Abstract Details
Abstracts
Author: Joe A Abbate
Requested Type: Poster
Submitted: 2023-04-24 22:49:48
Co-authors: E.Fable, A.Pankin, G.Tardini, E.Kolemen
Contact Info:
Princeton University
100 Stellarator Road
Princeton, NJ 08540
USA
Abstract Text:
For fusion reactor planning and many plasma dynamics modeling workflows, a combination of empirical scalings and physics equations are used to estimate the plasma response to external actuators. In prior work, we have made a machine learning model to predict tokamak plasma evolution using machine learning trained on DIII-D experimental data. We have also generated a database of TRANSP and ASTRA runs predicting temperature evolutions for around 100 DIII-D discharges; and compare the simulator results to simpler physics-based rule-of-thumb scalings. In this poster we outline methodologies for combining physics and experimental data to train machine learning models with the goal of learning dynamics and scalings with the same multimodal mechanism humans employ to get best estimates for plasma evolution. We also show initial results for DIII-D experimental data.
Comments: