A FUSION MACHINE LEARNING DATA SCIENCE PLATFORM TO SUPPORT THE DESIGN AND SAFE OPERATION OF A FUSION PILOT PLANT


B. Sammuli, General Atomics (Principal Investigator)
R. Nazikian, General Atomics (Co-Lead PI)
F. Wuerthwein, UCSD (Co-investigator)
M. Foltin, Hewlett Packard Enterprise (Co-investigator)
C. Michoski, SapientAI LLC (Co-investigator)


Project Objectives and Description:
Machine learning and artificial intelligence (ML/AI) methods hold great promise for accelerating scientific discovery and innovation for the design and safe operation of a Fusion Pilot Plant (FPP). Fully realizing this potential requires broadening community access to high-quality, curated, ML/AI-ready fusion data together with the tools and compute resources required for the efficient development of workflows utilizing fusion data at scale.

We will develop and deploy a feature-rich Fusion Data Platform (FDP) that will enable the creation of high-quality, reproducible AI/ML workflows using fusion data from experiments, as well as interpretative and predictive simulations. The project team will demonstrate the capabilities of the FDP by exercising it on unresolved problems that are critical for the successful design and operation of an FPP. The FDP will make fusion data easier to find, access, interpret, and analyze at scale for ML/AI model development by researchers and data scientists both inside and outside of the fusion community.

These capabilities will be demonstrated by applying the FDP to high-priority research needs relevant to the design of FPP scenarios and control, applicable to many possible realizations of an FPP. The use of very-large-scale simulation data will be demonstrated via simulations of specific fusion plasmas. Handling of experimental data will be demonstrated by creating models for identification of stable, efficient, reactor-relevant plasma regimes. These regime identification models will subsequently be incorporated into ML-augmented control algorithms for maintaining reliable control of a tokamak plasma.

The demonstrations will adhere to FAIR principles, with the resulting models, workflows, and curated data published on the FDP itself and made accessible for various applications, including educational/tutorial purposes, benchmarking, validation of models using common datasets, and model development building on the demonstration workflows themselves. These workflows will illuminate the power of the platform for assisting with critical reactor development tasks and will enable highly collaborative research through the sharing of all elements in the model development process.

Project’s Potential Impact:
The deployment and demonstration of the Fusion Data Platform will facilitate and accelerate the use of transformational data-driven methods in solving many critical problems in fusion science. The FDP deployed in this project will provide the capability to accommodate datasets from multiple fusion devices, as well as computational data, both domestic and international, and will serve as a resource for national laboratories, universities and the emerging private fusion industry.