Abstract
We introduce GHP-MOFassemble, a Generative artificial intelligence (AI), High Performance framework to accelerate the rational design of metal-organic frameworks (MOFs) with high CO2 capacity and synthesizable linkers. Our framework combines a diffusion model, a class of generative AI, to generate novel linkers that are assembled with one of three pre-selected nodes into MOFs in a primitive cubic (pcu) topology. The CO2 capacities of these AI-generated MOFs are predicted using a modified version of the crystal graph convolutional neural network model. We then use the LAMMPS code to perform molecular dynamics simulations to relax the AI-generated MOF structures, and identify those that converge to stable structures, and maintain their porous properties throughout the simulations. Among 120,000 pcu MOF candidates generated by the GHP-MOFassemble framework, a total of 102 structures completed molecular dynamics simulations at 1 bar with predicted CO2 capacity higher than 2 mmol/g at 0.1 bar, which corresponds to the top 5% of hMOFs in the hypothetical MOF (hMOF) dataset. Among these candidates, 18 have change in density lower than 1% during molecular dynamics simulations, indicating their stability. We also found that the top five GHP-MOFassemble's MOF structures have CO2 capacities higher than 96.9% of hMOF structures.
Introduction
To mitigate climate change, efficient carbon dioxide ($CO_2$) capture is essential. Metal-Organic Frameworks (MOFs)—porous crystalline materials composed of metal nodes and organic linkers—are promising "molecular sponges." However, the astronomical number of possible MOF configurations makes traditional screening insufficient. We present GHP-MOFassemble, a high-performance framework that leverages generative AI to design novel MOFs with high $CO_2$ capacity and structural stability.
Methods: The GHP-MOFassemble Pipeline
The framework employs a multi-stage approach to move beyond simple database screening toward de novo molecular design.
- Decomposition and Fragment Extraction: The process begins with the hMOF dataset (137,652 structures). High-performing MOFs ($CO_2$ capacity > 2 mmol/g) are identified, and their linkers are extracted and broken into chemically relevant fragments using the Matched Molecular Pairs Algorithm (MMPA).
- Generative Design: Using DiffLinker, an E(3)-equivariant diffusion model, the framework generates novel linkers by "filling in" atoms between fragments. This ensures the creation of unique structures rather than simple recombinations.
- Assembly and AI Screening: New linkers are assembled into MOFs with a primitive cubic (pcu) topology using three metal nodes (Cu paddlewheel, Zn paddlewheel, and Zn tetramer). A modified Crystal Graph Convolutional Neural Network (CGCNN) predicts the $CO_2$ capacity of the resulting 120,000 candidates.
- Physics-Based Validation: The top 6,020 candidates undergo Molecular Dynamics (MD) simulations via the LAMMPS code to evaluate structural stability at 300K and 1 atm.
Key Insights
- Expanding the Chemical Universe: The diffusion model generated linkers with a Tanimoto similarity of only 0.3 to 0.4 compared to known structures, indicating significant novelty.
- The Role of Catenation: Catenated (interpenetrated) frameworks often showed superior performance due to smaller pore sizes that enhance $CO_2$ trapping at low pressures.
- Stability as a Bottleneck: Only a small fraction of high-performing candidates maintained structural integrity during MD simulations, emphasizing the importance of physics-based validation in AI-driven discovery.
Results
The GHP-MOFassemble framework successfully generated 120,000 unique MOF candidates. AI screening identified 6,020 structures exceeding the performance threshold of 2 mmol/g.
Following rigorous MD simulations, 102 structures remained stable, with 18 candidates exhibiting exceptional stability (less than 1% density change). Notably, the top five generated structures outperformed 96.9% of the existing hMOF database.
Conclusion and Future Work
GHP-MOFassemble demonstrates the power of combining generative AI with extreme-scale computing for material discovery. Future work will expand the framework to include diverse topologies and metal nodes, alongside experimental validation of the most promising stable candidates.
Summary updated