Please Note: further content e.g. recordings, slides etc. will be added after the townhall has concluded
Machine Learning in materials modelling has revolutionised and transformed the field in recent years. Based on availability of good data from computations and/or experiment one can use machine learning to train expert systems using large language models (à la chatGPT), train surrogate models to predict properties or structure avoiding the use of simulations or simply speed up simulations by using machine learnt interatomic potentials (MLIPs). However, these advances require development of new infrastructure to capture, store and distribute ML models. Work in this pathfinder is dedicated to making possible the creation of these models by providing data infrastructure, workflows for creation and exploitation of machine learnt models.
This work begins with exploring the field of machine learning interatomic potentials which is a key technique in atomistic modelling. The development focuses around three key areas:
- Production of aiida-mlip plugin for AiiDA – enhances provenance tracking, reproducibility and sharing of processes and data
- Janus-core tools for integration of MLIPs
- abcd – specialised database for MLIP training data, faster searching through implementation of opensearch
You can find out more about some of the work and development below by exploring the poster, trying some tutorials or looking at the code in GitHub. If you want to provide feedback on the development please utilise the GitHub issues in the aiida-mlip or janus-core repositories.
The Team members at the June Townhall
Alin-Marin Elena
Elliot Kasoar
Additionally Jacob Wilkins (STFC) works on the Data to Knowledge pathfinder
Federica Zanca