FlowLLM for creating new possible materials

There are countless possible materials, which makes lab testing for all of the combinations difficult. And here comes practical AI:

AI at Meta FAIR and Universiteit van Amsterdam created FlowLLM to create novel materials, which combines LLMs with Riemannian flow matching (RFM) method.

Here's how it works:

• The process of generating materials in FlowLLM starts with a standard input prompt to let the LLM create an initial structure for a potential material.
• Then this text-based output is converted into a 3D representation. If it meets basic stability rules (like valid atomic types and physical lattice parameters), it moves forward to the RFM stage. Otherwise, we discard and restart the sample.
• The RFM model refines the LLM’s output, gradually adjusting atomic positions and lattice parameters for better stability.

Thanks to this approach FlowLLM handles both the discrete and continuous aspects of materials (particularly crystal) geometry. This allows to create stable, unique, and novel materials likely to be achievable in real-world.

Experiments show that FlowLLM generates stable materials 3x faster than previous models and creates unique materials at a 50% higher rate.

Paper: https://arxiv.org/pdf/2410.23405

Code for training: https://github.com/facebookresearch/flowmm