Machine learning control of DIII-D profiles using a linear profile predictor

Robust DIII-D profile control is essential for achieving desired plasma conditions and reaching difficult scenarios. We present a novel profile controller based on the Linear Recurrent Autoencoder Network [1, 2], which enables efficient calculation of optimal actuator trajectories in real time. We train a machine learning model using DIII-D data (as in [3]) by mapping the non- linear dynamics into a linear latent space, thereby allowing linear model predictive control methods to be used. The controller is input the present profiles (electron and ion temperature, density, q and rotation) and outputs the optimal actuator trajectory (NBI power and torque, ECH power, gas puffing, Ip, Bt) to approach the target profiles in our scenario. Simulations of DIII-D shot profile control are performed using the profile predictor presented in [3], showing the algorithm’s potential for robustly reaching difficult scenarios in DIII-D.

This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Fusion Energy Sciences, using the DIII-D

National Fusion Facility, a DOE Office of Science user facility, under Awards DE-AC02- 09CH11466, DE-FC02-04ER54698 and DE- SC0021275.

[1] S. E. Otto and C. W. Rowley 2017, arXiv:1712.01378

[2] J. Abbate et al, J. Plasma Phys. (2023), vol. 89, 895890102 [3] J. Abbate et al 2021 Nucl. Fusion 61 046027