Prof. Dr. Johannes T. Margraf, who leads the Chair of Physical Chemistry V at the University of Bayreuth, is at the center of a new international study examining why AI systems frequently misjudge the properties of advanced materials. His team’s work points to a long-standing issue in computational materials research: many simulation and AI models treat materials as if their atomic structures are perfectly ordered, even though real materials often contain substantial crystallographic disorder. The study highlights that this mismatch leads to meaningful inaccuracies in predictions, especially when these tools are used to identify compounds for batteries, photovoltaics or semiconductor devices.
Jakob, K. S., Walsh, A., Reuter, K., & Margraf, J. T. (2025). Learning Crystallographic Disorder: Bridging Prediction and Experiment in Materials Discovery. Advanced Materials. https://doi.org/10.1002/adma.202514226
Computer-aided materials discovery has become increasingly attractive because experimental synthesis and analysis can require significant time and resources. Researchers now rely on high-throughput simulations and machine-learning models to narrow large pools of potential compounds. These methods have accelerated progress, but they also depend heavily on idealized inputs. Crystalline materials, which include technologically important substances such as silicon and many metal oxides, are modeled with the assumption that their lattices contain no irregularities. The new work shows that this simplification overlooks common forms of disorder such as partial mixing of elements, missing atoms or local structural variations.
Prof. Dr. Johannes T. Margraf, from the University of Bayreuth stated,
“Our study shows that disorder can be a critical stumbling block in computational materials science if it is not accounted for in simulations. Fortunately, with the tools we provide, disordered materials can be detected even in large-scale workflows and addressed using the appropriate computational methods”.
The research team examined several major crystallographic databases to understand how widespread this issue may be. They reviewed materials that previous computational pipelines had flagged as promising candidates for various applications. Using a new machine-learning approach developed by the group, they evaluated whether these candidate materials would likely exhibit crystallographic disorder if synthesized. In every dataset examined, a large share of materials showed strong signatures of disorder. In one case, more than eighty percent of the predicted compounds were likely to deviate from the ideal structures used in the models that identified them.
This level of discrepancy has practical consequences. When disorder is present, properties such as conductivity, stability or catalytic activity can shift in ways that predictions fail to capture. These inaccuracies can cause researchers to pursue materials that appear promising on paper but perform differently in laboratory conditions. According to the authors, the problem is not that AI and simulations are unreliable by nature, but that the structural information they rely on needs to represent experimental reality more closely.
To address this gap, the team created a machine-learning tool that evaluates whether a given crystal structure is likely to be disordered. Instead of assuming perfect ordering, the model identifies compositions and atomic arrangements that historically show substitutional mixing or other deviations. By integrating this tool into existing computational workflows, researchers can filter out materials that are poorly represented by idealized structural models before investing time and resources in experimental testing. This creates a pathway for more efficient and more reliable materials discovery.
Margraf notes that crystallographic disorder does not need to be an obstacle if it is recognized early. When researchers know that disorder is likely, they can use appropriate computational techniques designed to handle more complex structural descriptions. This makes it possible to examine realistic materials directly rather than relying on ordered approximations. The team’s findings show that incorporating disorder awareness into computational pipelines can significantly improve the alignment between predictions and experimental outcomes.
The study underscores that AI and simulation remain powerful tools for materials science, but they must be paired with structural information that reflects how materials behave in practice. By providing a method to detect disorder at scale, the researchers aim to help the community move toward more accurate and experimentally grounded predictions. Their work suggests that the next steps in computational materials discovery will require closer integration of crystallography, data science and machine learning to better represent the complexity of real-world materials.
Adrian graduated with a Masters Degree (1st Class Honours) in Chemical Engineering from Chester University along with Harris. His master’s research aimed to develop a standardadised clean water oxygenation transfer procedure to test bubble diffusers that are currently used in the wastewater industry commercial market. He has also undergone placments in both US and China primarely focused within the R&D department and is an associate member of the Institute of Chemical Engineers (IChemE).
