Models of artificial intelligence that turn the text into images are also useful for generating new materials. Over the past few years, models of generative materials from companies such as Google, Microsoft and META have developed their training details to help researchers design tens of millions of new materials.
But when it comes to designing materials with exotic quantum properties, such as superconductivity or unique magnetic states, these models are struggling. It's a pity, because people could use help. For example, after a decade of material class research, which could revolutionize quantum calculations, called quantum spin liquids, only a dozen material candidates were identified. A bottleneck means that there are fewer materials that can serve as the basis of technological breakthroughs.
Now scientists MIT has developed a technique that allows popular models of generative materials to create promising quantum materials by observing specific design rules. Rules or restrictions direct the model of creating materials with unique constructions that cause quantum properties.
“The models of these large companies generate materials optimized for stability,” says Mingda Li, a professor of career development MIT from 1947. “Our perspective is that material learning is usually developing. We don't need 10 million new materials to change the world. We only need one really good material.”
The approach is described today in Article published by Natural materials. Scientists used their technique to generate millions of candidate materials consisting of geometric networks related to quantum properties. From this pool, they synthesized two real materials with exotic magnetic features.
“People in the quantum community really care about these geometric restrictions, such as kagome grilles, which are two overpowering triangles upside down. We have created materials with kagome networks, because these materials can imitate the behavior of rare earth elements, so they are of great technical importance.” Li says.
Li is the senior author of the newspaper. His co -authors myth are doctoral students Ryotaro Okabe, Mouyang Cheng, Abhijatmedhi Chotrattanapituk and Denisse Cordova Carrizales; Postdoc manasi mandal; Student researchers Kiran Mak and Bowen Yu; Visiting scholar Nguyen Tuan Hung; Xiang Fu '22, PhD '24; and a professor of electrical engineering and computer science Tommi Jaakkola, who is a partner of the IT and artificial intelligence laboratory (CSAIL) and the Institute of Data, Systems and Society. Additional co -authors are Yao Wang from Emory University, Weiwei Xie from Michigan State University, YQ Cheng from Oak Ridge National Laboratory and Robert Cava from Princeton University.
Models for the impact
The properties of the material are determined by its structure, and quantum materials are not different. Some atomic structures more often cause exotic quantum properties than others. For example, square networks can be used as a high temperature superconductors, while other shapes known as Kagome and Lieb grilles can support the creation of materials that can be useful for quantum calculations.
To help the popular class of generative models known as diffusion models produce materials in line with individual geometric patterns, scientists have created a scigen (an abbreviation of the integration of structural restrictions in the generative model). Scigen is a computer code that ensures that diffusion models adjacent to the limitations defined by the user at every stage of iterative generation. In the case of Scigen, users can provide any generative structural rules of the AI diffusion model, which should be observed when generating materials.
AI diffusion models work by sampling from their training data set to generate structures reflecting the distribution of structures found in the data set. Scigen blocks generations that are not in line with structural rules.
To test SCIGIG, scientists applied it to the popular model of producing AI materials known as DIFFCSP. They had a model equipped with a scigen generating materials with unique geometric patterns called Archimedean years, which are collections of 2D networks of various polygons. Archimedeńskie networks can lead to a number of quantum phenomena and were the subject of many studies.
“Archimedean networks cause the formation of quantum spin liquids and so -called flat strands that can imitate the properties of rare lands without rare lands, so they are extremely important,” says Cheng, correction of the author of the work. “Other Archimedan network materials have large pores that can be used to capture coal and other applications, so it is a set of special materials. In some cases there are no known materials with this grille, so I think that it will be really interesting to find the first material that matches this network.”
The model generated over 10 million material candidates with Archimedeans. A million of these materials survived the stability test. Using supercomputists at Oak Ridge National Laboratory, scientists then took a smaller sample of 26,000 materials and conducted detailed simulations to understand how the base atoms of the materials have survived. Scientists have found magnetism in 41 percent of these structures.
From this subset, scientists synthesized two previously undiscovered relationships, Tipdbi and TipBSB, in Xie and Cava laboratories. Subsequent experiments showed forecasts of the AI model largely even with the properties of real material.
“We wanted to discover new materials that could have a huge potential influence, including those structures that are known to cause quantum real estate,” says Okabe, the first author. “We already know that these materials with specific geometric designs are interesting, so it is natural to start with them.”
Acceleration of material breakthrations
Quantum spin liquids can unlock quantum calculations, enabling stable, resistant to cubic errors, which serve as the basis of quantum operations. But no quantitative spin materials have been confirmed. Xie and Cava believe that Scigen can accelerate the search for these materials.
“There is a large search for quantum and topological computer materials, and all are associated with the geometric patterns of materials,” says Xie. “But the experimental progress was very, very slow,” adds Cava. “Many of these quantum spin materials are subject to restrictions: they must be in a triangular network or kagome network. If the materials meet these restrictions, quantum researchers are excited; this is necessary, but insufficient condition. So, generating many, many materials, it immediately gives experimentalists hundreds or thousands of candidates for playing materials.
“This work presents a new tool using machine learning that can predict which materials will have specific elements in the desired geometric pattern,” says Drexel University professor Steve May, who was not involved in the study. “This should speed up the development of previously unexplored materials for applications in electronic, magnetic or new generation optical technologies.”
Scientists emphasize that experiments are still crucial for assessing whether the materials generated by AI can be syntated and how their real properties compare to model forecasts. Future work on SCIGIGE can take into account additional design principles in generative models, including chemical and functional restrictions.
“People who want to change the world care more about material properties than the stability and structure of materials,” says Okabe. “With our approach, the ratio of stable materials decreases, but opens the door to generate a whole bunch of promising materials.”
The works were partly supported by the US Energy Department, National Energy Scientific Computing Center, the National Science Foundation and Oak Ridge National Laboratory.
