How generative AI can help scientists synthesize complex objects | MIT News

The resulting artificial intelligence models have been used to build large libraries of theoretical objects that can help solve all kinds of problems. Now, scientists have to figure out how to do it.

In many cases, putting things together is not as simple as following a recipe in the kitchen. Factors such as temperature and length of processing can bring about significant changes in physical properties that make or break its performance. That has limited the ability of researchers to test millions of promising models.

Now, MIT researchers have created an AI model that guides scientists through the manufacturing process by suggesting promising ways to combine. In a new paper, they show the model brings state-of-the-art accuracy in predicting the effective aggregation modes of a class of materials called zeolites, which can be used to improve exchange, absorption, and ion exchange processes. Following their suggestions, the team synthesized a new zeolite material that exhibits thermal stability.

The researchers believe their new model can break a major bottleneck in the discovery process.

“To use an analogy, we know what kind of cake we want to make, but right now we don’t know how to bake the cake,” said lead author Elton Pan, a PhD candidate in MIT’s Department of Science and Engineering (DMSE). “Materials synthesis is currently done by domain experts and trial and error.”

The job description paper is out today on Nature Computational Science. Joining Pan on the paper are Soonhyoung Kwon ’20, PhD ’24; DMSE postdoc Sulin Liu; chemical engineering PhD student Mingrou Xie; DMSE postdoc Alexander J. Hoffman; Research Assistant Yifei Duan SM ’25; DMSE visiting student Thorben Prein; Killian Sheriff of DMSE PhD; MIT Robert T. Haslam Professor in Chemical Engineering Yuriy Roman-Leshkov; Professor Manuel Moliner University of Valencia Polytechnic; MIT Paul M. Cook Career Development Professor Rafael Gómez-Bombarelli; and MIT Jerry McAfee Professor of Engineering Elsa Olivetti.

Learning to bake

Massive investment in productive AI has led companies like Google and Meta to create huge databases filled with important recipes that, at least in theory, have properties like high thermal stability and special absorption of gases. But making those materials can require weeks or months of careful testing that examines specific reaction temperatures, times, precursor ratios, and other factors.

“People rely on their chemical knowledge to guide the process,” Pan said. “Humans are linear. If there are five of them, we might keep four unchanged and change one of them linearly. But machines are much better at thinking in a high-level environment.”

The compounding process of object discovery now often takes a lot of time on the object’s journey from hypothesis to implementation.

To help scientists navigate that process, MIT researchers trained a generative AI model on more than 23,000 combinations of ingredients described in 50 years of scientific papers. The researchers repeatedly added random “noise” to the recipes during training, and the model learned to subtract noise and sample from the random noise to find promising combinations.

The result is DiffSyn, which uses a technique in AI known as diffusion.

“Diffusion models are actually generative AI models like ChatGPT, but similar to the DALL-E image generation model,” Pan said. “In the process of thinking, it transforms the sound into an audible structure by removing a small amount of noise at each step. In this case, the ‘structure’ is the assembly line of the desired object.”

When a scientist using DiffSyn enters a desired material structure, the model provides promising combinations of reaction temperatures, reaction times, precursor ratios, and more.

“It tells you how to bake your cake,” Pan said. “You have a cake in mind, you feed it to the model, the model spits out fusion recipes. The scientist can choose any synthesis method he wants, and there are easy ways to measure the most promising synthesis method from what we offer, which we show in our paper.”

To test their system, the researchers used DiffSyn to propose novel synthesis methods for zeolite, a class of materials that is complex and time-consuming to create for the test object.

“Zeolites have a very high synthesis potential,” Pan said. “Zeolites also tend to take days or weeks to crystallize, hence the impact [of finding the best synthesis pathway faster] it is much higher than other luminaries in terms of hours.”

The researchers were able to create new zeolite materials using the synthesis methods proposed by DiffSyn. Subsequent tests revealed that the material had promising properties for catalytic applications.

“Scientists have been experimenting with different recipes one by one,” Pan said. “That makes them very time-consuming. This model can sample 1,000 of them in less than a minute. It gives you very good predictions for recipes for completely new combinations.”

Difficulty accounting

In the past, researchers have built machine learning models that draw something from a single recipe. Those methods ignore that there are different ways of doing the same things.

DiffSyn is trained to map material properties to many different assembly methods. Pan says that fits better with experimental reality.

“This is a paradigm shift from one-to-one mapping between architecture and integration to one-to-many mapping,” Pan said. “That’s a big reason why we’ve had strong gains in benchmarks.”

Going forward, the researchers believe that this method should be used to train other models that guide the synthesis of materials other than zeolites, including metal-organic frameworks, inorganic solids, and other materials that have more than one possible formation method.

“This method can be extended to other things,” said Pan. “Now, the bottleneck is obtaining high-quality data for different classes of materials. But zeolites are complex, so I can imagine that they are close to high complexity. Ultimately, the goal will be to combine these intelligent systems with independent real-world experiments, and agent reasoning about experimental feedback to significantly accelerate the process of designing materials.”

The work was supported by the MIT International Science and Technology Initiatives (MISTI), the National Science Foundation, Generalitat Vaslenciana, the Office of Naval Research, ExxonMobil, and the Agency for Science, Technology and Research in Singapore.