Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 18: SYMD contributed

MM 18.9: Vortrag

Mittwoch, 19. März 2025, 12:30–12:45, H23

Inverse Materials Design with Large Language Models — •Jan Janssen and Joerg Neugebauer — MPI for Sustainable Materials, Düsseldorf, Germany

Large language models (LLM) are trained on a vast amount of scientific literature to learn the included semantic, conceptional, and statistical relationships. The LLM applies these relationships to generate responses in natural language based on the context of the conversation. This raises the question: Can a LLM replace a scientist? Or how does the thought process of a scientist differ from the statistical approach of the LLM? Can the LLM make us better scientists?

We benchmark the capabilities of current LLMs to design new materials using atomistic simulations. While the required Python programming is challenging for the LLM and suffers from hallucination, this can be addressed with an agent-based approach by providing the LLM with a series of simulation workflows for the pyiron workflow framework. With these simulation workflows the LLM is not only capable to calculate material properties but can also invert the process and leverage statistical models to identify alloying compositions which match a pre-defined materials property, enabling inverse materials design.

Our benchmarks highlight the importance of developing scientific workflows. The more a workflow reduces the technical and scientific complexity of studying a given materials property the easier it is to use for LLMs and scientists alike. In this way LLMs also help us as scientists to validate and improve our scientific workflows. https://github.com/jan-janssen/LangSim

Keywords: Large language models; Inverse Materials Design; pyiron; Atomistic Simulation