Abstract Details
Abstracts
Author: Mark Cianciosa
Requested Type: Poster
Submitted: 2026-03-20 09:15:50
Co-authors:
Contact Info:
Oak Ridge National Laboratory
9973 Hummingbird Ln
Knoxville TN, 37923
United States
Abstract Text:
Compute capabilities limit the number of stellarator configurations that can be explored during optimization. While engineering considerations often reject cases, the solution space remains vast. Practical optimizations end up becoming a multistage validation exercise. Neural-Network (NN) surrogate models, once trained, can operate in fractions of a second while computing large batches of cases. Additionally, NNs have interpolation properties which can predict solutions not within the original training set. The challenge of training these models is the data sets needed to train them. In this project we aim to train a NN-model surrogate using the physics constraints directly. By training on the physics directly, we avoid the costly step of generating training data. At each step in the training process, the input space is randomly sampled, and the model is refined to produce a solution closer to equilibrium. Once trained, the NN model surrogate would represent the continuum of equilibrium solutions. From this continuum solution, we can identify manifolds of optimized configurations. Searching along these manifolds could identify configurations more amenable engineering without compromising confinement. Using a machine learning framework, we have developed equilibrium constraints for the nested flux surfaces, radial current, the energy, and radial force balance. We will demonstrate batch solving equilibria and initial results for training NN surrogate models.
This work is sponsored by US DOE under DE-AC05-00OR22725 with UT-Battelle, LLC.
Characterization: 1.0
Comments: