How can we predict how a material absorbs or reflects light - without the need for complex laboratory experiments or computationally intensive simulations? A research team at Technische Universität Ilmenau has now come a decisive step closer to answering this question, which has occupied materials scientists for decades. With the help of artificial intelligence, the scientists at the Institute of Physics have succeeded for the first time in making precise predictions about the optoelectronic properties of materials. In their study, published in the journal Nature Communications they use machine learning methods to decipher the complex relationships between atomic structures and their optical properties - faster, more efficiently and closer to reality than before, thus bridging the gap between computational efficiency and physical accuracy.
Every material reacts differently to light of certain wavelengths. These reactions can be represented as a spectrum - a kind of "fingerprint" of the material for light. In order to find out which wavelengths the material absorbs, for example which colors a material absorbs, which it reflects or transmits, and to understand its electronic and chemical properties, such spectra are usually measured experimentally or calculated theoretically. Traditional calculation methods such as the independent particle approximation (IPA), which only roughly captures some complex interactions between electrons, often reach their limits here and rarely correspond to reality. "Although better approximations are closer to reality, they quickly become exorbitantly expensive to calculate," explains Malte Grunert, Research Associate at the Group of Theoretical Physics 1 and first author of the study.
To solve this problem, the new model developed by the scientists at TU Ilmenau combines IPA data with high-precision data from the so-called random phase approximation (RPA) and relies on a so-called graph attention network, which is adaptive and flexible - and requires little expensive data to get close to reality.
"Important step towards interpretable artificial intelligence for materials science"
In an earlier study published in Small, the Ilmenau researchers had already shown that their graph attention network not only makes predictions, but can also be interpreted. Based on a data set of 10,000 optical spectra calculated quantum mechanically at the High-Performance Computing (HPC) Cluster at TU Ilmenau, the model generated a "map" of the material space that can be understood by humans. With the help of UMAP, a technique often used in biology or medicine to visualize high-dimensional data, the research team was able to visualize how the network "thinks" and intuitively groups materials according to chemical principles such as oxides or nitrides.
"We were therefore able to take an important step towards interpretable artificial intelligence for materials science," says Max Großmann, co-author of the study.
The highlight of the new investigations that have now been published: Through transfer learning - a method in which an already trained model is quickly adapted to new, similar tasks - this extensive, rough data is used to pre-train the model. High-precision RPA data then refines the predictions. The result is a tool that can identify materials efficiently and accurately - faster and more accurately than before.
"With these studies, we were able to show how modern algorithms can climb the so-called Jacob’s Ladder of optoelectronic properties and solve the classic challenges of materials science," explains Prof. Erich Runge, head of Theoretical Physics 1 and co-author of the study.
They not only provide precise predictions that are close to experimental results, but also help to understand the principles behind these predictions and thus really get close to reality - and with surprisingly little expensive data. This is an important step towards data-driven materials science.
The researchers also see the new method as having the potential to significantly accelerate the optimization and development of new sustainable materials, says Prof. Runge, "for example to convert sunlight into electricity more efficiently, to identify sustainable substitutes for critical elements or to make production more cost-effective."
Original publications
M. Grunert, M. Großmann, E. Runge, Machine learning climbs the Jacob’s Ladder of optoelectronic properties, Nat. Commun. 16, 8142 (2025). https://doi.org/10.1038/s41467-025-63355-9
M. Grunert, M. Großmann, E. Runge, Discovery of Sustainable Energy Materials Via the Machine-Learned Material Space. Small, 2412519 (2025). https://doi.org/10.1002/smll.202412519
Contact
Malte Grunert
Group of Theoretical Physics 1