Artificial intelligence (AI) can significantly accelerate the search for new materials. However, many models require substantial computing power or are difficult to adapt after training. The “Virtual Material Design” business area at Fraunhofer SCAI has now developed two methods designed to overcome the hurdles of AI-supported materials research: “Flexible Cutoff Learning” makes models for atomic simulations adaptable after training. “Materium” uses language model techniques to generate new crystal structures more quickly.

Better Control of Simulations

AI models can simulate atomic interactions and use this to predict material properties. For such calculations, they must determine which atoms interact. To do this, they use so-called cutoff radii: The model considers only atoms within a specified distance. A larger radius can increase accuracy but also raises the computational effort.

In conventional models, the cutoff radius is set at the beginning of training and cannot be changed afterward. Flexible Cutoff Learning takes a different approach: the cutoff radius is chosen randomly for each atom. This allows the model to learn how to handle different ranges. The trained model can then be adapted to various applications without retraining. In tests with molecular crystals, computational costs fell by over 60 percent, with only marginally higher error rates.

Generating Crystal Structures Like Sentences

Anyone searching for new materials with specific properties faces an overwhelming number of possible atomic combinations and crystal structures. Computers are intended to facilitate the search for suitable candidates.

“Materium” translates crystal structures into a sequence of tokens, like how a language model processes a sentence. The tokens encode information about atoms, oxidation states, spatial arrangements, and lattice parameters. This allows transformer architectures—which also underpin modern language models—to be applied in materials research.

Unlike diffusion-based methods, which require many incremental calculations, “Materium” generates structures directly rather than developing them through many successive computational steps. As a result, the method requires less computational effort: A model with 43 million parameters was trained in about four hours on a single Nvidia A100 GPU, generating a structure takes less than a second. By comparison, the diffusion-based model “MatterGen” required several days of training.

The publication shows that for properties with fluctuating values, such as density, space group, bandgap, and magnetic density, exact target values cannot yet be reliably achieved.

Practical Benefits for Research

Both methods aim to reduce the computational effort required for AI-assisted materials research. Flexible Cutoff Learning makes already trained simulation models more adaptable. “Materium” accelerates the computer-aided generation of new crystal structures. As a result, academic and industrial research teams could use existing hardware to investigate more variants and select material candidates more quickly for further simulations and material testing.

Read the full papers:

Flexible Cutoff Learning: https://doi.org/10.1088/2632-2153/ae9239

Materium: https://doi.org/10.1016/j.commatsci.2026.114980