Shadow Masks Predictions in SPARC Tokamak Plasma-Facing Components Using HEAT code and Machine Learning Methods

In this work, the power of machine learning (ML) is used to complement HEAT's (Heat flux Engineering Analysis Toolkit) capabilities, focusing specifically on the 3-D footprint surrogate models for accurate and fast heat load calculations in the divertor of the SPARC tokamak in seconds timescale. PFCs (plasma-facing components) surfaces shadowed from the incident heat flux due to the 3-D geometry of the PFC are referred to as magnetic shadows or shadow masked regions.

PFCs play a vital role in maintaining operational stability. However, simplistic axisymmetric assumptions often fail to capture the intricate interplay between 3-D PFC geometry and 2-D or 3-D plasmas, which can lead to to compromised performance or PFC failure such as melting. HEAT is a tool which addresses the critical need for high-precision 3-D predictions and analysis of PFCs in tokamaks.

ML techniques, particularly classifiers (ML method that categorizes data into different groups), are exploited to develop a surrogate model of the HEAT code, enabling efficient and accurate shadow mask predictions. Utilizing a diverse equilibrium variations database, including ranges of plasma current, q95 and incident magnetic flux angles as inputs to the ML model, this work aims to provide comprehensive insights into the behavior of the divertor system. Downstream implications of incorrect ML predictions will be analyzed from the engineering perspective of needed accuracy for design or operational choices.

By concentrating on shadowed regions within the divertor, this approach seeks to refine predictions and enhance the reliability of the surrogate model. The final goal is to use the surrogate model for real-time control, where a 3D plasma model provides inputs for a 3D temperature calculation. Furthermore, future work aims to implement the models in between shots or in real-time systems for control decisions.