Hall effect thrusters (HETs), electrostatic devices that have become increasingly vital for satellite propulsion due to their superior mass efficiency compared to chemical systems, frequently face operational challenges linked to discharge plasma instabilities. Among these, breathing mode oscillations—predatory–prey interactions between neutral gas and plasma—can drive current amplitudes exceeding the mean, contributing to wall erosion, electromagnetic interference, and potential damage to the power processing unit. While conventional active control methods, such as proportional-integral-derivative (PID) controllers, have been explored, their performance is limited by the highly nonlinear, time-varying, and hysteretic nature of the discharge plasma. A research team has now successfully demonstrated an advanced machine learning-based control framework, utilizing Echo State Networks (ESNs) and nonlinear model predictive control (NMPC), that provides a more predictive and flexible approach to suppressing these detrimental oscillations, paving the way for more intelligent, adaptive power processing systems for advanced space missions.
The study, conducted using a 6-kW class H6 Hall thruster in a specialized vacuum test facility, sought to develop a model that could characterize the plasma’s response to electrical perturbations. Researchers implemented a robust system identification process, applying a diverse set of voltage perturbations, including sine, triangle, square, ramp, and bandlimited Gaussian noise, to the discharge circuit. By recording the resulting discharge current trajectories, the team aimed to build a data-driven model capable of capturing the complex, nonlinear dynamics of the plasma. The facility used liquid nitrogen-cooled cryopumps to replicate space-like conditions, allowing the team to operate the H6 thruster at constant magnetic fields and propellant flow rates, providing a stable baseline for characterizing the thruster’s AC behavior at 250 V and 400 V operating points.
A critical observation during the system identification phase was the unexpected efficacy of Gaussian noise in damping the breathing mode. When applying 34 VRMS Gaussian noise perturbations, the researchers observed that the breathing mode amplitude in the discharge current frequency domain could be reduced by up to 84.6%. This damping effect, characterized by destructive interference through random phase interactions, suggests that continuous-range frequency signals are significantly more effective at altering the plasma’s breathing mode than single-frequency perturbations. This finding provided a simplified yet powerful method for characterizing the discharge plasma, yielding higher signal-to-noise ratio data that proved essential for training the subsequent machine learning models.
To model the plasma dynamics, the team utilized Echo State Networks, a class of recurrent neural networks noted for their efficiency in time-series prediction and their ability to handle nonlinear, nonstationary dynamics without the extreme computational overhead of long short-term memory (LSTM) networks. The ESNs were trained to map the control voltage and resulting discharge current, achieving a 90% prediction accuracy based on variance. This predictive capability allowed the researchers to integrate the ESN into a nonlinear model predictive control (NMPC) framework. Unlike reactive PID controllers, the NMPC approach solves a nonlinear optimization problem at each step to compute optimal control trajectories that actively minimize the root mean square (RMS) of discharge current oscillations, effectively anticipating the plasma’s behavior over a short look-ahead horizon.
Following the successful demonstration of this control architecture in simulation, the team transitioned to a real-world experimental setup. To enable real-time implementation, the learned NMPC control policy was condensed into a multilayer perceptron (MLP) trained through imitation learning. This MLP was then programmed onto an AMD ZCU104 field-programmable gate array (FPGA), enabling control inference on a microsecond time scale. By optimizing the neural network latency to 785 nanoseconds, the team ensured the controller could make decisions within a 10-microsecond loop period. The experimental validation included the use of a simultaneous perturbation stochastic approximation (SPSA) algorithm to refine the weights of the MLP controller during thruster operation, accounting for environmental variations that might deviate from the initial model.
Experimental results at the 250 V operating point highlighted the performance disparity between the machine learning controller and the conventional PID approach. While the ML-based controller achieved a 1.33% reduction in discharge current oscillations, the PID controller actually increased oscillations by 1.44% under the same conditions. These results underscore the limitations of PID control, which struggles to handle the inherent nonlinear coupling between the thruster’s discharge voltage and current. The successful experimental deployment of the MLP on the FPGA, reproducing simulation-predicted output waveforms, proves that an intelligent power processing unit can learn and adapt to the nonlinear and time-varying dynamics of the plasma discharge plasma load.
Physically, the research provided significant insights into the discharge plasma’s behavior. The current-versus-voltage curves obtained during perturbation testing revealed a saturation effect, where the discharge current levels off at higher applied voltages. This phenomenon is interpreted as an ionization limit, where the plasma can no longer draw additional current regardless of the increased discharge voltage, resulting in a nonlinearly increasing impedance. These findings suggest that the HET discharge plasma is inherently constrained, which may offer a degree of self-protection against overly large accidental control perturbations, making the proposed ML control strategy safer for space applications.
The study also employed extended convergent cross mapping (eCCM) and extended Kalman filter (EKF) state estimation to further analyze the control effects. These analyses demonstrated that reducing the electrical oscillations in the discharge current leads to a commensurate reduction in plasma parameter oscillations, including ion number density. It was confirmed that the control voltage exerts a reciprocal causal relationship with the discharge current, validating the theoretical framework behind the active oscillation suppression. While the proportional gain in PID control proved effective for electrical damping, the ML-based approach, which can account for more complex, high-frequency harmonics, showed superior performance in handling the nonlinear coupling induced by the breathing mode.
Despite these promising outcomes, the researchers acknowledged several limitations inherent in this initial study. The performance of the ML controller is highly dependent on the quality and comprehensiveness of the training dataset, and the turbulence inherent in HET plasma dynamics implies that a perfect model-reality match is difficult to achieve. Furthermore, the team noted that different ignitions of the thruster could lead to slight deviations in operating point behavior, suggesting that an online, adaptive learning capability—one that updates the ESN and MLP policy during active thruster operation—will be crucial for maintaining performance over the lifetime of a mission. The current study focused primarily on electrical parameters, and future research is planned to incorporate thrust measurements and direct plasma diagnostics to further clarify the physics of the control results.
Looking forward, the integration of rad-hardened field-programmable gate arrays and dedicated machine learning hardware will be critical for transitioning these laboratory results into flight-ready technology. The potential to operate Hall thrusters at higher power points, which are currently inaccessible due to strong oscillations, represents a significant opportunity to increase thrust, specific impulse, and efficiency in future electric propulsion systems. By shifting from static, reactive control circuits to intelligent, predictive machine learning controllers, the field of electric propulsion could enable more complex and optimized space missions, effectively turning the power processing unit into a truly adaptive “brain” for the thruster system.
Facility effects present a persistent challenge in characterizing Hall effect thrusters, as ground-based vacuum chambers cannot fully replicate the extremely low pressures found in space. The physical presence of conductive chamber walls introduces unintended current paths and sputtering, both of which can alter plasma behavior and affect the thruster’s operational lifetime. These interactions complicate the already complex nature of the discharge plasma, which exhibits nonlinearity, time variation, and hysteresis. Breathing mode oscillations, driven by predator–prey dynamics between neutral gas and plasma, exacerbate these issues by producing significant density variations that degrade performance, create electromagnetic interference, and damage power processing components.
The experimental setup for system identification utilized a pair of high-bandwidth transformers to inject voltage perturbations directly into the discharge circuit. By bypassing the traditional discharge filter—which normally attenuates signals above 3 kHz—researchers gained the necessary control authority to probe the plasma’s response at higher, oscillation-relevant frequencies. This methodology allowed for a precise mapping of control voltage to discharge current. In contrast, PID-based control experiments relied on a Texas Instruments C2000 microcontroller to compute adjustments. While proportional gain proved effective for electrical damping, the integration of integral and derivative terms necessitated careful tuning to avoid increasing oscillatory behavior, highlighting the inherent limitations of static control strategies when applied to nonlinear, time-varying plasma loads.
To investigate the physical impact of these controls, the team employed Extended Kalman Filter (EKF) state estimation. Analyses revealed that suppressing electrical oscillations in the discharge current led to a measurable reduction in plasma parameter oscillations, including a 7.48% decrease in ion number density oscillations at the 400 V operating point. Further theoretical validation confirmed that control voltage shares a reciprocal causal relationship with discharge current, a finding corroborated by extended convergent cross mapping (eCCM). This physical insight confirms that active electrical control can directly influence the underlying plasma instability.
Looking beyond simple oscillation suppression, this framework supports multiple operational modes. By refining control trajectories, these systems could potentially phase-align voltage and current to increase thrust or convert current oscillations into signal modulation for communications. This foundation suggests that future power processing units will not only be more efficient but also inherently adaptive, capable of self-correcting for ignition-to-ignition performance variations through the same dynamic system identification process demonstrated in this study.
Subject of Research: Machine Learning Control for Hall Effect Thrusters
Article Title: Real-time machine learning control of a Hall thruster discharge plasma
Article References: Krishnan, A., Vlahov, B., Gibson, J., Lev, D., Theodorou, E., Walker, M., & Saeedifard, M. (2026). Real-time machine learning control of a Hall thruster discharge plasma. Journal of Electric Propulsion, 5(1), Article 53. https://doi.org/10.1007/s44205-026-00229-2
Image Credits: AI Generated
DOI: 10.1007/s44205-026-00229-2
Keywords: Hall Effect Thruster, Electric Propulsion, Machine Learning, Echo State Networks, Nonlinear Model Predictive Control, Plasma Physics, Space Exploration, Breathing Mode, Oscillation Reduction, Real-time Control, Field-Programmable Gate Array, Thruster Lifetime, Ionization Limits, Feedback Control, Active Noise Cancellation
Cite Scienmag News
Teresa Odom. (November 16, 2026). Real-Time Machine Learning Controller Stabilizes Hall Thruster Discharge Plasma Oscillations. Scienmag. https://scienmag.com/real-time-machine-learning-controller-stabilizes-hall-thruster-discharge-plasma-oscillations/
Teresa Odom. "Real-Time Machine Learning Controller Stabilizes Hall Thruster Discharge Plasma Oscillations." Scienmag, 16 November 2026, https://scienmag.com/real-time-machine-learning-controller-stabilizes-hall-thruster-discharge-plasma-oscillations/. Accessed 16 November 2026.
Teresa Odom. "Real-Time Machine Learning Controller Stabilizes Hall Thruster Discharge Plasma Oscillations." Scienmag. November 16, 2026. https://scienmag.com/real-time-machine-learning-controller-stabilizes-hall-thruster-discharge-plasma-oscillations/








