In tokamaks, operators are able to control plasmas through a variety of “actuators” during shots: from neutral beams that heat and rotate the plasma, to coils which induce plasma current. Experimental proposals entail physicists specifying a desired plasma “state” of interest, which can be described by profiles of density, temperature, pressure, safety factor (q), and rotation. Operators and physicists usually work together to pre-specify a “path” of actuator signals through time that will successfully realize the state, using feedback control for realtime adjustments. However, the process of finding a successful actuator path is difficult and entails a lot of trial-and-error. “Model-predictive control” could make this process more efficient, saving physicists time and ensuring more successful physics experiments during future tokamak campaigns. In model-predictive control, a realtime model predicts the way the plasma state will evolve given various settings on actuators, then chooses settings which yield the plasma state closest to the physicist’s desired end-state. Realtime physics models are not always accurate for regimes of interest. We are therefore developing a machine-learning model which generates a single prediction in under 100 microseconds, using only realtime diagnostics.

Algorithm would choose to increase the “pinj” signal to move the plasma toward the highest temperature profile (shown in green).

Plasma Behavior Monitoring using High Resolution Diagnostics

Plasma Control Group is leading a multi-institutional research project to investigate machine learning for real-time fusion plasma behavior monitoring using high resolution diagnostics. We focus on two tracks of research; (1) detecting and classifying instabilities such as Alfven-Eigen (AE) modes in the core of plasma based on Electron Cyclotron Emission signals (2) developing a general framework for preprocessing such data, e.g., denoising spectrograms, before feeding them to the ML model.
We develop several ML models to classify five AE modes, namely, BAAE, BAE, EAE, RSAE, TAE in a dataset of ~1000 discharges. We particularly pay attention to the ML models such as Reservoir Computing Networks which are easy to train yet effective in processing temporal information in time-series data such as ECE. Figure below shows the performance of a 2-layer RC model which has been trained on down-sampled ECE signals (from 500KHz to 1KHz). Our preliminary experiments show a hit rate of ~90% in detecting AE modes.

The ECE processing pipeline for detecting AE activities of DIII-D shot #193348. 40 ECE spectrograms are enhanced using denoising auto-encoders and are fed to a recurrent neural network that has been trained to detect AE activities at each time step.

Diagnostic reduction and diagnostic upsampling using data-driven approaches

In fusion research devices and reactors, diagnostic data’s accuracy and breadth are critical to understanding and achieving optimal performance. However, the design and operational constraints of pilot plants, such as ITER, often limit the availability and scope of these diagnostics. We are working on pioneering machine learning-based approaches to synthetic diagnostics that aims to circumvent such limitations. Leveraging neural network (NN) and parallel GPU processing, we successfully recreated diagnostic signals, emphasizing both spectrograms and amplitude reconstructions. Our approach not only allows for enhanced signal visualization but also provides a potential solution to diagnostic restrictions in future fusion reactors. Preliminary results indicate promising accuracy in signal reconstructions, highlighting the feasibility and significance of integrating machine learning techniques in fusion research.

Reconstructing CO2 interferometer channels from ECE data. At each time step the LSTM reconstructs each CO2 channel by looking at ECE signals at time t along with the hidden state over the previous values of the three input channels.