Real-time machine learning control of a Hall thruster discharge plasma

Facility effects are known to alter Hall thruster performance between ground and space. Furthermore, discharge current oscillations are linked to impacting both the lifetime and thrust of Hall thrusters. In this work, we present a machine learning–based control framework designed for nonlinear, time-varying systems with hysteresis that can mitigate BM oscillations and adapt to environmental changes. Unlike reactive approaches such as PID control, our method is predictive and flexible across a range of control objectives. Specifically, we employ echo state networks (ESNs), a class of recurrent neural networks well-suited for time-series prediction. To train and test the model, we apply a diverse set of discharge voltage perturbations—including sine, triangle, square, ramp, chirp, bandlimited gaussian noise, pseudorandom binary sequence, and PID—on a 6-kW H6 Hall thruster operated at 400 V as well as 250 V at 4.3 A, and measure the resulting discharge current as a part of a process known as system identification. We found that 34 VRMS Gaussian noise perturbations can reduce discharge current oscillations locally by 84.6%, making them an ideal system identification signal. Using extended convergent cross mapping, we confirm causal relationships between the applied perturbations and discharge current response. The ESN achieves 90% prediction accuracy on previously unseen time-series data based on variance. We then integrate the ESN into a nonlinear model predictive control (NMPC) framework, creating a machine learning controller, to compute optimal control trajectories that suppress discharge oscillations. NMPC simulations indicate that discharge current oscillations can be reduced by over 90%. Finally, we employ a deployment strategy in which the learned control policy is embedded as a multilayer perceptron (MLP) on a field-programmable gate array, enabling real-time experimental implementation. The machine learning (ML) controller utilizing the MLP reduces oscillations by 1.33% while PID increases oscillations by 1.44% at the 250 V operating point. The ML controller has the potential to be applicable to other operating points by following a standard procedure of system identification and control optimization. Online and adaptive ML control can be used in the future to improve oscillation reduction.