[Paper Review] MOFDiff: Coarse-grained Diffusion for Metal-Organic Framework Design

Xiang Fu, Tian Xie+3 more|arXiv (Cornell University)|Oct 16, 2023

Metal-Organic Frameworks: Synthesis and Applications13 citations

TL;DR

MOFDiff introduces a coarse-grained diffusion model to generate 3D MOF structures and an assembly algorithm to recover all-atom structures, enabling novel MOFs for carbon capture.

ABSTRACT

Metal-organic frameworks (MOFs) are of immense interest in applications such as gas storage and carbon capture due to their exceptional porosity and tunable chemistry. Their modular nature has enabled the use of template-based methods to generate hypothetical MOFs by combining molecular building blocks in accordance with known network topologies. However, the ability of these methods to identify top-performing MOFs is often hindered by the limited diversity of the resulting chemical space. In this work, we propose MOFDiff: a coarse-grained (CG) diffusion model that generates CG MOF structures through a denoising diffusion process over the coordinates and identities of the building blocks. The all-atom MOF structure is then determined through a novel assembly algorithm. Equivariant graph neural networks are used for the diffusion model to respect the permutational and roto-translational symmetries. We comprehensively evaluate our model's capability to generate valid and novel MOF structures and its effectiveness in designing outstanding MOF materials for carbon capture applications with molecular simulations.

Motivation & Objective

• Overcome the limitations of template-based MOF design by enabling generation beyond predefined topologies and building blocks.
• Develop a coarse-grained representation of MOFs using building blocks and a contrastive embedding to capture diversity.
• Formulate a diffusion process over the coarse-grained MOF representation and an assembly algorithm to recover all-atom structures.
• Evaluate validity, novelty, and diversity of generated MOFs, and assess performance for carbon capture via molecular simulations.

Proposed method

• Derive a coarse-grained MOF representation with building-block identities and coordinates and lattice parameters.
• Employ a contrastive GemNet-OC encoder to embed building blocks into a compact latent space.
• Use a periodic GemNet-OC denoiser to perform conditional diffusion over CG MOF structures.
• Decode building-block identities by nearest-neighbor lookup in the learned embedding space.
• Apply an assembly algorithm that orients building blocks to maximize overlap of connection-point Gaussians, followed by force-field relaxation (UFF) to obtain all-atom MOFs.
• Optionally predict properties from latent codes with an MLP to enable property-driven inverse design.

Experimental results

Research questions

• RQ1 Can MOFDiff generate valid and novel MOF structures beyond predefined templates?
• RQ2 Can the diffusion-generated CG MOF structures be assembled into valid all-atom MOFs that relax to physically plausible structures?
• RQ3 Do MOFDiff-designed MOFs exhibit enhanced CO2 adsorption properties in molecular simulations compared to the BW-DB baseline?
• RQ4 Is latent-space optimization able to improve targeted properties such as CO2 working capacity?

Key findings

• Out of 10,000 random latent samples, 5,865 matches of connection points were obtained; 3,012 were valid; and 2,998 were valid, novel, and unique.
• Generated MOFs cover a distribution of structural properties similar to BW-DB, indicating diverse structural space coverage.
• MOFDiff-optimized MOFs show higher CO2 working capacity and greater CO2 selectivity and uptake in GCMC simulations compared to BW-DB baselines.
• Using MOFDiff reduces the number of required GCMC simulations to find a MOF with >2 mol/kg CO2 working capacity from 58.1 to 14.6, a substantial efficiency gain.
• Top MOFs discovered by MOFDiff include candidates competitive with literature MOFs for carbon capture (e.g., aligning with known top performers in BW-DB).