involving CO₂, CH₄, and N₂ at 298 K and 6–30 bar, the model’s CO₂ prediction accuracy are significantly improved: The coefficient of determination (R²) rises over 200% compared with graph attention networks, and overfitting is reduced by more than 90% compared with materials graph network (MEGNet). MEGNet previously had severe overfitting issues. In multi-component mixtures such as CO₂:N₂ and CH₄:CO₂, the model maintains high accuracy. R² for CO₂ in CO₂:N₂ mixtures improves approximately 15% compared with conventional GNNs. However, CH₄ prediction was more challenging due to its larger size and tetrahedral structure, resulting in a 10% lower accuracy improvement than CO₂.
Notably, the model not only improves predictive performance but also enhances interpretability through attention weight visualization. This visualization links high weights to key features such as Zn–O bonds at the 0–2 Å scale. This transparency addresses criticisms of black-box AI in materials science and helps researchers understand prediction logic.​
In conclusion, this study addresses the challenges of periodic adaptation and model overfitting in predicting the adsorption performance of MOFs, and proposes a relatively robust and interpretable solution for predicting adsorption performance. This prediction scheme is expected to accelerate the research and development process of material structures suitable for efficient gas separation and storage, thereby promoting the practical application of such functional materials.
The paper “A Multi-Scale Graph Neural Network for the Prediction of Multi-Component Gas Adsorption,” authored by Lujun Li, Haibin Yu. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.08.012.