Generative artificial intelligence models can create vast libraries of theoretical materials that could help solve a wide range of problems. However, scientists now face the challenge of actually producing these materials.
In many cases, material synthesis is not as simple as following a recipe in the kitchen. Factors such as temperature and processing time can cause enormous changes in a material's properties, which determine its performance. Until now, this has limited researchers' ability to test millions of promising materials generated by models.
MIT researchers have now created an AI model that guides scientists through the material production process by suggesting promising synthesis pathways. In a new study, they demonstrated that the model achieves state-of-the-art accuracy in predicting effective synthesis pathways for a class of materials called zeolites, which can be used to improve catalysis, absorption, and ion-exchange processes. Following its suggestions, the team synthesized a new zeolite material that exhibited improved thermal stability.
Learning to Bake a Material Cake
Elton Pan, a doctoral student in MIT's Department of Materials Science and Engineering and lead author of the study, explains the problem using an analogy: "We know what cake we want to bake, but right now we don't know how to bake the cake. Material synthesis is currently carried out through expertise and trial and error." The study describing this work was published in the journal Nature Computational Science.
Massive investments in generative AI have led companies such as Google and Meta to create enormous databases filled with recipes for materials that theoretically possess properties such as high thermal stability and selective gas absorption. However, producing these materials can require weeks or months of careful experiments testing specific reaction temperatures, times, precursor ratios, and other factors.
"People rely on their chemical intuition to guide the process," Pan says. "People are linear. If there are five parameters, we can keep four of them constant and vary one of them linearly. But machines are much better at reasoning in high-dimensional space." The synthesis process in materials discovery now often takes the most time on a material's journey from hypothesis to application.
The DiffSyn AI Model and How It Works
To help scientists manage this process, MIT researchers trained a generative AI model on more than 23,000 material synthesis recipes described in scientific papers spanning 50 years. During training, the researchers iteratively added random "noise" to the recipes, and the model learned to remove the noise and sample from random noise to find promising synthesis pathways. The result is DiffSyn, which uses an AI approach known as diffusion.
"Diffusion models are essentially generative AI models like ChatGPT, but more like the DALL-E image generation model," Pan explains. "During inference, it converts noise into a meaningful structure by subtracting a little noise at each step. In this case, the 'structure' is the synthesis pathway for the desired material."
When a scientist using DiffSyn enters the desired material structure, the model offers several promising combinations of reaction temperatures, reaction times, precursor ratios, and other parameters. "Essentially, it tells you how to bake your cake," Pan says. "You have a cake in mind, you put it into the model, and the model spits out synthesis recipes. A scientist can choose any synthesis pathway they want, and there are simple ways to quantify the most promising synthesis pathway from what we provide, as we show in our paper."
Zeolite Test Proves Successful
To test their system, the researchers used DiffSyn to propose new synthesis pathways for a zeolite, a complex class of materials that takes a long time to form into a testable material. Zeolites are crystalline hydrated aluminosilicates of alkali metals and alkaline earth metals. "Zeolites have a very high-dimensional synthesis space," Pan says. "Zeolites also tend to crystallize over days or weeks, so the impact of finding the best synthesis pathway more quickly is much greater than for other materials that crystallize within hours."
The researchers were able to produce a new zeolite material using synthesis pathways proposed by DiffSyn. Subsequent testing revealed that the material had a promising morphology for catalytic applications. "Scientists have been trying different synthesis recipes one by one," Pan says. "That makes them very time-consuming. This model can sample 1,000 of them in less than a minute. It gives you a very good initial idea of synthesis recipes for entirely new materials."
Accounting for Complexity
Previously, researchers created machine-learning models that mapped a material to a single recipe. These approaches do not account for the fact that there are different ways to produce the same material. DiffSyn is trained to map material structures to many different possible synthesis pathways. Pan says this better reflects experimental reality. "This is a paradigm shift from a one-to-one mapping between structure and synthesis to a one-to-many mapping," Pan says. "That is a major reason why we achieved strong gains in benchmarks."
A Promising Future for the Model
The researchers believe this approach should also work for training other models that guide the synthesis of materials beyond zeolites, including metal-organic frameworks, inorganic solids, and other materials that have more than one possible synthesis pathway.
"This approach could be extended to other materials," Pan says. "The bottleneck now is finding high-quality data for different classes of materials. But zeolites are complicated, so I can imagine they are close to the upper limit of difficulty. Ultimately, the goal would be to connect these intelligent systems with autonomous real-world experiments and agentic reasoning about experimental feedback to dramatically accelerate the materials design process."
Source: news.mit.edu
