System of Agentic AI for the Discovery of Metal-Organic Frameworks
Théo Jaffrelot Inizan, Sherry Yang, Aaron Kaplan, Yen-hsu Lin, Jian Yin, Saber Mirzaei, Mona Abdelgaid, Ali H. Alawadhi, KwangHwan Cho, Zhiling Zheng, Ekin Dogus Cubuk, Christian Borgs, Jennifer T. Chayes, Kristin A. Persson, Omar M. Yaghi · 2025
Read System of Agentic AI for the Discovery of Metal-Organic Frameworks by Théo Jaffrelot Inizan, Sherry Yang, Aaron Kaplan, Yen-hsu Lin, Jian Yin, Saber Mirzaei, Mona Abdelgaid, Ali H. Alawadhi, KwangHwan Cho, Zhiling Zheng, Ekin Dogus Cubuk, Christian Borgs, Jennifer T. Chayes, Kristin A. Persson, Omar M. Yaghi (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-2504-14110.txt.
Introduction
Théo Jaffrelot Inizan leads a large team presenting MOFGen, a crew of cooperating AI agents that dreams up new metal-organic frameworks and checks whether chemists could actually make them. Their paper appeared in two thousand twenty-five. These spongy crystals can trap carbon dioxide or harvest water from air. But the space of possible frameworks is vast, and many designs are unmakeable.
One language-model agent proposes fresh compositions. A diffusion model then builds the crystal structures in three dimensions. Quantum agents refine and filter the candidates. Feasibility agents apply expert rules and learned models to judge makability. Trained on known frameworks, the system generated hundreds of thousands of candidates. Five AI-dreamt frameworks were then synthesized in the lab. The authors call it a step toward automated discovery of makeable materials.
Catalog record
| CREATOR | Théo Jaffrelot Inizan, Sherry Yang, Aaron Kaplan, Yen-hsu Lin, Jian Yin, Saber Mirzaei, Mona Abdelgaid, Ali H. Alawadhi, KwangHwan Cho, Zhiling Zheng, Ekin Dogus Cubuk, Christian Borgs, Jennifer T. Chayes, Kristin A. Persson, Omar M. Yaghi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-14110
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
System of Agentic AI for the Discovery of Metal-Organic Frameworks Théo Jaffrelot Inizan, Sherry Yang, Aaron Kaplan, Yen-hsu Lin, Jian Yin, Saber Mirzaei, Mona Abdelgaid, Ali H. Alawadhi, KwangHwan Cho, Zhiling Zheng, Ekin Dogus Cubuk, Christian Borgs, Jennifer T. Chayes, Kristin A. Persson, Omar M. Yaghi Abstract Generative models and machine learning promise accelerated material discovery in MOFs for CO 2 capture and water harvesting but face significant challenges navigating vast chemical spaces while ensuring synthetizability. Here, we present MOFGen, a system of Agentic AI comprising interconnected agents: a large language model that proposes novel MOF compositions, a diffusion model that generates crystal structures, quantum mechanical agents that optimize and filter candidates, and synthetic-feasibility agents guided by expert rules and machine learning. Trained on all experimentally reported MOFs and computational databases, MOFGen generated hundreds of thousands of novel MOF structures and synthesizable organic linkers. Our methodology was validated through high-throughput experiments and the successful synthesis of five“AI-dreamt” MOFs, representing a major step toward automated synthesizable material discovery. Accelerating the discovery of novel, synthesizable materials is crucial for addressing major global challenges, including carbon capture [ 1 , 2 , 3 ] , water harvesting [ 4 , 5 , 6 , 7 ] . Historically, materials discovery has heavily relied on trial-and-error approaches, experimental heuristics, incremental improvements, chemical intuition, and computational simulations, notably with density functional theory (DFT) [ 8 , 9 ] . Significant advances in the development of open-access materials databases [ 10 , 11 , 12 , 13 , 14 , 15 ] , automated laboratory techniques [ 16 , 17 , 18 ] , machine-learning-based property predictors, generative models [ 19 , 20 , 21 ] and machine learning force fields (MLFF) [ 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 ] have collectively enabled exploration of chemical spaces and generated large databases of millions of computationally stable candidates [ 31 , 12 ] . However, synthesizability remains largely unexplored and presents a major limitation in translating predictions to reality. Indeed, a material that appears stable from a quantum mechanical perspective can be extremely difficult or too costly to synthesize. Metal–organic frameworks (MOFs) [ 32 , 33 , 6 ] , nanoporous materials formed from metal nodes linked by strong bonds to organic molecules, are an example of materials synthesized in crystalline form. Their synthesis typically requires extensive screening of experimental conditions (e.g., metal and linker concentrations, temperature) and involves multi-step organic linker syntheses, rendering the process labor-intensive and limiting exploration of the chemical space. To overcome these challenges, integrating generative models with quantum mechanical predictions, and rigorous experimental constraints specific to MOFs can streamline candidate selection and reduce reliance on conventional heuristics [ 34 , 21 , 35 ] . Recent breakthroughs in generative models, exemplified by AlphaFold in protein structure prediction [ 36 , 37 ] and diffusion-based models in inorganic crystal generation [ 20 , 38 , 19 ] , highlight the potential of these models to discover structures beyond human intuition. However, even if these models can generate a large number of DFT-stable materials, translating these advances to MOFs remains a challenge. For example, conventional stability metrics [ 39 ] are not always applicable to MOFs. In addition, MOFs typically involve large complex structures, chemical composition diversity, and a wide range of topologies, not trivial to capture by machine learning models. As a result, only a handful of inorganic material discovered by generative models has been experimentally verified to date [ 19 ] , emphasizing the significant gap between computational prediction and real synthesis. This report outlines the design of a system of Agentic AI, MOFGen, combining large language models (LLMs) [ 40 , 41 ] , diffusion algorithms [ 42 , 43 , 38 ] , MLFFs, quantum mechanical computations [ 44 , 45 ] , machine learning-based synthesis predictors [ 46 , 47 ] and high-throughput synthesis. Trained across all reported MOF chemistry, our interconnected system of specialized agents enabled rapid exploration and discovery of synthesizable MOFs, significantly shortening the path from computational predictions to realization in the laboratory. Within just a few weeks, we have synthesized five “AI-dreamt” MOFs, with some completed in just few days. By leveraging generative models and AI-driven methodologies, our system accelerates the creation of “AI-dreamt” synthesizable materials, paving the way towards human-in-the-loop foundation model for MOF discovery. Results Overview of the system of Agentic AI, MOFGen Our approach, MOFGen, consists of a modular system of AI and computational agents, each defined as an entity capable of taking actions based on its environment to accelerate MOF discovery. The architecture of our system, shown in Fig. 1 , is composed as follows. First, the main LLM agent, MOFMaster , provides three generation options: (i) generate a specified number of MOF structures without chemical information, (ii) generate MOF structures containing a specific metal secondary building unit (SBU), and (iii) generate MOF structures with a given chemical composition. Second, these instructions are then passed to LinkerGen , an LLM chemical-composition generator leveraging in-context learning and chain-of-thought [ 48 ] on experimental MOF organic linkers, and which produces MOF chemical formulae based on the constraints provided by MOFMaster . Third, the generated formulae are served as inputs to CrystalGen which is a chemical composition–conditioned, all-atom, denoising-diffusion probabilistic model for crystal structure generation, trained on both experimental and computational MOF crystal structures. Fourth, the diffusion-generated structures are subjected to a series of geometry optimization and filtering steps through a low-level quantum mechanical agent, QForge . Fifth, SynthABLE decomposes each MOF into its building blocks and evaluates their synthesizability using multi-level fidelity synthetic rules and machine learning synthesizability predictors. Sixth, the remaining candidates are further refined by a high-level quantum mechanical agent, QHarden , which performs successive geometry optimizations ranging from PBE-D4 to the higher-level r 2 SCAN-D4 levels. Seventh, SynthGen ranks the most likely-to-be-synthesized MOF candidates, which are subsequently validated through our high-throughput synthesis platform, powered by LLM–guided synthesis planning and characterized using single crystal or powder X-ray diffraction (SCXRD or PXRD). Chemical Formula Generation with LinkerGen A dedicated LLM agent, LinkerGen , is employed to generate organic linker chemical formulae. The agent is prompted with chemical representations extracted from the experimental databases through an in-context learning strategy using chemical constraints with desired metal SBU formula. We tested several strategies: (i) chemical formula \rightarrow SMILES, and (ii) SMILES \rightarrow SMILES. Our analysis shows that the SMILES-to-SMILES strategy exhibits the highest success rate of generating valid linkers. In addition, the synthesizability scores (SA score and SCScore) show (see the supplementary materials, section 0.2.1 ) a shift in the distribution with a double peak matching the experimental data. Once generated the organic linker formulae are then concatenated with metal SBU formulae, using the correct stoichiometric ratios. To enhance structural complexity, supercell formulations are also generated by scaling the number of total atoms, linkers and metal SBUs, with a maximum of 256 atoms, atom limit that CrystalGen can target. The resulting database of chemical formulae serves as the input for downstream diffusion-based structure generation for CrystalGen . Crystal Structure Generation with CrystalGen Due to their large size and complex three-dimensional frameworks, MOFs present a