Using AI to explore new research topics in materials science
The number of scientific publications is growing so rapidly that scientists can no longer keep track of all the work, even in their own field. In a recent study, researchers from the Karlsruhe Institute of Technology (KIT) and partners are showing how new research ideas can nevertheless be derived from this abundance: with the help of artificial intelligence (AI), they are systematically evaluating publications in materials science and identifying possible research approaches. Publication in Nature Machine Intelligence. (DOI: 10.1038/s42256-026-01206-y)
Materials form the basis of numerous technologies - from batteries and solar cells to electronic components and medical applications. Materials science is therefore considered a cross-sectional discipline and influences many fields of research and technology. The number of scientific publications is correspondingly large, but their findings can only be used if relevant developments and correlations are recognized. With this in mind, the authors of the study investigated how scientific publications can be systematically evaluated. "Our aim is to support researchers in creative thought processes by making new questions and possible collaborations between disciplines visible," says Professor Pascal Friederich from the Institute of Anthropomtics and Robotics at KIT.
Combining large language models and machine learning
In their current work, the researchers are combining large language models (LLMs) and machine learning (ML) methods. First, the LLMs identify key technical terms and scientific concepts in the specialist articles. This information is used to create a concept graph, i.e. a knowledge network in which each keyword forms a node. A second machine learning model links these nodes together if terms are mentioned together particularly often in scientific publications.
"If our language model recognizes, for example, that terms such as 'perovskite' and 'solar cell' occur together more and more frequently, a new connection is created in the concept graph," says Thomas Marwitz, first author of the study and Informatics student at KIT. "An ML model then analyzes the development of these connections in order to predict which combinations of scientific concepts could become more important in the next two to three years." To do this, the ML model evaluates how links between concepts change over many years. If certain concepts become increasingly linked, this may indicate that a new field of research is developing. Conversely, weakening links can indicate that certain topics are becoming less important.
A knowledge network of technical terms created by AI analyses visualizes trends and new research ideas in materials science.
research ideas in the materials sciences. (Graphic: Thomas Marwitz, KIT)
Impetus for new research ideas
The results can provide researchers with indications of combinations of topics that have received little attention to date. Interviews with experts revealed that some of the AI-generated proposals were actually rated as innovative and promising. "We don't want to replace the researchers," emphasizes Friederich. "The results are not a machine for inventions, but an analysis tool that can help to discover new ideas and potential collaborations in a more targeted way. We want to provide targeted support for scientific creativity."
The study shows how large volumes of scientific literature can be systematically evaluated with the help of AI. This approach could also help to make emerging research trends visible earlier in other specialist areas.
Original publication
Thomas Marwitz, Alexander Colsmann, Ben Breitung, Christoph Brabec, Christoph Kirchlechner, Eva Blasco, Gabriel Cadilha Marques, Horst Hahn, Michael Hirtz, Pavel A. Levkin, Yolita M. Eggeler, Tobias Schlöder, Pascal Friederich: Predicting new research directions in materials science using large language models and concept graphs. Nature Machine Intelligence, 2026. DOI 10.1038/s42256-026-01206-y.