| Oct 26, 2025 |
AI model predicts gas adsorption in MOFs with accuracy and transparency
A new AI model accurately predicts gas adsorption in metal-organic frameworks and explains its results, offering faster, clearer materials discovery.
(Nanowerk News) Evaluating how metal–organic frameworks (MOFs) absorb gases has typically been a slow, expensive process. Laboratory experiments take time and resources, while computer simulations demand heavy computational power that limits large-scale testing. Even advanced deep learning models, such as graph neural networks (GNNs), have struggled with two major issues: adapting to the repeating nature of crystal structures and avoiding overfitting that weakens predictions for new materials.
|
|
Researchers have now developed a model that tackles both problems (Engineering, "A Multi-Scale Graph Neural Network for the Prediction of Multi-Component Gas Adsorption"). Their approach combines two ideas: a multi-scale crystal graph and a multi-scale multi-head attention crystal graph network. The first captures MOF structures across three key spatial ranges tied to gas interaction.
|
|
At the smallest scale, 0–2 angstroms, it analyzes local bonding like metal–oxygen links. The 2–3 angstrom range captures functional groups and surface effects, while the 3–5 angstrom range maps pore shapes and connectivity. To reflect the endless repetition of crystal lattices, the graph includes self-connecting edges that encode periodicity—something traditional GNNs often miss.
|
|
The second component, a network with a multi-head self-attention mechanism, ranks which atomic connections matter most. It gives higher scores to influential bonds such as metal–ligand interactions and filters out background noise. A graph pooling layer, tuned at a 0.6 ratio, keeps the most important features and cuts the rest, reducing overfitting that can skew predictions.
|
 |
| For the detection of structural and functional features across different scales, the sensing distances of MOF atoms are defined. The geometric lengths of these sensing distances correspond to the detection of open metal sites/bonds, functional groups/surfaces, and pore structures/topologies, respectively. (Image: Reprinted from DOI:10.1016/j.eng.2025.08.012, CC BY)
|
|
To test the model, the team used a large database of real MOF structures and gas adsorption data generated through simulation. Results showed striking improvements. For single gases like CO₂, CH₄, and N₂ under standard conditions, prediction accuracy for CO₂ rose more than 200 percent compared with earlier graph attention networks. Overfitting dropped by more than 90 percent relative to another leading model. In gas mixtures such as CO₂:N₂ and CH₄:CO₂, accuracy stayed high. CO₂ predictions improved by about 15 percent, though CH₄ proved harder to model because of its larger size and tetrahedral shape.
|
|
Beyond higher accuracy, the model also offers clearer insights. By visualizing attention weights, researchers could see which structural features drove the results. Bonds like Zn–O at the smallest scale consistently received the highest weights, revealing how the model reached its conclusions. This level of transparency helps address the “black box” problem that often surrounds AI in materials research.
|
|
Overall, this method provides a more reliable and interpretable way to predict gas adsorption in MOFs. By linking atomic structure to performance with greater precision, it could speed up the design of materials for cleaner gas separation and energy storage.
|
