A New AI Tool Can Design Crystals from Simple Text Prompts

📄 Full Paper

💬 Ask

Scientists are increasingly relying on new materials for cutting-edge technologies, from semiconductors and solar cells to batteries. But designing those materials is often a time-consuming, expensive, and laborious process. Now, researchers at Google DeepMind have developed an AI tool that can generate new crystal structures based on simple text instructions.

The AI, called GenMS (Generative Hierarchical Materials Search), is the first end-to-end system that takes natural language as input and generates complete crystal structures that satisfy user requirements. GenMS is a hierarchical search engine that leverages three components:

• A large language model (LLM) that takes a user’s text instructions and generates candidate chemical formulae.
• A diffusion model that takes these formulae as input and generates 3D structural representations of crystals.
• A graph neural network (GNN) that predicts properties like formation energy (how stable the structure is) for the generated crystals.

GenMS employs a multi-objective optimization framework. It learns to identify the best structure by balancing the following two factors:

• High-level heuristics: How well does the chemical formula generated by the LLM satisfy the user’s request?
• Low-level heuristics: How well does the crystal structure generated by the diffusion model satisfy the user’s request?

GenMS uses a tree search algorithm to explore the space of possible crystal structures. It starts with a set of initial chemical formulae generated by the LLM. For each formula, GenMS samples structures from the diffusion model and uses the GNN to predict their properties. The algorithm then selects the top-scoring structures based on a combination of the high-level and low-level heuristics.

The researchers tested GenMS on three major crystal families — perovskites, pyrochlores, and spinels — and found that GenMS significantly outperformed baseline methods in generating unique, valid, and potentially stable crystal structures. They demonstrated that GenMS can generate structures for a wide range of user requests, including those for complex structures like layered structures, double perovskites, and spinels.

The researchers also examined the individual components of GenMS. They found that the use of compact representations for crystals led to more efficient generation, and that the GNN effectively guided the selection of low-energy structures.

The study highlights the promising potential of GenMS for revolutionizing materials discovery. It shows that AI can not only automate the process of designing crystals, but also accelerate the pace of innovation in materials science. The research team believes GenMS will eventually be able to generate highly complex structures, like Mxenes and Kagome lattices, and that it could potentially be extended to generate molecules and protein structures from natural language.

“These results show that we’re moving towards a future where AI can be a powerful tool for designing materials, not just for generating basic structures,” says lead author Sherry Yang. “GenMS is still under development, but it already has the potential to revolutionize the way we design and discover new materials.”

Chat about this paper

To chat about this paper, you'll need a free Gemini API key from Google AI Studio.

Your API key will be stored securely in your browser's local storage.