Materials are at the foundation of many technologies, such as batteries, solar cells, electronic components, and medical applications. This makes materials science an interdisciplinary field with strong impact on many areas of research and technology and a correspondingly large volume of research papers. However, the findings described in these papers are only useful if relevant trends and relationships can be identified. Against this backdrop, the study’s authors examined ways of systematically analyzing scientific papers. “Our goal is to support researchers in their creative thought processes by shedding light on new avenues of research and opportunities for interdisciplinary cooperation,” said Professor Pascal Friederich from KIT’s Institute of Nanotechnology.
Combining Large Language Models and Machine Learning
In their project, the researchers combined large language models (LLMs) with machine learning (LM) methods. The LLMs begin by identifying key terminology and scientific concepts in the journal articles. This information is used to generate a concept graph, a knowledge network in which each keyword forms a node. A second machine learning model connects nodes when terms are mentioned together particularly often in scientific papers.
“For example, if our LLM observes that terms like ‘perovskite’ and ‘solar cell’ appear more often together, it will draw a new link in the concept graph,” said Thomas Marwitz, a computer science student at KIT and lead author of the study. “Then an ML model analyzes trends in these links to predict which combinations of scientific concepts could become more significant in the next two or three years.” The ML model does this by analyzing how links between terms change over many years. When certain concepts are linked with increasing frequency, this can be an indication that a new field of research is developing. Conversely, a decrease in the number of links can be an indication that certain topics are losing relevance.
Ideas for New Avenues of Research
The results of the analyses can direct researchers toward topic combinations that previously received little attention. Interviews with experts showed that they did indeed see some of the AI-generated suggestions as innovative and promising. “We’re not trying to replace researchers,” Friederich emphasized. “Our findings aren’t an invention machine, they’re an analytic tool that can help to identify new ideas and opportunities for collaboration more effectively. Our aim is to provide targeted support for scientific creativity.”
The study shows how large amounts of scientific literature can be systematically analyzed using AI. This approach could also help to reveal emerging research trends in other scientific fields.
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 .
In close partnership with society, KIT develops solutions for urgent challenges – from climate change, energy transition and sustainable use of natural resources to artificial intelligence, sovereignty and an aging population. As The University in the Helmholtz Association , KIT unites scientific excellence from insight to application-driven research under one roof – and is thus in a unique position to drive this transformation. As a University of Excellence, KIT offers its more than 10,000 employees and 22,800 students outstanding opportunities to shape a sustainable and resilient future. KIT – Science for Impact.
Nature
Predicting new research directions in materials science using large language models and concept graphs
1-Apr-2026