Presenter: Yang Hu from EMPA
OptiMat Alloys reframes a data-coverage problem as a software one, and the talk is aimed at research software engineers from any domain — no materials science background needed, though some familiarity with LLM tool-calling and FAIR data helps. Instead of shipping a static dataset, it pairs a natural-language LLM agent with a simulation backend: the agent parses a request, dispatches molecular-dynamics jobs driven by foundation ML interatomic potentials, and commits every result to a shared, versioned database with full provenance, while running multiple potentials and structural realisations per composition to yield built-in uncertainty estimates.
In doing so it extends FAIR principles from pre-computed repositories to on-demand generation, built on open scientific-Python tooling. Listeners will take away a working pattern for putting an LLM agent in front of expensive scientific compute so non-experts can drive it in plain language, and how to keep machine-generated results trustworthy through provenance, versioning, and uncertainty quantification.
A public demo is available at https://youtu.be/lQzuorkzPMc