Abstract
Zeolites are a class of crystalline nanoporous materials known for their ability to discriminate molecules based on size and shape. Such molecular shape selectivity arises from the precise 3-dimensional arrangement of zeolite framework atoms and the resulting non-covalent interactions. In this work, Henry’s constants ($k_H$) for n-octadecane adsorption in all-silica zeolites were used as the target property in a systematic effort to optimize and evaluate various machine-learning representations widely used for materials modeling. ZeoNet was found to vastly outperform other representations, achieving a correlation coefficient of $r^2 = 0.973$ and a mean-squared error (MSE) of 4.4 in $\ln k_H$. In comparison, the best-performing graph model for this task, M3GNet, obtained $r^2 = 0.888$ and MSE = 18.5, reflecting the difficulty of graph models to capture subtle structural variations and long-ranged spatial correlations. ZeoNet also exhibits excellent transferability to other hydrocarbon molecules, including mono-and di-branched C18 isomers and linear C24 and C30 alkanes. Finally, multi-task learning was shown to improve the accuracy in predicting adsorption selectivity without compromising the prediction of Henry’s constants for individual molecules.
Introduction
Zeolites serve as critical "molecular sieves" in the chemical industry, enabling the separation of molecules based on subtle differences in size and shape. This shape selectivity is fundamental to the production of high-quality fuels and lubricants. Traditionally, identifying the optimal zeolite framework for a specific molecular target requires computationally intensive physics-based simulations, such as Monte Carlo methods.
While machine learning (ML) offers a path toward accelerating these discoveries, the optimal structural representation for zeolites remains a subject of debate. Most materials science ML models rely on Graph Neural Networks (GNNs), which emphasize local atom-to-atom connectivity. However, this study investigates whether such local representations are sufficient for predicting the adsorption of long, flexible hydrocarbon chains, where the global "void space" and pore topography are paramount.
Methods and Data
The researchers performed a comprehensive comparison of various structural representations to predict Henry’s constants ($k_H$) for n-octadecane and other long-chain alkanes (C18 to C30). The representations evaluated included:
- 3D Volumetric Grids (ZeoNet): Utilizing 3D voxels to represent distance grids (distance to the nearest surface), atom occupancy, or nearest-atom types.
- Graph Neural Networks (GNNs): Industry standards such as CGCNN, MEGNet, M3GNet, and MACE.
- Alternative Architectures: 2D multi-view images, Vision Transformers (ViTs), and Point Clouds.
The models were trained on a massive dataset of approximately 100,000 zeolite structures sourced from the IZA and PCOD databases. Ground truth labels were generated using high-fidelity Monte Carlo simulations for six different linear and branched hydrocarbon isomers.
Results and Insights
The study found that 3D Convolutional Neural Networks (ConvNets), specifically the ZeoNet architecture using distance grids, significantly outperformed all other models. ZeoNet achieved a correlation coefficient ($r^2$) of 0.973 and a Mean-Squared Error (MSE) of 4.4 in $\ln k_H$. In contrast, the best-performing graph model, M3GNet, achieved an $r^2$ of 0.888 and an MSE of 18.5.
The superior performance of ConvNets suggests that capturing long-range spatial correlations and the global "envelope" of the pore is more critical for long-chain adsorption than local connectivity. Distance grids proved particularly effective by providing a smooth field representing pore topography.
Furthermore, the researchers demonstrated the power of advanced training strategies:
- Transfer Learning: Pre-training ZeoNet on one molecule (e.g., linear C18) allowed for high-accuracy fine-tuning on different molecules (e.g., branched isomers) using 80% less data.
- Multi-task Learning: Training a single model with a shared representation for multiple molecules improved the accuracy of adsorption selectivity predictions without sacrificing individual molecule performance.
Discussion and Conclusion
While ZeoNet shows remarkable accuracy and efficiency—predicting properties in 0.13 seconds compared to hours for traditional simulations—some challenges remain. The model's performance slightly decreased for the longest chains (C30), likely due to the fixed 45 Å grid size. Additionally, while data augmentation was used to address rotational invariance, future architectures might incorporate this property natively.
In conclusion, this work highlights that for complex molecular sieving applications, 3D volumetric representations are superior to graph-based models. By focusing on the global shape of the void space, researchers can more accurately and rapidly design zeolites for critical industrial separations.
Summary updated