MLIP School 2026 — Dataset craft for machine-learned interatomic potentials using ACEpotentials.jl
You will build a silicon dataset from scratch, fit ACE potentials, and learn why a low RMSE can still be a bad potential.
The opening talk is a five-minute read of the slides from the room, published flat so it keeps working after the session.
Two ways to work
In your browser
- E1: Oracle, first fit, and why RMSE is not the whole truth
- E1x: Choosing a basis (optional extension)
- E2: Surfaces — extrapolation bites, dataset curation fixes it
- E3: Automating curation — sampling versus selection
- C: The truth about the truth
Work through the day-one exercises in order: E1, E2, E3, then C. E1x is an optional extension — take it after E1 if you are ahead. It resumes nothing and spends none of your labels, so skipping it costs you nothing.
Keep one exercise open at a time. Each of these pages carries its own Python runtime and uses about 1.5 GB of memory, so close the tab for the exercise you have finished before opening the next. Two open at once is fine on most laptops; three is not.
Day two — bring your own system
A separate one-hour session the following day. It resumes nothing from day one: you bring a system of your own and run the same loop on it.
On your own machine
Run:
uvx mograder student https://mlipschool.uk/mograder.tomlYou will need your API key. You were given one on a printed slip at registration. Paste it into the notebook's API key field, and use the same key throughout — E2, E3 and C resume the potential you fit in E1, and models belong to the key that fitted them.