Scientists Develop AI Tool for Designing Novel Materials

© iStock

An international team of scientists, including researchers from HSE University, has developed a new generative model called the Wyckoff Transformer (WyFormer) for creating symmetrical crystal structures. The neural network will make it possible to design materials with specified properties for use in semiconductors, solar panels, medical devices, and other high-tech applications. The scientists will present their work at ICML, a leading international conference on machine learning, on July 15 in Vancouver. A preprint of the paper is available on arxiv.org, with the code and data released under an open-source license.

New materials form the foundation of modern technologies, from electronics to medicine. Thanks to AI, the development timeline for novel materials has been reduced from decades to just a few months. However, scientists must not only rapidly generate new compounds but also accurately predict their properties—such as whether the material will conduct electricity or be durable. 

The properties of a material are primarily determined by the internal symmetry of the crystals that compose nearly all solid substances. However, many modern generative models do not explicitly account for symmetry.

Researchers from HSE University, the National University of Singapore, Nanyang Technological University, and Constructor University have developed the Wyckoff Transformer (WyFormer), a novel machine-learning algorithm that rapidly generates materials with specified symmetries and predicts their stability and performance.

The model is based on representing a crystal using Wyckoff positions—mathematically precise coordinates that specify where atoms can be located based on crystal lattice symmetry. This approach enables a concise and universal representation of the structure

'Imagine your reflection in a mirror. While our face is symmetrical, some features come in pairs belonging to two distinct classes, such as the right and left eyes. Other features belong to a single class, like the tip of the nose. In mathematical terms, the nose corresponds to Wyckoff position A, while the eye corresponds to Wyckoff position B. In other words, Wyckoff positions are the key points that define symmetry and enable us to recognise a human face in the mirror,' explains Ignat Romanov, co-author of the paper and Junior Research Fellow at the Faculty of Computer Science, HSE University. 

Figure 1. Schematic illustration of Wyckoff positions (top) and crystal representation based on Wyckoff positions using WyFormer (bottom).

© Nikita Kazeev et al. Wyckoff Transformer: Generation of Symmetric Crystals

The new model was trained on an open database of real materials from the Materials Project. The AI leverages a transformer architecture to generate novel crystal synthesis recipes that automatically conform to symmetry rules. 

'There are countless ways atoms can combine. Trying to find useful combinations and design new materials without understanding their symmetry rules is like sticking LEGO blocks together without a plan. While such improvisation may occasionally produce interesting results, it often leads to unstable structures,' says Nikita Kazeev, co-author of the paper, Research Fellow at the Institute for Functional Intelligent Materials (I-FIM), National University of Singapore, and graduate of the HSE Faculty of Computer Science. 'Our AI tool has effectively learned from all good LEGO instructions—it knows how to generate models that are stable, aesthetically pleasing, and functional. Returning from the metaphor to materials, it understands the full range of symmetry options and can predict a material's properties even without knowing the exact atomic positions within the unit cell.' 

Compared to other models, WyFormer produces a higher proportion of stable and unique structures, generates crystals with more accurate symmetry, and achieves a generation rate an order of magnitude faster. 

The researchers plan to apply the model to develop practical materials for solid electrolytes and materials with specified thermal conductivity.