“`html
The machine learning frameworks that convert text into visuals are equally valuable for producing innovative substances. In recent years, generative substance models from corporations such as Google, Microsoft, and Meta have utilized their training datasets to assist researchers in developing millions of novel materials.
However, when it comes to creating substances with rare quantum characteristics like superconductivity or distinct magnetic states, those frameworks face challenges. This is unfortunate, as human assistance is greatly needed. For instance, after a decade of investigation into a category of materials that could transform quantum computing, termed quantum spin liquids, only about twelve material options have been recognized. This limitation results in fewer substances to serve as the foundation for significant technological advancements.
Now, researchers at MIT have established a method that enables prominent generative substance models to produce promising quantum materials by adhering to specific design principles. The principles, or limitations, guide models to construct materials with distinctive structures that yield quantum properties.
“The frameworks from these major companies produce materials optimized for durability,” remarks Mingda Li, MIT’s Class of 1947 Career Development Professor. “Our viewpoint is that this is not typically how materials science progresses. We don’t require 10 million new materials to transform the world. We only need one exceptional material.”
The method is detailed today in a publication by Nature Materials. The researchers applied their method to generate millions of candidate materials comprising geometric lattice configurations associated with quantum properties. From this collection, they synthesized two genuine materials with unique magnetic characteristics.
“Individuals in the quantum community emphasize these geometric limitations, such as the Kagome lattices formed by two overlapping, inverted triangles. We fabricated materials with Kagome lattices because those structures can replicate the behavior of rare earth elements, thus holding significant technological relevance,” Li explains.
Li serves as the senior author of the document. His MIT co-authors encompass PhD students Ryotaro Okabe, Mouyang Cheng, Abhijatmedhi Chotrattanapituk, and Denisse Cordova Carrizales; postdoctorate Manasi Mandal; undergraduate researchers Kiran Mak and Bowen Yu; visiting scholar Nguyen Tuan Hung; Xiang Fu ’22, PhD ’24; and professor of electrical engineering and computer science Tommi Jaakkola, who is affiliated with the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Institute for Data, Systems, and Society. Additional co-authors include Yao Wang from Emory University, Weiwei Xie from Michigan State University, YQ Cheng from Oak Ridge National Laboratory, and Robert Cava from Princeton University.
Guiding models towards impact
The characteristics of a material are dictated by its structure, and quantum materials are no exception. Certain atomic arrangements are more prone to exhibiting rare quantum characteristics compared to others. For example, square lattices can act as a platform for high-temperature superconductors, while other forms known as Kagome and Lieb lattices can facilitate the creation of materials relevant for quantum computing.
To assist a widely used class of generative frameworks recognized as diffusion models in producing materials that align with specific geometric designs, the researchers developed SCIGEN (an abbreviation for Structural Constraint Integration in GENerative model). SCIGEN is a software code that ensures diffusion models comply with user-specified constraints at every iterative generation step. With SCIGEN, users can impose any generative AI diffusion model with geometric structural rules to follow while generating materials.
AI diffusion models operate by sampling from their training dataset to produce structures that represent the distribution of structures within that dataset. SCIGEN prevents generations that do not correspond with the structural rules.
To evaluate SCIGEN, the researchers utilized it on a popular AI materials generation framework known as DiffCSP. They instructed the SCIGEN-augmented model to create materials with distinctive geometric patterns referred to as Archimedean lattices, which are assemblies of 2D lattice tilings consisting of various polygons. Archimedean lattices can result in a variety of quantum phenomena and have been the subject of extensive research.
“Archimedean lattices give rise to quantum spin liquids and so-called flat bands, which can replicate the properties of rare earths without involving rare earth elements, making them extremely significant,” states Cheng, a co-corresponding author of the study. “Other Archimedean lattice materials possess large pores that could facilitate carbon capture and other uses, resulting in a collection of extraordinary materials. In several instances, there are no known materials with that lattice; hence, I think it will be immensely intriguing to discover the first material that conforms to that lattice.”
The model generated more than 10 million material options with Archimedean lattices. One million of those materials passed a stability screening. Utilizing the supercomputers at Oak Ridge National Laboratory, the researchers then selected a smaller sample of 26,000 materials and conducted detailed simulations to comprehend how the underlying atoms of the materials behaved. The researchers identified magnetism in 41 percent of those structures.
From that subset, the researchers synthesized two previously unknown compounds, TiPdBi and TiPbSb, in the laboratories of Xie and Cava. Subsequent experiments demonstrated that the predictions made by the AI model largely corresponded with the actual properties of the materials.
“Our goal was to uncover new materials that could have a significant potential impact by incorporating these structures known to produce quantum properties,” asserts Okabe, the paper’s primary author. “We are already aware that these materials with certain geometric patterns are captivating, so it’s logical to begin with them.”
Expediting material advancements
Quantum spin liquids could enable quantum computing by creating stable, error-resistant qubits that serve as the basis for quantum operations. Yet, no quantum spin liquid materials have been confirmed. Xie and Cava believe SCIGEN could hasten the quest for these materials.
“There exists a substantial search for quantum computing materials and topological superconductors, all of which are connected to the geometric patterns of materials,” Xie states. “However, experimental progress has been exceedingly slow,” Cava adds. “Many of these quantum spin liquid materials must meet specific conditions: They must be in a triangular lattice or a Kagome lattice. If the materials meet those criteria, quantum researchers become enthusiastic; it’s a necessary yet insufficient condition. Thus, by generating numerous materials this way, it immediately provides experimentalists with hundreds or thousands more candidates to work with to hasten quantum computing material research.”
“This endeavor offers a new tool, utilizing machine learning, that can forecast which materials will exhibit specific elements in a desired geometric arrangement,” comments Drexel University Professor Steve May, who was not part of the research. “This should accelerate the discovery of previously unexplored materials for applications in next-generation electronic, magnetic, or optical technologies.”
The researchers emphasize that experimentation remains essential to determine whether AI-generated materials can be synthesized and how their actual characteristics compare with model predictions. Future developments on SCIGEN could integrate additional design constraints into generative models, including chemical and functional limitations.
“Individuals aiming to transform the world prioritize material properties over the stability and structure of materials,” Okabe remarks. “With our approach, the proportion of stable materials diminishes, but it opens the door for generating a plethora of promising materials.”
This work was partially funded by the U.S. Department of Energy, the National Energy Research Scientific Computing Center, the National Science Foundation, and Oak Ridge National Laboratory.
“`