Machine learning toolset for disruption prediction

Facilitating model development with standardized workflow and evaluation benchmarks

   Yumou Wei, Fernando Valenzuela, Cristina Rea    •     Postdocs

In tokamak plasmas, a disruption is an abrupt and catastrophic event in which the magnetically confined plasma rapidly loses its stored thermal and magnetic energy, leading to a sudden termination of the plasma discharge. Disruptions can impose serious damage to the reactor and the plasma-facing components, making the ability to reliably predict the onset of disruption events necessary for long-term safe operation of a tokamak reactor. Nevertheless, modeling disruptions through first-principle models or physics simulations can be challenging due to the complex interaction between various physics phenomena in addition to unknown physics. This has motivated the use of advanced statistical and AI/ML techniques for analyzing and predicting disruptions.

Leveraging the extensive expertise in machine-learning-driven disruption research of the PSFC Disruption Studies Group, this project aims to develop toolsets and standardized benchmarks for developing and evaluating disruption prediction models. Specifically, it focuses on:

  1. Consolidating previous research products into an open-sourced python library to facilitate rapid prototyping of disruption prediction models. This library, currently named “scikit-disruption,” will offer tools for preprocessing training data, initializing models, and computing performance metrics that are relevant to disruption prediction applications. It is designed to complement standard data science and machine learning libraries such as pandas, scikit-learn, and keras. It is also designed to primarily utilize the dataset produced by DisruptionPy, the multi-machine data retrieval and processing framework currently being developed by the PSFC Disruption Studies Group.

  2. Establishing standardized benchmarks for disruption prediction models. Continuing the research work of DisruptionBench, this part of the project aims to develop standardized protocols for evaluating models using the improved multi-machine dataset produced by DisruptionPy. It aims to encompass multiple scenarios with a focus on transfer learning and online adaptive learning in order to assess a model’s potential performance on a future device.

Application of scikit-disruption in the disruption prediction model development workflow.

The project is currently led by postdoctoral associate William Wei, under Dr. Cristina Rea’s supervision.