EFIT-AI Database Creation and Storage for Machine Learning

The desired application of Machine Learning (ML) and Data Science techniques to advance fusion research has motivated the need for high quality and open access databases. A EFIT-AI database has been assembled and curated: a collection of multiple equilibrium reconstructions for a variety of tokamak discharges. The database features the entirety of the 2019 DIII-D campaign (approx. 2500 discharges) and contains 3 different types of EFIT reconstructions: 1. with magnetics-only constraints; 2. with magnetics + MSE(Motional Stark Effect) constraints; and 3. with OMFIT-automated kinetic constraints (for a subset of shots). The database is currently being used for training of equilibrium reconstruction neural networks aimed at improving real time reconstructions. To incorporate Findable, Accessible, Interoperable, and Reusable (FAIR) data principles, the data was organized according to the ITER IMAS (Integrated Modeling and Analysis Suite) data schema (ontology), and then stored as self-descriptive HDF5 binary files that will be made publicly available. This mapping to IMAS was carried out within the OMFIT framework [https://omfit.io], leveraging the functionalities of the OMAS library [https://gafusion.github.io/omas]. A collection of validation scripts ensures the quality of the training dataset by verifying that all necessary data is available, the equilibria have low Grad-Shafranov equilibrium error, and the equilibrium reconstructions match experimental diagnostic signals. To track different versions of EFIT data generated for this database construction, this work utilized Git and the Open-source Data Version Control (DVC) system, which was developed intentionally for Data Science and Machine Learning projects. Work supported by the US Department of Energy DE-SC0021203, and DE-FC02-04ER54698.