New International Study Uncovers Major Limitations in AI-Driven Materials Discovery 

A new study led by the University of Bayreuth and our Institute´s Theory Department reveals that widely used computer simulations and artificial intelligence tools often make significant errors when predicting the properties of next-generation, high-performance materials. Published in the prestigious journal Advanced Materials, the research identifies the root of these inaccuracies and introduces new methods to overcome them.

Introduction to Current Prediction Models

Computational tools have become indispensable in the search for innovative materials—from faster semiconductors to more efficient solar cells. In particular, AI-powered prediction models help narrow down the number of promising candidates long before they reach a laboratory. However, these models frequently rely on idealised representations of crystal structures that do not accurately reflect how materials form in real-world conditions.

One common feature of real crystalline materials is crystallographic disorder, especially substitutional disorder, where atoms of similar elements mix within the lattice. These natural deviations are often overlooked by current AI systems, leading to inaccurate predictions about how a material will behave once synthesized.

Key Aspects of the Study

1. Identifying the Source of Prediction Errors
The study demonstrates that current AI and simulation workflows often assume perfect crystal lattices. These idealized models fail to account for common structural imperfections—especially substitutional disorder—which significantly influence material properties.
2. New Machine-Learning Tool to Detect Disorder
The international team—bringing together experts from the Fritz Haber Institute of the Max Planck Society, Imperial College London, and the University of Bayreuth—developed a machine-learning model capable of reliably detecting disorder in crystalline materials.
3. Large-Scale Screening of Materials Databases
Applying the new tool to existing databases revealed that many previously “promising” materials are likely to behave very differently in experimental conditions. In one dataset, over 80% of materials predicted by simulations showed signs of disorder, calling into question their predicted performance.
4. Improving Future AI-Based Material Discovery
The authors stress that disorder must be incorporated into the next generation of computational workflows. “Our study shows that disorder can be a critical stumbling block in computational materials science if it is not accounted for,” says Prof. Johannes T. Margraf. “With the tools we provide, disordered materials can be detected even in large-scale workflows and treated with suitable computational methods.”

Why It Matters

“Using our models, we can we can predict whether a crystal is affected by disorder and steer material discovery towards computationally well-represented areas,” says Konstantin Jakob, first author of the study and PhD student in the Theory Department of our Institute.
Everyday technologies—from smartphone batteries to rooftop solar panels—depend on highly optimized materials. As society faces urgent challenges such as climate change and energy transition, the demand for new functional materials continues to rise. Yet, discovering such materials experimentally is slow, costly, and complex.

While AI and simulation-based discovery promised to accelerate this process, the new findings show that ignoring structural disorder results in unreliable predictions. In some databases screened by the research team, more than 80% of AI-reccomended materials were likely to exhibit disorder in real experiments, meaning their true properties could deviate dramatically from theoretical expectations.
By identifying this critical gap and providing tools to address it, the study marks an important step toward more trustworthy and efficient computational materials discovery.