 |
Center for Nonlinear Studies |
While the predictive capabilities of a Large Language Model (LLM) are limited by its training set, LLM agents provide the LLM with the option to request information outside the training set. For the specific case of atomistic simulations, we implemented LLM agents to calculate material properties, e.g., the Equation of State. With these agents, the LLM is able to invert the materials design process and identify the alloying compositions which fulfill the user-defined requirements for the target properties.
This level of autonomous materials design is achieved with simulation workflows developed by materials science experts including feedback loops to handle edge cases. These workflows are based on the experience from high-throughput screening studies and aim to handle both scientific challenges like the uncertainty based on the choice of basis set as well as technical challenges like calculations which are prematurely terminated by reaching the memory limit. In this context, the LLM can orchestrate the execution of the workflows and the coupling of the individual workflow steps.
The workflows are developed using the pyiron workflow framework, based on the Python workflow definition, an interoperable workflow format to exchange workflows between AiiDA, jobflow and pyiron. Based on the Python programming language, the pyiron workflow framework can be directly coupled to the LLM, which enables this level of autonomous materials discovery.
Bio: Jan Janssen is the group leader for Materials Informatics at the Max Planck Institute for Sustainable Materials. His group focuses on applying methods from computer science including machine learning to discover novel sustainable materials with applications ranging from machine-learned interatomic potentials to large language model agents for atomistic simulation. Previously, Jan was a director’s postdoctoral fellow in the T-division at LANL as well as an invited postdoctoral fellow at the University of Chicago and the University of California Los Angeles. Besides his research work, Jan maintains over 900 open-source materials informatics software packages for the conda-forge community and is a regular contributor to open-source software on Github.
Host: Danny Perez (T-1)