Symmetry-Informed Graph Neural Networks for Carbon Dioxide Isotherm and Adsorption Prediction in Aluminum-Substituted Zeolites

TL;DR

This work presents SymGNN, a symmetry-informed graph neural network that leverages crystal symmetries to predict CO2 adsorption properties in aluminum-substituted zeolites. The model uses symmetry-guided message passing with FiLM conditioning to share parameters across different zeolite topologies and predicts both the heat of adsorption and the derivative of loading as a function of pressure, enabling isotherm construction. A large synthetic dataset of 27,648 structures across 106 zeolite topologies, generated with four aluminum-placement schemes, is used to evaluate generalization to unseen topologies and to characterize adsorption via structure inference from experimental data using a genetic algorithm. The authors discuss limitations such as data diversity and outline future directions including experimental fine-tuning and generative inverse-design approaches, highlighting the potential impact for rapid adsorption prediction and materials discovery in nanoporous materials.

Abstract

Accurately predicting adsorption properties in nanoporous materials using Deep Learning models remains a challenging task. This challenge becomes even more pronounced when attempting to generalize to structures that were not part of the training data.. In this work, we introduce SymGNN, a graph neural network architecture that leverages material symmetries to improve adsorption property prediction. By incorporating symmetry operations into the message-passing mechanism, our model enhances parameter sharing across different zeolite topologies, leading to improved generalization. We evaluate SymGNN on both interpolation and generalization tasks, demonstrating that it successfully captures key adsorption trends, including the influence of both the framework and aluminium distribution on CO$_2$ adsorption. Furthermore, we apply our model to the characterization of experimental adsorption isotherms, using a genetic algorithm to infer likely aluminium distributions. Our results highlight the effectiveness of machine learning models trained on simulations for studying real materials and suggest promising directions for fine-tuning with experimental data and generative approaches for the inverse design of multifunctional nanomaterials.

Figures (15)

• Figure 1: Heat of adsorption for all datapoints as a function of the aluminium proportion. Note that the color is in log-scale.
• Figure 2: Distribution of loading values at each simulated pressure.
• Figure 3: Loading values for all datapoints with isotherms as a function of the aluminium proportion, at varying pressures.
• Figure 4: Comparison of parameter sharing in symmetry-informed message passing (top row) and symmetry-based parameter sharing (bottom row) across five different zeolite topologies. In the top row, nodes with the same generators are assigned the same color, while in the bottom row, nodes with the same node-update parameters (belonging to the same orbit) share a color. Notably, while symmetry-based parameter sharing results in more distinct colors, symmetry-informed message passing allows certain generator sets to be shared across different zeolites, enabling better transferability.
• Figure 5: The SymGNN architecture. $\square$ denotes the layer input, $\|$ denotes concatenation and $\odot$ denotes elementwise multiplication. $f$ is the ELU activation function, $\sigma$ is the sigmoid activation and $\textbf{sp}$ is the Softplus activation. In the model, atoms and distances between atoms are embedded, following which symmetry-informed message passing takes places. In the output module, the final hidden state is used to predict the heat of adsorption ($y$). By combining the final hidden state, the predicted heat of adsorption and the pressure, the model predicts the derivative of the loading.