A large-scale automated EFIT recomputation workflow for disruption studies at 1 kHz

The open-source DisruptionPy library [1, 2] is routinely used by the MIT PSFC Disruption Studies Group to directly interface the data storage servers and retrieve, process, and organize data in a highly efficient and scalable fashion into convenient databases for subsequent downstream analysis and training for AI/ML models. In recent years, a rising need for denser time bases of fusion data has clashed with the infrequent nature of typical equilibrium computations. Example applications include radiative collapses and fast impurity injections, e.g. UFOs, for which rapid changes in equilibrium can more accurately inform model development. We have therefore embarked in a thorough 1 kHz recomputation of the EFIT equilibrium reconstruction code for 20,000+ discharges of Alcator C-Mod, to be leveraged by all the usual DisruptionPy physics methods in order to retrieve massive datasets on a regular time base for training AI, with explicit focus on time-dependent ML models. The present contribution analyzes the functional and technical pitfalls of such a workflow, and details the validation steps put in place so as to be able to trust the freshly-recomputed equilibria and any subsequent downstream computation.

[1] Trevisan, et al, Zenodo (2024) 10.5281/zenodo.13935223

[2] Trevisan, et al, submitted to JOSS (2025)

Work supported by the DOE FES under Award DE-SC0024368.