For over a century, scientists have leveraged the power of X-ray crystallography to decipher the intricate structures of crystalline materials, including metals, rocks, and ceramics. This sophisticated method thrives on analyzing pristine crystals, but researchers often grapple with powdered forms that consist of random crystal fragments, complicating their efforts to reconstruct overall structures.
Now, a team of chemists at MIT has unveiled an innovative generative AI model designed to simplify the determination of powdered crystal structures. This predictive model offers researchers valuable assistance in characterizing materials that hold promise for applications in batteries, magnets, and a range of other fields.
“Understanding the structure of material is fundamental,” asserts Danna Freedman, the Frederick George Keyes Professor of Chemistry at MIT. “It plays a crucial role in superconductivity, magnetism, and photovoltaic applications, impacting virtually every materials-centric field.”
Freedman, along with Jure Leskovec, a computer science professor at Stanford University, are the senior authors of a new study published in the Journal of the American Chemical Society. The leading contributors to the research include MIT graduate student Eric Riesel and undergraduate Tsach Mackey from Yale University.
Unveiling Unique Patterns
Crystalline materials, which encompass metals and various inorganic solids, are formed from lattices composed of identical, repeating units. Imagine these units as distinct “boxes” containing atoms organized with precision. When X-rays interact with these lattices, they scatter at varying angles, providing insight into atomic positioning and bonding.
Since the dawn of the 20th century, this technique has played a vital role in material analysis, including biological molecules like DNA and certain proteins. However, determining the structures of materials existing solely in powdered form poses significant challenges, as these fragments fail to convey the complete 3D structure of the original crystal.
“Although the primary lattice remains intact, a powder consists of microcrystals that are randomly oriented,” Freedman explains. This randomness is one reason thousands of materials still have unexplained diffraction patterns despite existing data.
To tackle this issue, Freedman and her team developed a machine-learning model trained on data sourced from the Materials Project, which contains information on over 150,000 materials. The process began by feeding tens of thousands of materials into an existing model that can simulate the X-ray diffraction patterns. Subsequently, they trained their AI model, dubbed Crystalyze, to predict structures based on these patterns.
The Crystalyze model systematically breaks the structure prediction process into manageable tasks. It first identifies the size and shape of the atomic “box” and the specific atoms to be included. Next, it predicts how these atoms are arranged within the box. For each X-ray diffraction pattern, the model generates multiple potential structural designs, which can be tested against a model simulating diffraction patterns.
“Our generative AI model creates predictions based on unseen data, allowing us to formulate numerous possibilities,” Riesel notes. “If the predicted powder pattern matches the actual input pattern, we can confidently say we’ve identified the correct structure.”
Unraveling Unknown Crystalline Structures
The researchers tested their model using thousands of simulated diffraction patterns from the Materials Project and over 100 experimental patterns from the RRUFF database, which hosts X-ray diffraction data for nearly 14,000 natural crystalline minerals. The model demonstrated accuracy in approximately 67% of cases. Subsequently, they applied the model to diffraction patterns that had not yet been deciphered, sourced from the Powder Diffraction File, which encompasses data for over 400,000 materials, both solved and unsolved.
Using their AI model, the team successfully proposed structures for over 100 previously unresolved patterns. Moreover, they identified structures for three novel materials synthesized in Freedman’s lab by compelling elements that typically do not react at regular atmospheric pressure to form compounds under extreme pressure conditions. This innovative method can lead to the creation of new materials with distinct crystal structures and physical properties, even with identical chemical compositions; graphite and diamond are classic examples of such substances.
The newly discovered materials, which encompass bismuth along with another element, may significantly contribute to the advancement of new permanent magnet designs.
“Our research unlocked numerous new material possibilities from existing data and accomplished the first identification of three unknown structures from our laboratory,” Freedman states.
Successfully determining the structures of powdered crystalline materials holds great potential across various fields, according to the MIT team, and they have launched a web interface for the Crystalyze model available at crystalyze.org.
This research was made possible through funding from the U.S. Department of Energy and the National Science Foundation.
Photo credit & article inspired by: Massachusetts Institute of Technology
