Ilmenau researchers are revolutionizing materials science with AI!

Ilmenau researchers are revolutionizing materials science with AI!

On September 23, 2025, researchers at the Ilmenau University of Technology presented significant advances in materials research. By developing a graph attention network that can both make predictions and deliver interpretable results, they are ushering in a new era of artificial intelligence in science. The study, which is based on a comprehensive data set of 10,000 quantum mechanically calculated optical spectra, was carried out at the HPC cluster at TU Ilmenau and represents an innovative approach to analyzing materials. TU Ilmenau reports that this model is capable of generating an understandable “map” of the material space.

The team uses UMAP (Uniform Manifold Approximation and Projection) to visualize high-dimensional data. This reveals how the network categorizes materials based on their chemical principles. Max Großmann, co-author of the study, emphasizes that this represents significant progress towards interpretable AI for materials science. These new methods enable more precise identification of materials, improving not only the speed but also the accuracy of predictions.

Innovative techniques for materials research

The study uses transfer learning to adapt already trained models to new tasks. Coarse data is used to pre-adjust the model. High-precision RPA data then refines the predictions. According to Prof. Erich Runge, another co-author, modern algorithms show promising approaches to solving the challenges in materials science. Their predictions are not only accurate, but also close to experimental results, promoting understanding of the underlying principles.

Another crucial aspect of the study is the potential to accelerate the development of new sustainable materials. This could result in materials that, for example, enable sunlight to be converted more efficiently into electricity, which is particularly important in times of climate change and the energy transition.

Graph Attention Networks – A new dimension in AI

A fundamental part of the research is the foundation on which the graph attention network is based, established by other scientists such as Petar Veličković and his colleagues. Their work, published in a paper titled “Graph Attention Networks,” describes new neural network architectures for graph-structured data. These architectures use masked self-attentive layers to overcome some of the drawbacks of previous methods based on graph convolutions. The results are impressive; The GAT models have achieved outstanding results in four major transductive and inductive graph benchmarks, such as Cora and Pubmed. arxiv documents this remarkable development.

The combination of these innovative approaches not only represents a step into the future of materials science, but also shows how AI and machine learning can revolutionize existing processes. The new methods and models create a clear perspective on the challenges and opportunities that the responsible use of resources will offer in the coming years.

The original publications that explain these developments in more detail are:
– M. Grunert, M. Großmann, E. Runge, “Machine learning climbs the Jacob’s Ladder of optoelectronic properties”, Nat. Commun. 16, 8142 (2025).
– M. Grunert, M. Großmann, E. Runge, “Discovery of Sustainable Energy Materials Via the Machine-Learned Material Space”, Small, 2412519 (2025).