Chemists have created a generative AI model that simplifies the process of identifying the structures of powdered crystalline materials. This model could assist researchers in characterizing materials suitable for various applications, including batteries and magnets.
For over a century, X-ray crystallography has been utilized by scientists to ascertain the structures of crystalline materials like metals, rocks, and ceramics.
This technique is most effective when working with intact crystals; however, many times, scientists only have access to powdered versions of these materials, which feature random crystal fragments. This randomness complicates the ability to reconstruct the overall structure.
Researchers at MIT have now introduced a new generative AI model designed to facilitate the analysis of powdered crystals. This predictive model stands to benefit researchers aiming to characterize materials for various uses, including in batteries and magnets.
“Knowing the structure is fundamental for any material. It is crucial for applications related to superconductivity, magnets, and photovoltaic materials. It plays a vital role in any materials-centered application,” states Danna Freedman, the Frederick George Keyes Professor of Chemistry at MIT.
Freedman and Jure Leskovec, a professor of computer science at Stanford University, are the senior authors on the study reported in the Journal of the American Chemical Society. The leading authors include MIT graduate student Eric Riesel and Yale undergraduate Tsach Mackey.
Unique Patterns
Crystalline materials encompass metals and other inorganic solids, constructed from lattices composed of identical, repeating units. These units can be seen as “boxes” of unique shapes and sizes, with precisely arranged atoms inside.
When X-rays hit these lattices, they diffract at various angles and intensities due to their interaction with different atoms, delivering insights into the positioning of atoms and their interconnections. Since the early 1900s, this method has been employed to study various materials, including biological molecules with crystalline structures like DNA and some proteins.
However, for materials present only in powdered form, determining their structures becomes significantly more challenging due to the random fragments, which do not represent the full three-dimensional structure of the original crystal.
“The original lattice is still present; what we refer to as powder is a collection of microcrystals. So even though they’re in random orientations, they maintain the same lattice structure as a large crystal,” Freedman explains.
Thousands of materials have available X-ray diffraction patterns that remain unsolved. To decode these structures, Freedman and her team trained a machine-learning model using data from the Materials Project database, which includes over 150,000 materials. They initially input tens of thousands of materials into an existing model that simulates the expected X-ray diffraction patterns. They then used these patterns to further train their AI model, named Crystalyze, to predict structures based on the X-ray patterns.
The model divides the structure prediction task into several parts. First, it ascertains the size and shape of the lattice “box” and selects which atoms fit into it. Next, it predicts how the atoms are arranged within the box. For each set of diffraction patterns, the model generates multiple potential structures that can be evaluated by inputting them into another model that predicts diffraction patterns for those structures.
“Our generative AI model creates output that it has not encountered before, allowing us to generate numerous different hypotheses,” Riesel mentions. “We can produce a hundred guesses and then compare them to the actual powder pattern. If the predicted pattern matches the input perfectly, we recognize that we achieved the correct structure.”
Unraveling Unknown Structures
The research team tested the model on various simulated diffraction patterns from their database. They also evaluated it against over 100 experimental diffraction patterns from the RRUFF database, which houses powdered X-ray diffraction data for nearly 14,000 natural crystalline minerals that were excluded from the training set. In these trials, the model was correct roughly 67 percent of the time. They subsequently tested the model on diffraction patterns that had not been previously solved, with data sourced from the Powder Diffraction File, which includes information for over 400,000 materials, both solved and unsolved.
Using their model, the team identified structures for more than 100 previously unsolved patterns. They also applied their model to discover structures in three new materials created in Freedman’s lab by synthesizing elements under high pressure, which typically do not react at atmospheric conditions. This method can lead to new materials exhibiting dramatically different crystal structures and physical properties while maintaining the same chemical composition.
For instance, graphite and diamond, both composed of pure carbon, illustrate this principle. The newly developed materials in Freedman’s lab, each containing bismuth and another element, could pave the way for designing innovative materials for permanent magnets.
“We have uncovered numerous new materials from existing data, and most notably, solved three structures unknown until now that represent the first new binary phases for those elements,” Freedman states.
The ability to identify the structures of powdered crystalline materials holds potential benefits for researchers in nearly any field related to materials, as indicated by the MIT team, who have made a web interface for the model available at crystalyze.org.
This research received funding from the U.S. Department of Energy and the National Science Foundation.
