Building-Block Aware Generative Modeling for 3D Crystals of Metal Organic Frameworks
Chenru Duan, Aditya Nandy, Sizhan Liu, Yuanqi Du, Liu He, Yi Qu, Haojun Jia, Jin-Hu Dou · 2025
Read Building-Block Aware Generative Modeling for 3D Crystals of Metal Organic Frameworks by Chenru Duan, Aditya Nandy, Sizhan Liu, Yuanqi Du, Liu He, Yi Qu, Haojun Jia, Jin-Hu Dou (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-08531.txt.
Introduction
Chenru Duan and seven co-authors present a 2025 generative model that designs metal-organic frameworks, porous crystals built from metal nodes joined by organic links. The design space is far too large to search by making each candidate in a lab. Older models either reused known building blocks or handled only small crystal cells. The new model learns the three-dimensional shapes of the blocks themselves and builds crystals block by block.
The model produced crystals with cells of up to 1,000 atoms that looked valid, new, and varied. It invented metal nodes and links not seen in its training data. One high-scoring design was then made in a real lab. Tests confirmed that the made crystal matched the design, showing the tool can guide real synthesis.
Catalog record
| CREATOR | Chenru Duan, Aditya Nandy, Sizhan Liu, Yuanqi Du, Liu He, Yi Qu, Haojun Jia, Jin-Hu Dou |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-08531
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Building-Block Aware Generative Modeling for 3D Crystals of Metal Organic Frameworks Chenru Duan, Aditya Nandy, Sizhan Liu, Yuanqi Du, Liu He, Yi Qu, Haojun Jia, Jin-Hu Dou 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 ] \mathrm{[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. 1. Introduction Metal-organic frameworks (MOFs) are a type of functional inorganic-organic hybrid material. These materials connect inorganic metal centers (metal ions or metal-oxyhydroxy clusters) through coordination bonds to form an infinite, periodic extension of the network structure. Due to their unique, modular structure and performance advantages, they have received a great deal of attention in recent years. [ 1 , 2 , 3 , 4 , 5 , 6 , 7 ] To begin, a highly ordered and programmable aperture network is constructed due to its special coordination mode through metal nodes and organic ligands. Its pore size can be precisely [ 8 ] regulated by changing the ligand length and the size of the metal cluster. [ 9 , 10 , 11 ] Secondarily, MOFs exhibit remarkably high specific surface areas, significantly exceeding those of conventional inorganic materials (e.g. zeolites) due to their fully penetrating three-dimensional channels and the efficient use of space by the hybrid inorganic-organic frame. [ 12 , 13 ] More importantly, the dynamic tunability arising from the metal-ligand coordination endows MOFs with exceptional multifunctionality. This characteristic enables nearly unlimited modular combination strategies through rational selection of building blocks. Specifically, metal nodes serve dual functions as both structural anchors and catalytically active centers, while organic ligands allow programmable modification of pore characteristics through precise functional group engineering. [ 14 , 15 ] This "structure-function integration" design strategy allows MOFs to show irreplaceable advantages in the field of intelligent response materials and precise molecular recognition. Its unique and excellent internal structure determines its position as a core material for a wide range of applications in interdisciplinary fields. MOF materials have a wide range of applications in the fields of energy, [ 16 , 17 , 18 ] catalysis, [ 19 , 20 , 21 , 22 ] sensing, [ 23 , 24 , 25 ] gas storage, [ 26 , 7 ] and so on. Over more than two decades of study, experimental structures of MOFs have been collected, sanitized, and curated in the Cambridge Structural Database (CSD). [ 27 ] Only a subset of MOFs are porous; accordingly, different sets of experimentally accessed MOFs have been further sanitized and chosen for usable databases such as CoRE MOF [ 28 , 29 ] and CSD MOF. [ 30 ] Due to their modular, building-block nature for construction, MOFs can also be assembled from their constituent parts. For this reason, many works over the past decade have constructed hypothetical databases of MOFs with varying degrees of diversity and detail. [ 31 , 32 , 33 , 34 , 35 , 36 ] These databases first began as exhaustive combinations of a small number of building blocks on a small number of topological nets [ 31 , 32 ] , but later grew in diversity with increasingly complicated nets and building blocks. [ 33 , 34 , 35 , 36 ] As MOF chemical space becomes increasingly sampled, constructing realistic MOFs [ 37 ] that are synthetically accessible remains an open challenge. This challenge is further compounded by the fact that synthetically accessing a MOF is insufficient for evaluating its properties. Prior to its use as a functional material, solvent must be removed from the pores of the MOF. Additionally, the MOF must be thermally stable, stable to conditions with water, and mechanically stable to enable practical use. With the rapidly expanding availability of experimental and synthetic data for MOFs, machine learning (ML) has emerged as a powerful, cost-effective, and efficient alternative to density functional theory (DFT) for designing functional MOFs. [ 38 , 39 , 40 , 41 , 42 ] Supervised ML models trained on high-throughput computational data have achieved remarkable success in predicting key properties such as adsorption capacities [ 43 ] , band gaps [ 44 ] , conductivity [ 45 ] , and stability metrics such as mechanical stability. [ 46 , 34 ] Additionally, ML models trained on mined experimental data have been able to predict activation stability [ 47 , 48 ] , thermal stability [ 47 , 48 ] , and most recently, water stability. [ 49 ] Transformer-based models pretrained on extensive MOF datasets have further advanced the state-of-the-art in property prediction, significantly improving generalizability and data efficiency, particularly when large-scale datasets are available [ 50 , 51 , 52 ] . Concurrently, generative models such as variational autoencoders [ 53 ] , generative adversarial networks [ 54 ] , and large language models [ 55 , 56 , 57 , 58 ] have enabled the de novo design of MOFs tailored for specific functions. More recently, diffusion models [ 59 ] , which have significantly enhanced the quality of generated images [ 60 ] , molecules [ 61 ] , and chemical reactions [ 62 , 63 ] , have emerged as promising tools for directly sampling and generating three-dimensional MOF structures. For example, MOFDiff [ 64 ] employs a diffusion model to generate coarse-grained MOF structures—comprising the identities and central positions of building blocks—that are subsequently assembled using an algorithmic approach. This assembly step was later transformed into a learnable flow-matching process in MOFFlow [ 65 ] . However, both of these approaches primarily assemble existing building blocks without generating novel 3D structures for the building blocks themselves, which often constitutes the main mode of innovation from experimental chemists. Furthermore, the generation of materials using all-atom diffusion models typically faces size constraints of around 200 atoms, owing to computational limitations and GPU memory demands [ 66 ] . Such a constraint has limited the all-atom generation of MOF unit cells, which frequently contain hundreds or thousands of atoms. In a concurrent work [ 67 ] , Inizan et al. developed the first agentic AI systems for end-to-end MOF generation and experimental validation, showing great promise for automated synthesizable MOF discovery. We address these challenges by introducing building-block aware (BBA) MOF Diffusion, a diffusion model framework that leverages a BBA representation of MOFs combined with SE(3) equivariance [ 62 , 53 ] . Our BBA MOF