New paper: Generative Hierarchical Materials Search

[maykcaldas]

Paper: Generative Hierarchical Materials Search

Authors: Sherry Yang, Simon Batzner, Ruiqi Gao, Muratahan Aykol, Alexander L.

Abstract: Generative models trained at scale can now produce text, video, and morerecently, scientific data such as crystal structures. In applications ofgenerative approaches to materials science, and in particular to crystalstructures, the guidance from the domain expert in the form of high-levelinstructions can be essential for an automated system to output candidatecrystals that are viable for downstream research. In this work, we formulateend-to-end language-to-structure generation as a multi-objective optimizationproblem, and propose Generative Hierarchical Materials Search (GenMS) forcontrollable generation of crystal structures. GenMS consists of (1) a languagemodel that takes high-level natural language as input and generatesintermediate textual information about a crystal (e.g., chemical formulae), and(2) a diffusion model that takes intermediate information as input andgenerates low-level continuous value crystal structures. GenMS additionallyuses a graph neural network to predict properties (e.g., formation energy) fromthe generated crystal structures. During inference, GenMS leverages all threecomponents to conduct a forward tree search over the space of possiblestructures. Experiments show that GenMS outperforms other alternatives ofdirectly using language models to generate structures both in satisfying userrequest and in generating low-energy structures. We confirm that GenMS is ableto generate common crystal structures such as double perovskites, or spinels,solely from natural language input, and hence can form the foundation for morecomplex structure generation in near future.

Link: https://arxiv.org/abs/2409.06762

Reasoning: produce the answer. We start by examining the title and abstract for any mention of language models. The title "Generative Hierarchical Materials Search" does not explicitly mention language models, but it does suggest the use of generative models. Moving to the abstract, we see that it describes a system called GenMS, which includes a language model as one of its components. Specifically, the language model takes high-level natural language as input and generates intermediate textual information about a crystal. This indicates that the paper involves the use of a language model as part of its methodology.