New Study Reveals Limitations of AI-Based Material Prediction

Graphical representation of the diamond crystal structure. Natural deviations from ideal crystal structures are often overlooked by AI applications, which can lead to significant errors in predicting suitable new materials. © Adobe Stock/Molecular Science

Identifying new technological materials among the vast number of possible combinations of elements and structural arrangements is a major challenge in materials science. Experimentally, this search is often limited by the complexity and cost of synthesising and analysing promising candidates. As a result, the use of computer simulations for material discovery has become increasingly attractive in recent years. Particularly in the field of crystalline materials, substantial progress has been achieved. This class of materials, which includes many important compounds such as silicon, steel, and diamond, features atoms arranged in highly regular lattices. However, current computational workflows are based on idealised crystal structures that do not accurately reflect experimental reality. In particular, they ignore crystallographic disorder, which is common in real materials. For example, many crystalline materials contain similar elements that are “mixed” within their lattice—a phenomenon known as substitutional disorder. When information about this disorder is missing or not properly considered, AI or simulation methods make errors in predicting the properties of such materials.

An international research team comprising scientists from the Theory Department at the Fritz Haber Institute, Imperial College London, and the University of Bayreuth—led by Prof. Dr Johannes T. Margraf, Chair of Physical Chemistry V at the University of Bayreuth—has now developed a machine-learning tool that can reliably detect such disordered materials. “With this tool, we can predict whether a crystal is affected by such disorder and steer material discovery towards computationally well-represented areas,” says Konstantin Jakob, first author of the study and research associate at the Fritz Haber Institute.

Using the new model, the team screened databases containing materials previously identified as promising by computer simulations. In all cases, the researchers found that a significant proportion of these predictions are likely to exhibit disorder in experiments—in one instance, more than 80% of the proposed materials were affected. This means that the vast majority of materials found in this database could have markedly different properties in practice than predicted.

“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,” says Prof. Margraf. In future, this tool will significantly enhance both the reliability and efficiency of computer-assisted material discovery.