Building-Block Aware Generative Modeling for 3D Crystals of Metal Organic Frameworks

TL;DR

This work tackles the combinatorial MOF design problem by introducing Building-Block-Aware MOF Diffusion, a $SE(3)$-equivariant diffusion model that learns 3D all-atom representations of MOF building blocks and nets from the CoRE MOF dataset. By operating on building blocks rather than full unit cells, it samples large MOFs (up to ~1000 atoms) with high geometric validity and novelty, including new inorganic nodes and organic edges. The authors validate their approach by synthesizing a high-scoring MOF, [Zn(1,4-TDC)(EtOH)2], whose structure is confirmed by PXRD, TGA, and N2 sorption, demonstrating practical synthesizability. The method expands accessible MOF chemical space and offers a scalable path for AI-guided material design, with future potential for broader nets and property-guided generation.

Abstract

Metal-organic frameworks (MOFs) marry inorganic nodes, organic edges, and topological nets into programmable porous crystals, yet their astronomical design space defies brute-force synthesis. Generative modeling holds ultimate promise, but existing models either recycle known building blocks or are restricted to small unit cells. We introduce Building-Block-Aware MOF Diffusion (BBA MOF Diffusion), an SE(3)-equivariant diffusion model that learns 3D all-atom representations of individual building blocks, encoding crystallographic topological nets explicitly. Trained on the CoRE-MOF database, BBA MOF Diffusion readily samples MOFs with unit cells containing 1000 atoms with great geometric validity, novelty, and diversity mirroring experimental databases. Its native building-block representation produces unprecedented metal nodes and organic edges, expanding accessible chemical space by orders of magnitude. One high-scoring [Zn(1,4-TDC)(EtOH)2] MOF predicted by the model was synthesized, where powder X-ray diffraction, thermogravimetric analysis, and N2 sorption confirm its structural fidelity. BBA-Diff thus furnishes a practical pathway to synthesizable and high-performing MOFs.

Figures (11)

• Figure 1: Schematic for building block aware MOF diffusion models.a. MOFs in the CoRE MOF 2019Chung2019 dataset composed of one inorganic node and one organic edge are selected and disassembled into their respective building blocks to be used for diffusion model training. During the sampling (generation) phase, the generated building blocks are assembled to form a MOF. b. Learning the joint distribution of three building block components in a MOF: 1) inorganic node (red), 2) organic edge (green), and 3) topological net (blue). A forward independent diffusion process corrupts the joint distribution of CoRE MOF dataset at $t=T$ to an independent normal distribution at $t=0$, during which an object-aware SE(3) GNN is trained with the denoising objective function. In the backwards direction, this scoring network is applied to denoise from samples in the normal distribution to the original joint distribution. In inorganic nodes and organic edges, atoms are colored as follows: Zn in purple, Cu in green, C in gray, N in blue, O in red, S in yellow, and H in white. Gray for inorganic node and red for organic edge for coloring the topological nets.
• Figure 1: Distribution for sources of invalidity in generated edges. A structure is characterized as insane if it has nonphysical connectivity, such as floating atoms.
• Figure 2: Statistics for the generated MOFs.a. Validity (orange), novelty (green) and combined (blue) for the three building blocks (edge, node, and net) independently and the final MOF assembled. b. Source of invalidity (left) and novelty (right) broken down to the contribution of node, edge, or both. c. Distribution for the number of atoms in the unit cell of generated MOFs. Both all scatters and a notched box plot are shown.The quarter 1 and 3 are the edges of the box, and fences corresponding to the edges +/- 1.5 times the interquartile range. d. Cumulative probability for the edge angle (top) and node RMSD after net optimizations (bottom) for the CoRE MOF dataset (green) and generated MOFs (blue). A box plot for the same distribution is shown at the top margin, correspondingly. e. Conditional MOF generation with dicopper tetracarboxylate “paddle-wheel” cluster as the known inorganic node. Desired chemical compositions for edge design can be proposed, with examples as C14O8NS (top), C13ONF (middle), and C13ON3F (bottom). With edge composition as inputs, BBA MOF diffusion conditionally samples 3D structures for organic edges, which are assembled together with the known inorganic node and topological net (here, nbo) to the final MOF structures.
• Figure 2: Distribution for sources of invalidity in generated nodes. A structure is characterized as insane if it has nonphysical connectivity, such as floating atoms.
• Figure 3: Overlaid revised autocorrelation (RAC) features of MOFs on RAC subsets. Opacity is utilized to visualize data overlap, and z-order is selected for visualization. a. Metal-centered RAC features for BBA Diffusion generated MOFs (red), CoRE MOF 2019 (green), and ToBaCCo MOF (blue). b. Linker-centered RAC features. c. Non-C functional-group-centered RAC features. d. Full-linker RAC features.