New tool harnesses AI to navigate expanding world of metal–organic frameworks

A new open-access tool created by University of Toronto Engineering researchers provides a systematic way to organize and synthesize knowledge about metal–organic frameworks (MOFs) — a class of materials with applications in drug delivery, catalysis, carbon capture and more. 

Metal–organic frameworks (MOFs) are an exceptionally versatile class of materials, distinguished by their ultra-high surface area and precisely tunable chemistry. Some MOFs have surface areas reaching up to 7,000 m²/g, meaning that a gram of this material contains enough internal surface area to cover a football field. 

This unique structure enables a wide range of promising applications. Some can be used as molecular sieves, separating carbon dioxide from other gases so it can be captured and sequestered. Others grab onto tiny molecules, enabling them to be detected at extremely low concentrations. Still others can help speed up industrially important reactions, or deliver drugs to certain areas of the body. 

The growing importance and transformative potential of MOFs in science and technology is underlined by the fact that they were the subject of the 2025 Nobel Prize in Chemistry. 

But with studies on MOFs accelerating across more than 25 application domains, keeping track of the field’s rapidly growing body of knowledge has proven increasingly challenging — not just for researchers, but also for the AI tools intended to support scientific discovery.

A team led by Professor Mohamad Moosavi in the Department of Chemical Engineering & Applied Chemistry, and the Vector Institute, has developed a new system to help address that challenge.

Their new tool is named Unifying Chemical Data for MOFs, abbreviated to MOF-ChemUnity. The work has been published in the Journal of the American Chemical Society, one of the most prestigious journals in chemistry; the study was selected for the cover of a recent issue.

“Scientific discovery begins with reading and synthesizing the literature, but this remains one of the most difficult steps to automate,” says Moosavi.   

“MOF-ChemUnity creates a unified foundation that both researchers and AI systems can build on.” 

A structured map of MOF knowledge 

The remarkable tunability of MOFs makes them suitable for a wide range of technologies, but the breadth and diversity of research across disciplines have made the field increasingly complex to navigate. 

MOF-ChemUnity addresses this challenge using a structured and scalable knowledge graph that systematically extracts and links information from MOF research papers, crystal structure repositories and computational materials databases.

At the core of the system is a multi-agent large language model (LLM) workflow designed to connect chemical names in the literature to the correct crystal structures. This enables synthesis procedures, material properties and potential applications to be represented in a consistent, machine-readable format. 

“A knowledge graph connects pieces of information like a web, linking things, like a MOF, its metal node, synthesis protocol, and adsorption property through their relationships — ‘made from’, ‘synthesized’,  ‘used for’,” says Moosavi. 

 “This lets AI not just store data but understand and reason about how materials, properties and applications are connected — exactly what MOF-ChemUnity enables.”  

Reducing AI hallucination through literature grounding 

The team demonstrated the system’s impact by integrating the knowledge graph with large language models to build a literature-informed AI assistant for MOFs. Unlike standard AI systems, which can produce plausible-sounding but incorrect statements, the literature-informed assistant draws on verified experimental and computational records. 

In blind evaluations performed by MOF experts from multiple institutions, the assistant’s responses were consistently rated as more accurate, interpretable and trustworthy than those produced by baseline LLMs such as GPT-4o. 

“This approach reduces hallucination, which is one of the major obstacles in applying large language models to scientific domains,” Moosavi says. 

“By grounding AI responses in curated and linked literature, we can support more reliable scientific reasoning.”

A foundation for future materials discovery 

The U of T team — Moosavi and his graduate students, Thomas Pruyn and Amro Aswad, who were key contributors to the work — have made the dataset and code openly available on GitHub, aiming to support continued progress in materials science and AI-driven research.

The main funder is the National Research Council of Canada’s Materials for Clean Fuels Challenge Program, and U of T’s Acceleration Consortium and Data Science Institute.  

Moosavi says the project lays groundwork for a broader shift in how scientific knowledge is organized and accessed. 

“This work will help break down silos in scientific research,” Moosavi says.  

“Human researchers are limited by the number of papers they can read, but MOF-ChemUnity takes a first step toward enabling AI systems that can process data across fields.  

“It establishes a new paradigm for literature-informed discovery, and we envision it as the beginning of generalized knowledge systems that can accelerate research across many fields.”

---

---