Dynamic modeling of Alfvén eigenmodes using Machine Learning on DIII-D

Controlling unstable Alfvén eigenmodes (AE) are required to achieve a sustainable burning plasma in magnetically confined fusion devices. This work develops a deep neural network-based dynamical system to simulate the plasma environment for future reinforcement learning (RL) control of AEs at DIII-D. The model uses multimodal inputs including plasma profiles, actuator parameters, and plasma shaping to predict normalized beta, neutron rates, and an unstable toroidal Alfvén eigenmode (TAE) likelihood. These outputs define the reward system that an RL agent would use to autonomously control TAEs in real-time at DIII-D. We present the methodology for database development, unstable TAE likelihood classification, neural network architecture design, and training optimization. Results show the model's capability to simulate the evolution of the plasma state in response to variations in neutral beam power, gas puffing, and electron cyclotron heating. Supported by the U.S. Department of Energy under DE-FC02-04ER54698, DE-AC02-09CH11466, DE-SC0021275, DE-SC0020337, DE-SC0014664.