Each year, researchers around the world create thousands of new materials — but many of them never reach their full potential. A new AI tool from U of T Engineering could change that by predicting how a new material could best be used, right from the moment it’s made.
In a study published in Nature Communications, a team led by Professor Seyed Mohamad Moosavi (ChemE) introduces a multimodal AI tool that can predict how well a new material might perform in the real world.
The system focuses on a class of porous materials known as metal-organic frameworks (MOFs). Moosavi says that last year alone, materials scientists created more than 5,000 different types of MOFs, which have tunable properties that lead to a wide range of potential applications.
For example, MOFs can be used to separate CO2 from other gases in a waste stream, preventing the carbon from reaching the atmosphere and contributing to climate change. They can also be used to deliver drugs to particular areas of the body, or to add new functions to advanced electronic devices.
According to Moosavi, one major challenge facing the field is that a MOF created for one purpose often turns out to have the ideal properties for a completely different application.
For example, in one of their previous studies, it was found that a material originally synthesized for photocatalysis was instead very effective for carbon capture — but this discovery was only made seven years later.
“In materials discovery, the typical question is, ‘What is the best material for this application?’” says Moosavi.
“We flipped the question and asked, ‘What’s the best application for this new material?’ With so many materials made every day, we want to shift the focus from ‘what material do we make next’ to ‘what evaluation should we do next.’”
This approach aims to reduce the time lag between discovery and deployment of MOFs.
To help make this possible, ChemE PhD student Sartaaj Khan developed a multimodal machine learning system trained on various types of data typically available immediately after synthesis — specifically, the precursor chemicals used to make the material, and its powder X-ray diffraction (PXRD) pattern.
“Multimodality matters,” says Khan. “Just as humans use different senses — such as vision and language — to understand the world, combining different types of material data gives our model a more complete picture.”
The AI system uses a multimodal pretraining strategy to gain insights into a material’s geometry and chemical environment, enabling it to make accurate property predictions without needing post-synthesis structural characterization.
This can speed up the discovery process and help researchers recognize promising materials before they’re overlooked or shelved.
To test the model, the team conducted a ‘time-travel’ experiment. They trained the AI on material data available before 2017 and asked it to evaluate materials synthesized after that date.
The system successfully flagged several materials — originally developed for other purposes — as strong candidates for carbon capture. Some of those are now undergoing experimental validation in collaboration with the National Research Council of Canada.
Looking ahead, Moosavi plans to integrate the AI into the self-driving laboratories (SDLs) at U of T’s Acceleration Consortium, a global hub for automated materials discovery.
“SDLs automate the process of designing, synthesizing and testing new materials,” he says.
“When one lab creates a new material, our system could evaluate it — and potentially reroute it to another lab better equipped to assess its full potential. That kind of seamless inter-lab coordination could accelerate materials discovery.”
