Generative artificial intelligence models have been used to create vast libraries of theoretical material to help solve all kinds of problems. Now all that’s left is for scientists to figure out how to make it.
Synthesizing ingredients is often not as easy as following a recipe in the kitchen. Factors such as temperature and length of processing can cause significant changes in a material’s properties and affect its performance. This limits researchers’ ability to test the millions of promising materials generated in their models.
Now, MIT researchers have created an AI model that guides scientists through the process of making materials by suggesting promising synthetic routes. In a new paper, they showed that the model provides state-of-the-art accuracy in predicting effective synthetic routes for a class of materials called zeolites, and can be used to improve catalysis, absorption, and ion exchange processes. Following their suggestion, the team synthesized a new zeolite material with improved thermal stability.
The researchers believe their new model has the potential to break through the biggest bottleneck in the materials discovery process.
“By analogy, we know what kind of cake we want to make, but right now we don’t know how to bake it,” says lead author Elton Pan, a doctoral candidate in MIT’s Department of Materials Science and Engineering (DMSE). “Materials synthesis is currently done through disciplinary expertise and trial and error.”
A paper describing this research is published today. natural computational science. Joining Pan on this paper is Soonhyoung Kwon ’20, Ph.D. ’24. DMSE Postdoctoral Researcher Sulin Liu. Minrou Xie, a chemical engineering doctoral student. DMSE Postdoc Alexander J. Hoffman; Research Assistant Yifei Duan SM ’25; DMSE Visiting Student Thorben Prein. DMSE PhD Candidate Sheriff Killian. Yuri Roman-Leshkov, MIT Robert T. Haslam Professor of Chemical Engineering; Professor Manuel Molinar, Polytechnic University of Valencia. Rafael Gomez Bombarelli, MIT Paul M. Cook Professor of Career Development; and MIT Jerry McAfee Professor of Engineering Elsa Olivetti.
learn to make bread
With massive investments in generative AI, companies like Google and Meta have created huge databases filled with material recipes that, at least in theory, have properties such as high thermal stability and selective absorption of gases. However, producing these materials can require weeks or months of careful experimentation testing specific reaction temperatures, times, precursor ratios, and other factors.
“People rely on their chemical intuition to guide the process,” Pan says. “Humans are linear. If you have five parameters, you might keep four of them constant and one of them varies linearly. But machines are much better at reasoning in high-dimensional spaces.”
Currently, the synthetic process of materials discovery is often the most time-consuming step in taking a material from hypothesis to use.
To help scientists navigate the process, MIT researchers trained a generative AI model based on more than 23,000 material synthesis recipes described in 50 years of scientific literature. The researchers repeatedly added random “noise” to the recipe during training, and the model learned how to remove the noise and sample from the random noise to find promising synthesis routes.
The result is DiffSyn, which uses an AI approach known as diffusion.
“The diffusion model is essentially a generative AI model like ChatGPT, but similar to the DALL-E image generation model,” says Pan. “During inference, we transform noise into meaningful structure by subtracting a little bit of noise at each step. In this case, ‘structure’ becomes the synthesis root of the desired material. ”
When scientists using DiffSyn input their desired material structure, the model provides several promising combinations of reaction temperatures, reaction times, precursor ratios, and more.
“They basically teach you how to bake a cake,” Pan said. “You picture a cake in your head, feed it into the model, and the model spits out a synthesis recipe. Scientists can choose the synthesis path they want. And there’s a simple way to quantify the most promising synthesis path from what we provide, and we’ll show that in the paper.”
To test the system, the researchers used DiffSyn to propose a new synthesis route for zeolites, a class of materials that are complex and take time to form into testable materials.
“Zeolites have a very high-dimensional synthesis space,” Pan says. “Also, zeolites tend to take days to weeks to crystallize, so the impact[of finding the optimal synthetic route faster]is much greater than for other materials that crystallize in hours.”
The researchers were able to create a new zeolite material using the synthetic route proposed by DiffSyn. Subsequent tests revealed that the material had a promising morphology for catalytic applications.
“Scientists have tried different synthetic recipes one by one,” Pan says. “So it’s very time-consuming. This model can sample 1,000 samples within a minute. This gives us very good initial guesses for synthetic recipes for entirely new materials.”
Complexity considerations
Researchers have previously built machine learning models that map ingredients to a single recipe. These approaches do not take into account that there are many different ways to make the same material.
DiffSyn is trained to map material structures to different possible synthesis paths. Pan says that better matches the reality of the experiment.
“This is a paradigm shift from a one-to-one mapping to a one-to-many mapping between structure and synthesis,” says Pan. “That’s a big reason why we achieved such big gains on the benchmark.”
Going forward, the researchers believe this approach will work to train other models to guide the synthesis of materials other than zeolites, including metal-organic frameworks, inorganic solids, and other materials with multiple synthetic routes.
“This approach could potentially be extended to other materials,” Pan says. “Currently, the bottleneck is finding high-quality data for different material classes. However, given the complexity of zeolites, we can imagine that we are near the upper limit of the difficulty. Ultimately, the goal will be to interface these intelligent systems with autonomous real-world experiments, perform agentic inference based on experimental feedback, and dramatically accelerate the material design process.”
This research was supported by the MIT International Science and Technology Initiative (MISTI), the National Science Foundation, Generalitat Vaslenciana, ExxonMobil Office of Naval Research, and the Singapore Agency for Science, Technology and Research.
