Generative AI models have been used to create huge libraries of theoretical materials that could help solve all kinds of problems. Now scientists just need to figure out how to make them.
In many cases, synthesizing materials is not as simple as following a recipe in the kitchen. Factors such as temperature and processing length can cause enormous changes in the material's properties that affect or disrupt its performance. This limits researchers' ability to test millions of promising model-generated materials.
Now, MIT researchers have created an artificial intelligence model that guides scientists through the materials-making process, suggesting promising synthetic paths. In the new paper, they show that the model provides state-of-the-art accuracy in predicting efficient synthetic pathways for a class of materials called zeolites, which can be used to improve catalysis, absorption and ion exchange processes. Following his suggestions, the team synthesized a new zeolite material that showed improved thermal stability.
Scientists believe their new model could break the biggest bottleneck in the materials discovery process.
“To use an analogy, we know what kind of cake we want to bake, but right now we don't know how to bake it,” says lead author Elton Pan, a graduate student in MIT's Department of Materials Science and Engineering (DMSE). “Materials synthesis is currently based on domain expertise and trial and error.”
Article describing the work is published today in Nature Computational Science. Joining you in the article were Soonhyoung Kwon '20, PhD '24; DMSE postdoc Sulin Liu; chemical engineering graduate student Mingrou Xie; DMSE postdoc Alexander J. Hoffman; Research Assistant Yifei Duan SM '25; DMSE visiting student Thorben Prein; DMSE PhD student Killian Sheriff; MIT Robert T. Haslam Professor of Chemical Engineering Yuri Roman-Leshkov; Professor of the Polytechnic University of Valencia, Manuel Moliner; MIT Paul M. Cook Career Development Professor Rafael Gómez-Bombarelli; and MIT engineering professor Jerry McAfee Elsa Olivetti.
Learning to bake
Massive investments in generative artificial intelligence have led companies like Google and Meta to create huge databases filled with material recipes that, at least in theory, have properties such as high thermal stability and selective gas absorption. However, making these materials can require weeks or months of careful experiments that test specific reaction temperatures, timing, precursor ratios, and other factors.
“People rely on their chemical intuition to guide this process,” says Pan. “Humans are linear. If there are five parameters, we can keep four of them constant and change one of them linearly. But machines are much better at reasoning in a multidimensional space.”
The materials discovery synthesis process currently takes the longest time to take a material from hypothesis to application.
To help scientists navigate this process, MIT researchers trained a generative artificial intelligence model based on more than 23,000 materials synthesis recipes described in 50 years of scientific publications. The researchers iteratively added random “noise” to the recipes during training, and the model learned to remove noise and sample from the random noise to find promising synthesis paths.
The result is DiffSyn, which uses an approach in artificial intelligence known as diffusion.
“Diffusion models are basically a generative artificial intelligence model like ChatGPT, but more like the DALL-E image generation model,” says Pan. “During inference, it transforms noise into meaningful structure by subtracting a small amount of noise at each step. In this case, 'structure' is the route to synthesizing the desired material. “
When a scientist using DiffSyn enters the desired material structure, the model offers promising combinations of reaction temperatures, reaction times, precursor ratios, and more.
“He basically tells you how to bake a cake,” says the Lord. “You make a dough, you put it into a model, the model spits out recipes for synthesis. A scientist can choose any synthesis path they want, and there are simple ways to quantify the most promising synthesis path based on what we provide, which we show in our paper.”
To test their system, the researchers used DiffSyn to propose new pathways for synthesizing zeolite, a class of materials that is complex and takes time to transform into a testable material.
“Zeolites have a very multidimensional synthesis space,” says Pan. “Zeolites also typically crystallize over several days or weeks, so the impact (of finding the best synthesis path more quickly) is much greater than with other materials that crystallize within hours.”
Scientists managed to produce a new zeolite material using the synthesis paths proposed by DiffSyn. Subsequent research showed that the material had a promising morphology for catalytic applications.
“Scientists tried different synthesis recipes one by one,” says Pan. “This makes them very time-consuming. This model can sample 1,000 of them in less than a minute. This gives a very good initial guess at recipes for synthesizing completely new materials.”
Complexity accounting
Previously, researchers built machine learning models that mapped a material to a single recipe. These approaches do not take into account the fact that there are different ways of producing the same material.
DiffSyn is trained to map material structures into many different possible synthesis paths. You say this fits experimental reality better.
“It is a paradigm shift from a one-to-one mapping between structure and synthesis to a one-to-many mapping,” says Pan. “This is a big reason why we achieved significant gains in benchmarks.”
In the future, the researchers believe this approach should allow training of other models that guide the synthesis of materials other than zeolites, including organometallic frameworks, inorganic solids, and other materials that have more than one possible synthesis pathway.
“This approach can be extended to other materials,” says Pan. “The bottleneck now is finding high-quality data for different classes of materials. However, zeolites are complicated, so I can imagine that they are approaching the upper limit of difficulty. Ultimately, the goal would be to combine these intelligent systems with autonomous real-world experiments and agentic inference based on experimental feedback to dramatically accelerate the materials design process.”
The work was supported by MIT International Science and Technology Initiatives (MITI), the National Science Foundation, Generalitat Vaslenciana, the Office of Naval Research, ExxonMobil, and the Singapore Agency for Science, Technology and Research.
