In this paper
AbstractIntroductionMotivation for electromagnetic feedback controlStability analysis and structure-informed controlMacroscopic observations for feedback controlA cancellation-dissipation structure and time-delay compensationTraining of the control networkAdjoint state method for controller trainingSemi-Lagrangian solverNumerical experimentsTwo-stream instabilityWeibel-type current-filamentation instabilityEffect of feedback-observation designLinear versus nonlinear operator architecturesFourier and Laplace transform of linearized equationReferencesWorksMathematics
Structure-Informed Electromagnetic Feedback Control of the Vlasov-Maxwell System with Delayed Observations
Abstract
Feedback control of kinetic plasma instabilities is challenging because observing the full phase-space distribution is costly, while sensing and actuation inevitably introduce delay. We develop a structure-informed electromagnetic feedback framework for the Vlasov-Maxwell system that uses only macroscopic observations to control the full kinetic distribution. We first show that magnetic actuation alone does not directly regulate radial velocity moments, motivating the use of combined electric and magnetic controls. A nonlinear perturbation estimate yields an instantaneous cancellation-dissipation feedback law. To account for delayed observations, we then use Maxwell's equations to identify the macroscopic quantities governing the resulting cancellation errors, namely currents and spatial field gradients. These quantities serve as inputs to a low-rank neural operator that learns a residual correction to the delayed cancellation baseline. Thus, while the feedback depends only on low-dimensional macroscopic observables, the control objective remains the stabilization of the full distribution toward a prescribed kinetic equilibrium. Numerical experiments for two-stream and Weibel-type instabilities demonstrate effective long-time stabilization under substantial feedback delays.
Introduction
Controlling plasma instabilities is an important challenge in applications ranging from fusion energy to charged-particle beam manipulation. External electromagnetic fields provide a means of influencing the plasma, but their effectiveness depends on the dynamics and the control objective. Magnetic confinement and macroscopic stability have long been central topics in fusion research [28]. At the kinetic level, the Vlasov–Poisson (VP) and Vlasov–Maxwell (VM) equations describe the interaction between the particle distribution and self-consistent fields. Instabilities such as the two-stream, bump-on-tail, and Weibel instabilities illustrate how small perturbations of an unstable distribution can generate growing fields and substantial changes in the plasma state. Their suppression calls for control strategies that account for this coupled evolution.
The mathematical literature on kinetic plasma control includes both controllability and optimal-control formulations. Controllability results for the VP system using localized interior source controls were established in [10] and extended to systems with prescribed external force fields in [11]. For electromagnetic actuation, a variational framework for optimal magnetic control of a VP plasma was developed in [13], with an extension to magnetic fields generated by a finite collection of fixed coils in [14]. Adjoint-based optimization has also been combined with Monte Carlo particle-in-cell discretizations for magnetic confinement [4]. Optimal control has been studied for the two-dimensional relativistic VM system, with external currents acting as controls in Maxwell’s equations [26].
For kinetic instability suppression, a PDE-constrained optimization framework was proposed in [9], in which a parameterized time-independent external electric field is chosen to maintain a desired distribution and delay the onset of VP instabilities. To reduce the computational cost of repeated kinetic solves, a moment-based formulation uses a reduced moment system during optimization, with the resulting controls assessed against the full kinetic dynamics [21]. The choice of objective is also consequential: numerical investigations of the associated optimization landscapes indicate that objectives incorporating time-integrated information can be more favorable for gradient-based optimization [12]. These studies highlight the importance of both the forward model and the objective in constructing effective controls.
A complementary approach uses linear stability analysis to identify and suppress unstable modes. Fourier–Laplace analysis of the VP system leads to electric-field controls designed to cancel unstable contributions in the linear response [8]. This perspective has been extended to a plasma subject to a uniform background magnetic field, with electric controls designed through the dispersion relation [5]. For electromagnetic kinetics, control by prescribed transverse laser fields has been studied in a reduced VM model [6]. This approach combines linear analysis for transverse instabilities with nonlinear PDE-constrained optimization for longitudinal instability suppression. In particular, the nonlinear coupling can be exploited when the corresponding longitudinal mode is unaffected by the laser control in the linearized system. The control fields in these constructions are determined from the specified initial data and subsequently applied without updating them from observations of the evolving plasma.
Feedback control provides a way to incorporate such observations into the control action. At the kinetic level, instantaneous magnetic feedback laws have been derived from short-time optimization of particle dynamics to maintain spatial confinement [1]. This approach was extended to collisional VP–BGK dynamics with uncertain parameters in [2]. Learning-based feedback has also been demonstrated in tokamak experiments: deep reinforcement learning has been used for magnetic control of plasma shape and position [7] and for avoidance of tearing instabilities [24]. These developments motivate the construction of feedback policies that can be evaluated rapidly while adapting to the observed state.
More directly related to the present work, an operator-based feedback framework for the VP system was developed in [22]. The framework uses a low-rank neural operator that maps the distribution perturbation to an external electric field. The operator is trained through an adjoint method and demonstrates effective stabilization numerically beyond the optimization horizon. An analytical cancellation-based control law guided by a perturbation energy estimate was also introduced in that work. This operator viewpoint is consistent with the broader development of neural operators for maps between function spaces, including Fourier neural operators, DeepONet, and integral-kernel architectures [19, 23, 16]. Extending the VP feedback framework to electromagnetic kinetics raises additional questions about the respective roles of electric and magnetic actuation, the information required by the controller, and the effect of feedback latency.
To be more specific, we consider the nonrelativistic VM system. In normalized variables, the uncontrolled system reads
Here is the particle distribution, and denote position and velocity, and and are the self-consistent electric and magnetic fields. The density and current are and , respectively, and is the fixed neutralizing background. We introduce external electric and magnetic fields in the Lorentz force and seek feedback laws that suppress the growth of perturbations around a prescribed spatially homogeneous equilibrium.
Two issues guide the controller design. First, electric and magnetic actuation have different effects on the kinetic state. The magnetic Lorentz force performs no work and makes no direct contribution to the evolution of radial velocity moments. Although magnetic control can influence these moments indirectly through the self-consistent fields, it does not provide a general mechanism for regulating an arbitrary kinetic-energy profile. This motivates combining electric and magnetic control channels. Second, the controller must operate with limited and delayed observations. The learned feedback controller in [22] takes the full phase-space perturbation as its input. In contrast, the feedback law developed here depends only on macroscopic observables, including current densities, electromagnetic fields, and spatial gradients of the transverse fields. Importantly, these macroscopic quantities are used only as feedback information: the control objective remains kinetic, namely, to suppress the perturbation of the full distribution and drive it toward the prescribed equilibrium . Thus, the proposed framework seeks to control a high-dimensional kinetic state using only low-dimensional macroscopic observations, without requiring direct observation or reconstruction of the full phase-space distribution during deployment. Moreover, sensing, computation, and actuation introduce latency, which can undermine the effectiveness of instantaneous field-cancellation strategies when they are implemented using delayed observations.
We address these issues through a structure-informed feedback design. The nonlinear perturbation balance provides an instantaneous cancellation–dissipation prototype. Maxwell’s equations then identify the currents and transverse field gradients that generate the cancellation errors under delayed feedback. We use these quantities as inputs to a low-rank linear neural operator that learns a residual correction to the delayed field-cancellation baseline. The deployed controller uses current and field observations, while its offline training uses kinetic simulations and an adjoint-based optimization method. This construction connects the choice of feedback inputs and the learned correction to the governing dynamics.
The main contributions are as follows:
We identify a structural limitation of purely magnetic actuation for kinetic-state stabilization. In particular, magnetic fields act through velocity-space rotations and do not directly modify radial velocity moments of order , including the kinetic energy. This shows that magnetic fields’ ability to alter particle trajectories does not, in general, provide sufficient control authority to drive the distribution toward a prescribed equilibrium, motivating the use of hybrid electromagnetic feedback.
Based on the Vlasov–Maxwell energy structure, we construct a cancellation–dissipation feedback framework that combines electric and magnetic control channels. To accommodate delayed observations, we further derive a structure-informed compensation strategy from Maxwell’s equations, using current and field-gradient observables as feedback inputs and a low-rank neural operator to learn the corresponding residual correction.
Numerical experiments demonstrate that the proposed feedback remains effective well beyond the optimization horizon and under substantial feedback delays. Comparisons with alternative observable choices further show that incorporating the governing Vlasov–Maxwell structure into the controller design is essential for long-time stabilization.
Motivation for electromagnetic feedback control
A fundamental question in plasma feedback control is the choice of actuation. Magnetic fields play a central role in plasma confinement and equilibrium shaping [28, 1]: they bend particle trajectories without doing work on individual particles. Here, our objective is to suppress the growth of kinetic perturbations around a prescribed equilibrium. This raises the question of whether purely magnetic actuation provides sufficient control authority for general kinetic-state regulation.
To answer this question, we denote the external electric and magnetic controls by and , then the controlled Vlasov equation is
We assume smooth solutions, periodic spatial boundary conditions, and sufficient decay in velocity to justify the following integrations by parts and moment calculations. Let be a stationary equilibrium of the uncontrolled system, and define
Subtracting the equilibrium equation from (2) gives
For every velocity-independent vector field ,
In particular, the phase-space transport is divergence free. Multiplying (3) by and integrating over yields the exact nonlinear perturbation balance
This identity measures the evolution of the perturbation. Electric actuation couples to the velocity gradient of the equilibrium, while magnetic actuation couples to its angular derivatives through . The latter vanishes when the equilibrium is isotropic in velocity.
The distinction between electric work and magnetic rotation is also visible in the radial velocity moments
Using the orthogonality and divergence identities above, integration by parts gives
Thus, the magnetic force makes no direct contribution to these moments, including the kinetic energy . Magnetic actuation can nevertheless affect them indirectly by changing the distribution, current, and self-consistent electric field.
A simple example illustrates the resulting limitation on general kinetic-state regulation. Consider a neutral, spatially homogeneous, isotropic initial distribution with zero initial self-consistent fields:
With , the magnetic transport vanishes because is parallel to , and the current is zero by isotropy. Hence and remain a solution for every admissible external magnetic field in the control model (2). This state cannot be driven toward a different target distribution by purely magnetic actuation.
These observations suggest that magnetic actuation alone is generally insufficient to stabilize the kinetic state. While magnetic control plays an important role in regulating electromagnetic dynamics and particle confinement, an additional mechanism is needed to modify the particle energy distribution and suppress the growth of perturbations. This motivates a combined electromagnetic feedback strategy, in which electric control directly influences the particle energy distribution, while magnetic control regulates the transverse electromagnetic dynamics. Accordingly, throughout this work, we design feedback laws for both the external electric field and the external magnetic field .
Stability analysis and structure-informed control
In this section, we develop a structure-informed feedback framework for suppressing instabilities in the Vlasov–Maxwell system. We first analyze the linearized longitudinal and transverse dynamics to identify macroscopic quantities relevant to instability growth. We then use the nonlinear perturbation balance to construct a cancellation–dissipation feedback law and examine how delayed observations introduce errors into the field cancellation. The resulting error identities guide the choice of current and field-gradient inputs for a low-rank neural operator that provides a learned correction to the delayed feedback.
For simplicity, we focus on the one-dimensional, two-velocity (1D–2V) Vlasov–Maxwell system for the controller construction, subsequent analysis, and numerical experiments. More specifically, , the spatial variable lies in a periodic interval, and . In vector identities involving cross products, we identify with and with . The self-consistent fields have the form and , and the uncontrolled system is
Here
For zero-mean longitudinal fields, one may also write with a periodic potential satisfying . The external controls are and . Their contributions to the Vlasov equation are obtained by replacing by and by in (3.1). The general perturbation equation and energy balance in Section 2 apply to this reduction with spatial integrals over the periodic interval and velocity integrals over .
Macroscopic observations for feedback control
Designing a feedback controller based on partial observations requires identifying macroscopic quantities that convey information about the coupled kinetic and electromagnetic dynamics. The feedback controller in [22] uses the full phase-space perturbation as its input. Here, we seek observations that can be used to evaluate the feedback without reconstructing the full distribution. Fourier–Laplace analysis characterizes the amplification of small perturbations, but the resulting spectral information alone does not specify which macroscopic quantities should be supplied to a feedback controller. We first recall the longitudinal response to clarify this limitation, and then integrate Maxwell’s equations to expose the dependence of the fields on macroscopic currents.
We consider perturbations around a spatially homogeneous equilibrium , with , , and . Linearizing the perturbation equation (3) and retaining first-order contributions from the control fields yields
Following [8], we use Fourier and Laplace transforms to describe the longitudinal linear response. We adopt the conventions
for the Fourier and Laplace transforms, respectively.
For a nonzero spatial Fourier mode, the wavevector is ; longitudinal and transverse refer to directions parallel and perpendicular to this vector, respectively. Longitudinal instabilities, such as the two-stream and bump-on-tail instabilities, can appear in the density perturbation and the longitudinal electric field perturbation . Applying the Fourier and Laplace transforms to (8) yields
Here denotes the free-streaming density perturbation, where satisfies the longitudinal free-streaming transport equation . The functions and are defined through
where denotes the Fourier transform with respect to . The derivation of (9) follows standard arguments and is provided in Appendix § 6.
The longitudinal response in (9) involves the function
If there exist and with such that
then, provided that the corresponding numerator in (9) does not vanish to sufficient order at , has a pole in the right half-plane. The inverse Laplace transform then contains a growing contribution proportional to . The term involving also allows the magnetic response to influence the density evolution when this coupling kernel is nonzero.
This representation describes how growth can appear in the longitudinal response and how the control fields enter that response. It does not, by itself, determine an appropriate set of feedback observations.
In particular, remains a self-consistent field in (9), and the formula does not explicitly identify the kinetic moment that drives the transverse field dynamics. Examining the poles of therefore leaves a key part of the observation design unresolved.
To make this dependence explicit, we now integrate the transverse Maxwell equations in physical space and time. This provides a direct representation of the fields in terms of the macroscopic current, which is more useful here for selecting feedback observations. The Maxwell equations retain the same form in the full nonlinear kinetic system, so the following identities are not restricted to the linearized dynamics. Introducing the characteristic variables
we rewrite the last two equations of (8) as
Integrating these equations along the characteristic curves and , respectively, gives
The transverse electric and magnetic fields can then be recovered from and , yielding
Equation (10) explicitly separates the transverse field evolution into freely propagating initial fields and contributions from the history of . It identifies the transverse current as the moment through which the kinetic distribution enters the transverse Maxwell dynamics. In particular, a growing current mode can generate corresponding growth in and . The value of this representation for feedback design is that it exposes the current–field dependence directly; it does not require a separate transverse dispersion calculation to identify as a relevant observation.
The integral representation thus supplies information about the field dynamics that the longitudinal pole criterion alone does not provide. Together with the density–electric-field relation in the longitudinal direction, it motivates the macroscopic quantities , , , , and as candidate feedback observations. These relations do not yet prescribe how the control fields should depend on the observations or establish that the observations suffice for stabilization. In the next subsection, we use the nonlinear perturbation balance to construct a feedback law and then analyze the cancellation errors caused by delayed observations. The resulting identities further identify the currents and spatial field derivatives used as inputs to the learned correction.
A cancellation-dissipation structure and time-delay compensation
Equations (3.3)–(3.4) reveal how the self-consistent electromagnetic fields are coupled to moments of the distribution function perturbation . These observations suggest that feedback control based on the macroscopic quantities , , , , and may be effective in suppressing both longitudinal and transverse instabilities. However, the appropriate functional dependence of the control fields on these quantities remains unclear. Moreover, the complex coupling between kinetic and electromagnetic dynamics makes it challenging to derive a theoretically optimal feedback control law.
For practical implementation, we instead pursue a suboptimal feedback control strategy with a simple structure. Motivated by the perturbation energy identity (2.3), derived from the original nonlinear Vlasov–Maxwell system, we consider a linear feedback control law of the form
which yields the following dissipation identity:
In particular, for sufficiently small initial perturbations, pure field cancellation (),
provides a neutral stabilization mechanism, satisfying . In practice, however, the controller may not respond instantaneously to changes in the system due to factors such as the computational time required by the numerical solver and delays in adjusting the control fields. Such delayed responses can significantly degrade the effectiveness of pure cancellation control, as demonstrated in § 5.
To account for the effects of time delay, we assume that the control applied at time is computed using information available at an earlier time (or before), where denotes the time lag. The resulting cancellation errors can be expressed as
Motivated by these identities, we introduce operator-based dissipative corrections, , , and , that utilize the observations , , , and to compensate for delay-induced cancellation errors. Specifically, we propose the following cancellation-residual feedback control structure:
Here, are neural operators that map the observed quantities to the corresponding control corrections. Possible architectures include the Fourier Neural Operator (FNO) [19, 15, 18], Low-Rank Neural Operator (LNO) [16], Graph Neural Operator (GNO) [3, 20], and DeepONet [23, 25, 17]. Although selecting an optimal operator architecture is nontrivial, motivated by the linear dissipation structure of the feedback law (11), we propose a linear neural operator with a low-rank structure:
Here, forms an basis on the computational domain, which can either be learned or chosen analytically (e.g., a trigonometric basis under periodic boundary conditions). The kernels are parameterized by neural networks. The structure in (14) resembles a DeepONet-type architecture with a linear branch network, where the branch coefficients are computed through kernel integrals rather than a conventional branch network, allowing greater flexibility in handling different computational meshes.
In summary, the proposed feedback controller (13) integrates three key components:
Instability-informed macroscopic observations identified through the linear stability analysis;
A field-cancellation baseline motivated by the energy structure of the Vlasov–Maxwell system;
Delay-informed current and field-gradient observations, together with learned residual corrections motivated by Maxwell’s equations.
Training of the control network
In this section, we describe the training procedure for the neural feedback controller. The network parameters are optimized using the adjoint-state method within an optimize-then-discretize framework. We also introduce a semi-Lagrangian discretization of the kinetic equations to enable efficient long-time simulations.
Adjoint state method for controller training
Training the feedback operator (13) requires solving the following PDE-constrained optimization problem over the time interval :
To obtain an effective control operator, the optimization time horizon must be sufficiently long to capture the growth of instabilities while remaining computationally manageable. Moreover, the controller should be exposed to diverse training trajectories to promote its effectiveness beyond the training interval . To this end, we incorporate a variety of initial perturbations during training, as described in § 5.
We solve the optimization problem (4.1) using the adjoint-state method. To this end, we introduce the Lagrangian
When satisfies the constraint (16), the constraint term vanishes, so for arbitrary and , and consequently . The resulting gradient can then be used in standard optimization algorithms, such as SGD or Adam, to update the control parameters.
Applying integration by parts yields
If satisfies the adjoint equation
the gradient of the objective function simplifies to
Substituting the feedback operator (13) into the above expression, we obtain
where, for , the operator terms can be written explicitly as
and
In general, the state sensitivities , , , , and are not available in closed form, and computing them numerically through automatic differentiation can be expensive. To reduce the computational cost, we treat the state-dependent observations as fixed during each gradient evaluation, leading to the approximate gradient
We emphasize that the approximation (19) leads to a sequential optimization procedure: the state trajectory is held fixed during each gradient evaluation and recomputed after each parameter update. The exact closed-loop gradient (18) contains two contributions. The first arises from the explicit dependence of the learned control on its parameters along the current trajectory, while the second accounts for changes in the feedback observations induced by variations in the state trajectory. The approximation (19) retains only the first contribution during each gradient evaluation. Rather than explicitly differentiating the second contribution, its effects are incorporated indirectly through the updated trajectories in subsequent iterations.
This approach avoids differentiating through the entire closed-loop dynamics while retaining the direct sensitivity of the objective function to the learned control. It is particularly suitable for the proposed cancellation-residual formulation, in which the learned dissipation operator provides a moderate correction to a prescribed field-cancellation baseline. Although the approximation sacrifices the exact closed-loop gradient in favor of computational efficiency, our numerical results indicate that the sequential optimization procedure produces stable parameter updates when the initial perturbation is sufficiently small and an appropriate optimization step size is used. Its performance will be further evaluated through numerical experiments in Section 5.
Semi-Lagrangian solver
Solving the PDE-constrained optimization problem (4.1) requires repeatedly solving high-dimensional kinetic equations. To relax the time-step restrictions and improve the efficiency of long-time simulations, we employ a semi-Lagrangian scheme for the kinetic equation, coupled with a leapfrog-type update for the electromagnetic fields. Let , , , and denote the numerical solutions at time . We assume that is represented on the space-velocity mesh , and on the spatial mesh , and on the staggered mesh , where . The numerical solution is advanced from to through the following steps:
First, compute .
Compute from and . Then update on the primal mesh and on the staggered mesh according to
Next, obtain the nodal values of on the primal mesh by averaging, , and use these values to compute
Finally, compute . Obtain from and complete the updates of and using
The updates of and can be expressed in the following leapfrog form:
which recovers the classical Yee scheme [27]. The scheme (4.5) provides a second-order time discretization of the semi-discrete system
An important property of the semi-discrete scheme (4.6) is that it satisfies a discrete energy balance. To see this, define the transverse electromagnetic energy by
Differentiating with respect to time and substituting (4.6), we obtain
Using the discrete summation-by-parts identity
we see that the last two terms on the right-hand side of (22) cancel. Consequently,
In the vacuum case, where , the energy is conserved. However, this energy conservation property does not generally carry over to the fully discrete scheme (20). To achieve energy conservation at the fully discrete level, implicit time integrators, such as the Crank–Nicolson scheme, may be employed.
The adjoint equation (17) can be integrated backward in time using a similar semi-Lagrangian procedure:
First, compute .
Next, compute
Finally, compute .
The overall PDE-constrained optimization procedure is summarized in Algorithm 1.
Algorithm 1 PDE-constrained optimization for VM control
Specify the terminal time and the maximum number of optimization iterations .
Initialize the control parameters .
for do
Solve the controlled VM system (16) forward in time over using the current control laws , , and .
Solve the adjoint equation (17) backward in time to obtain .
Evaluate the approximate gradient using (19).
Update the control parameters , , and .
end for
Numerical experiments
We evaluate the performance of the electromagnetic feedback control (13) through numerical experiments. In all test cases, periodic boundary conditions are imposed at both endpoints of the spatial interval. To construct the dissipation operator (14), the integral kernels
are represented by a fully connected multi-output neural network with architecture 1-32-32-. The output of each hidden layer is activated by the ReLU function before being passed to the next layer. The controllers are optimized with respect to the running loss
following the procedure described in § 4.
To stabilize the optimization during the early iterations, we initialize the parameters in the output layer of to zero, so that the learned dissipation initially vanishes. The network parameters are then updated using the approximate gradient (19) for 1000 iterations. We first perform 200 Adagrad iterations as a preprocessing stage, followed by Adam iterations for further acceleration.
To expose the controller to a richer family of off-equilibrium states and improve generalization beyond the optimization horizon , we add an artificial perturbation to the initial data and compute the gradient along the resulting perturbed trajectory at each iteration. The perturbation takes the form
Here , , and are orthonormal bases in the , , and directions, respectively. In our computations, we use the normalized trigonometric functions on ,
and the orthonormal Hermite basis functions on ,
The Hermite polynomials are generated recursively by
We set the perturbation magnitude to .
The full tensor-product family contains basis functions. To reduce the sampling dimension, we randomly select basis functions from this family and sample their coefficient vector uniformly from the -dimensional unit ball.
During training, we use the future loss as a measure of long-horizon generalization and select the best-performing iteration accordingly.
Two-stream instability
We first consider the two-stream equilibrium
with . The system is initialized with
The two-stream equilibrium is unstable, so a small initial perturbation is amplified as the system evolves. Our goal is to stabilize the full kinetic distribution around the prescribed equilibrium, with particular attention to the effect of feedback delay.
We take the reference feedback delay to be and optimize the controller over the time interval . The computational domain is
The learning rate is set to during the Adagrad preprocessing stage and to during the subsequent Adam iterations. Figure 5.1 compares the uncontrolled solution, delayed cancellation control,

Figure 5.1. Two-stream instability. Histories of the perturbation and electromagnetic energy over . No external field (red) vs. delayed cancellation control (blue) vs. cancellation–dissipation control (green). The feedback delay is . The vertical black line marks the terminal time of optimization.
and the cancellation–dissipation feedback control (13). Without control, both the kinetic perturbation
and the electromagnetic energy
grow rapidly. For instantaneous observations, the pure cancellation law , , and keeps the perturbation norm constant, as predicted by (4). Its effectiveness, however, deteriorates when the cancellation is evaluated using delayed observations. In comparison, the learned cancellation–dissipation controller substantially improves the suppression of the instability and remains effective well beyond the optimization horizon.
In applications, the actual feedback delay may differ from the reference value used during training. To assess the robustness of the controller to such mismatch, we optimize (13) at and then apply the learned controller at different values of . As shown in Figure 5.2, when , the learned dissipation underestimates the delay-induced error and therefore provides insufficient compensation. The more delicate case occurs when , for which the learned dissipation can overcompensate and degrade stability. A simple correction is to rescale the output of the dissipation networks by , without retraining. This scaling is consistent with the first-order approximation of the time-integrated cancellation error in (12), which is proportional to the delay. Figure 5.3 shows that this correction is effective for both and .
Figure 5.2. Two-stream instability. History of the perturbation under varying feedback delays . No external field (red) vs. delayed cancellation control (blue) vs. cancellation–dissipation control (green). The dissipation networks (3.8) are trained at and then applied directly at different values of without correction. The vertical black line marks the terminal time of optimization.

Figure 5.3. (a) (b) (c) (d) . Two-stream instability. History of the perturbation under varying feedback delays . No external field (red) vs. delayed cancellation control (blue) vs. cancellation–dissipation control (green). The dissipation networks (14) are trained at , and their outputs are rescaled by in each test case.
We further display the marginal distribution
in Figure 5.4 for the larger delay . Delayed cancellation alone provides noticeably weaker suppression of the kinetic deformation, whereas the operator-based controller better preserves the two-stream structure over long times. Together, these results indicate that the learned correction can compensate for a non-negligible feedback delay and that moderate delay mismatch can be handled by a simple rescaling of the learned dissipation.

Figure 5.4. (a) , no control; (b) , no control; (c) , no control; (d) , delayed cancellation; (e) , delayed cancellation; (f) , delayed cancellation; (g) , operator control; (h) , operator control; (i) , operator control. Figure 5.4: Two-stream instability. Evolution of the marginal distribution for feedback delay . No external field (first row) vs. delayed cancellation (second row) vs. cancellation–dissipation control (third row). The dissipation networks are trained at over , consistent with Figures 5.1–5.3, and their outputs are rescaled by at selected times. Delayed field cancellation suppresses the transverse field growth over a moderate time interval but gradually loses effectiveness. In contrast, the cancellation–dissipation controller continues to suppress the transverse perturbation well beyond the optimization horizon.
Weibel-type current-filamentation instability
We next consider the equilibrium
with and . This distribution consists of two equal counterstreaming populations in the direction. Since the spatial variation is along , perpendicular to the streaming direction, the resulting transverse electromagnetic instability is more precisely described as a current-filamentation instability, closely related to the Weibel instability. We therefore refer to this test as a Weibel-type current-filamentation instability.
The computational domain is
The learning rate is set to during the Adagrad preprocessing stage and to during the subsequent Adam iterations. The instability is seeded by the small initial magnetic perturbation
The anisotropic counterstreaming distribution responds to this magnetic perturbation by developing transverse current structures, which reinforce the magnetic perturbation and lead to rapid growth of the kinetic deviation, as shown in Figure 5.5.

Figure 5.5. Weibel-type current-filamentation instability. Histories of the perturbation and electromagnetic energy over . No external field (red) vs. delayed cancellation control (blue) vs. cancellation–dissipation control (green). The feedback delay is . The vertical black line marks the terminal time of optimization.
The feedback controller (3.7) is trained with reference delay over . As in the two-stream test, the learned controller substantially improves upon delayed field cancellation. Figure 5.6 shows that when , the learned dissipation is too weak to fully compensate for the larger delay, whereas can lead to overcompensation. In both cases, rescaling the learned dissipation by substantially reduces the effect of delay mismatch.

Figure 5.6. Weibel-type current-filamentation instability. History of the perturbation under varying feedback delays . No external field (red) vs. delayed cancellation control (blue) vs. cancellation–dissipation control (green). The dissipation networks (3.8) are trained at . In the first row, the learned dissipation is applied directly at without correction. In the second row, the network outputs are rescaled by . The vertical black line marks the terminal time of optimization.
Figure 5.7 shows the spatial profiles of the local transverse electromagnetic energy density

Figure 5.7. Weibel-type current-filamentation instability. Evolution of the transverse electromagnetic energy density for feedback delay . No external field (red) vs. delayed cancellation control (blue) vs. operator-based feedback control (green). The dissipation networks (14) are trained at , with their outputs rescaled by .
Effect of feedback-observation design
To assess the role of observation design in long-time stabilization and delay compensation, we compare the proposed gradient-based feedback variables
with the component-based alternative
The component-based choice provides a natural baseline because the electromagnetic field components directly characterize the instantaneous field state. However, the delay-compensation identities (12) show more specifically that the cancellation errors are generated by , , , and . These quantities therefore encode the instantaneous field evolution responsible for the delay-induced residual and are expected to provide more informative feedback variables. In the following comparison, all network architectures and training settings are kept unchanged; only the last two feedback variables are replaced.
Figure 5.8 presents the comparison for the two-stream and Weibel-type current-filamentation instabilities. In both cases, the component-based controller provides substantial short-time suppression but degrades significantly beyond the optimization horizon. In contrast, the gradient-based controller maintains substantially stronger long-time stabilization of both the kinetic perturbation and the electromagnetic energy. The consistency of this behavior across the two instability mechanisms supports the observation design derived directly from the Vlasov–Maxwell dynamics.

(a)

(b)

(c)

(d)
Figure 5.8. Comparison of gradient-based and component-based feedback for the two-stream instability (first row) and the Weibel-type current-filamentation instability (second row). Histories of the perturbation and electromagnetic energy over for the two-stream instability and for the Weibel-type instability. No external field (red) vs. delayed cancellation control (blue) vs. gradient-based delay compensation (green) vs. component-based delay compensation (pink). Both learned controllers use the operator structure (13)–(14); the component-based controller replaces and by and , respectively. The feedback delay is .
We note that spatial differentiation can amplify measurement noise, and its practical implementation depends on sensor resolution and the procedure used to reconstruct derivatives. Incorporating noisy observations and regularized derivative estimation into the training procedure is left for future work.
Linear versus nonlinear operator architectures
The main experiments above use dissipation networks with a linear operator structure. We now examine whether introducing nonlinearity into the operator architecture changes the control performance. As a representative nonlinear architecture, we use a Fourier neural operator (FNO) with two Fourier layers, 15 retained modes, and width 9 in each intermediate layer. The number of retained modes is consistent with the spectral resolution used in the preceding LNO experiments, and the biases are set to zero. With this configuration, the FNO- and LNO-based controllers have comparable numbers of trainable parameters.
We replace the linear dissipation operator in (14) by the FNO while keeping the remainder of the training procedure unchanged. To account for the different optimization behavior of the nonlinear architecture, the learning rates in the Adagrad and Adam stages are set to and , respectively.
At the reference delay, Figure 5.9 and Figure 5.10 show that the LNO and FNO provide comparable stabilization. The difference becomes more pronounced when the actual delay differs from the value used during training; see Figure 5.11 and Figure 5.12. In particular, the FNO is less sensitive in the tested cases when the actual delay is smaller than the reference value, , and avoids the pronounced overcompensation observed for the LNO. The precise source of this difference is not yet clear and may reflect both architectural and optimization effects.

Figure 5.9. Comparison of LNO- and FNO-based controllers for the two-stream instability. Histories of the perturbation and electromagnetic energy over . No external field (red) vs. LNO-based control (green) vs. FNO-based control (blue). Both controllers are trained at over and show comparable performance at the reference delay.

Figure 5.10. Comparison of LNO- and FNO-based controllers for the Weibel-type current-filamentation instability. Histories of the perturbation and electromagnetic energy over . No external field (red) vs. LNO-based control (green) vs. FNO-based control (blue). Both controllers are trained at over .

(a) (a)

(b) (b)
Figure 5.11. Comparison of LNO- and FNO-based controllers for the two-stream instability under varying feedback delays . Histories of the perturbation for no external field (red), LNO-based control (green), FNO-based control (blue), corrected LNO-based control (green dashed), and corrected FNO-based control (blue dashed). The FNO is more robust for , whereas the corrected LNO gives slightly stronger suppression in this test.

(a) (a)

(b) (b)
Figure 5.12. Comparison of LNO- and FNO-based controllers for the Weibel-type current-filamentation instability under varying feedback delays . Histories of the perturbation for no external field (red), LNO-based control (green), FNO-based control (blue), corrected LNO-based control (green dashed), and corrected FNO-based control (blue dashed). In these tests, the FNO is less sensitive to delay mismatch.
The delay-informed rescaling by improves both architectures. After this correction, the relative performance is problem-dependent: the LNO gives slightly stronger suppression in the two-stream test, whereas the FNO remains moderately more effective in the Weibel-type current-filamentation test. These results suggest a trade-off between the structural simplicity of the LNO and the greater robustness to delay mismatch observed for the FNO in the present experiments. The LNO has a more interpretable structure and, through a flexible choice of basis functions, can be adapted naturally to prescribed actuation modes and nonuniform spatial discretizations. Without the delay-rescaling correction, however, it is more sensitive to mismatch between the actual delay and the value used during training. The standard FNO is less sensitive to such mismatch in the present tests, although its Fourier representation is most natural on regular grids.
Fourier and Laplace transform of linearized equation
In this section, we derive the relation (9). We denote
Performing Fourier transform in in the linearized equation (8) yields
Note that , integrating the above equation with respect to eliminates the last two terms on the right,
Integration in time leads to
Further integrating the above equation with respect to , we notice that
and
with being the free-stream density perturbation. Thus,
where and are defined through Fourier transforms,
Finally, performing Laplace transform in time in (A.1) yields
The second line is obtained by using
Then the desired relation (9) is obtained by rearranging (A.2).
Acknowledgment: We would like to thank Prof. Lukas Einkemmer for insightful discussions on the 1D2V model and linear stability analysis.
References
References
- [1]Giacomo Albi, Giacomo Dimarco, Federica Ferrarese, and Lorenzo Pareschi. Instantaneous control strategies for magnetically confined fusion plasma. Journal of Computational Physics, 527:113804, 2025.arxiv.org/abs/2403.16254
- [2]Giacomo Albi, Giacomo Dimarco, Federica Ferrarese, and Lorenzo Pareschi. Robust feedback control of collisional plasma dynamics in presence of uncertainties. Journal of Computational Physics, 546:114512, 2026.arxiv.org/abs/2505.19992
- [3]Anima Anandkumar, Kamyar Aziz zad enesh eli, Kaushik Bhattacharya, Nikola Kovachki, Zongyi Li, Burigede Liu, and Andrew Stuart. Neural operator: Graph kernel network for partial differential equations. In ICLR 2020 workshop on integration of deep neural models and differential equations, 2020.arxiv.org/abs/2003.03485
- [4]Jan Bartsch, Patrik Knopf, Stefania Scheurer, and Jörg Weber. Controlling a Vlasov–Poisson plasma by a particle-in-cell method based on a Monte Carlo framework. SIAM Journal on Control and Optimization, 62(4):1977–2011, 2024.DOI
- [5]Peiyi Chen, Rogerio Jorge, Qin Li, and Yukun Yue. Control of a uniformly magnetized plasma with external electric fields. Journal of Statistical Physics, 193(4):40, 2026.arxiv.org/abs/2509.07988
- [6]Nicolas Crouseilles, Lukas Einkemmer, Qin Li, Uri Shumlak, and Yukun Yue. Control of kinetic plasma instabilities by laser fields. arXiv preprint arXiv:2509.23080, 2025.arxiv.org/abs/2509.23080
- [7]Jonas Degrave, Federico Felici, Jonas Buchli, Michael Neunert, Brendan Tracey, Francesco Carpanese, Timo Ewalds, Roland Hafner, Abbas Abdolmaleki, Diego de las Casas, Craig Donner, Leslie Fritz, Cristian Galperti, Andrea Huber, James Keeling, Maria Tsimpoukelli, Jackie Kay, Antoine Merle, Jean-Marc Moret, Seb Noury, Federico Pesamosca, David Pfau, Olivier Sauter, Cristian Sommariva, Stefano Coda, Basil Duval, Ambrogio Fasoli, Pushmeet Kohli, Koray Kavukcuoglu, Demis Hassabis, and Martin Riedmiller. Magnetic control of tokamak plasma through deep reinforcement learning. Nature, 602(7897):414–419, 2022.DOI
- [8]Lukas Einkemmer, Qin Li, Clément Mouhot, and Yukun Yue. Control of instability in a Vlasov–Poisson system through an external electric field. arXiv preprint arXiv:2407.15008, 2024.arxiv.org/abs/2407.15008
- [9]Lukas Einkemmer, Qin Li, Li Wang, and Yunan Yang. Suppressing instability in a Vlasov–Poisson system by an external electric field through constrained optimization. Journal of Computational Physics, 498:112662, 2024.arxiv.org/abs/2305.17994
- [10]Olivier Glass. On the controllability of the Vlasov–Poisson system. Journal of Differential Equations, 195(2):332–379, 2003.DOI
- [11]Olivier Glass and Daniel Han-Kwan. On the controllability of the Vlasov–Poisson system in the presence of external force fields. Journal of Differential Equations, 252(10):5453–5491, 2012.
- [12]Martin Guerra, Qin Li, Yukun Yue, and Leonardo Zepeda-Núñez. What metric to optimize for suppressing instability in a Vlasov–Poisson system? Journal of Computational Physics, 564:115169, 2026.arxiv.org/abs/2504.10435
- [13]Patrik Knopf. Optimal control of a Vlasov–Poisson plasma by an external magnetic field. Calculus of Variations and Partial Differential Equations, 57(5):134, 2018.
- [14]Patrik Knopf and Jörg Weber. Optimal control of a Vlasov–Poisson plasma by fixed magnetic field coils. Applied Mathematics & Optimization, 81(3):961–988, 2020.
- [15]Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. On universal approximation and error bounds for fourier neural operators. Journal of Machine Learning Research, 22(290):1–76, 2021.arxiv.org/abs/2107.07562
- [16]Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to PDEs. Journal of Machine Learning Research, 24(89):1–97, 2023.
- [17]Samuel Lanthaler, Siddhartha Mishra, and George E Karniadakis. Error estimates for deeponets: A deep learning framework in infinite dimensions. Transactions of Mathematics and Its Applications, 6(1):tnac001, 2022.arxiv.org/abs/2102.09618
- [18]Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for pdes on general geometries. Journal of Machine Learning Research, 24(388):1–26, 2023.DOI
- [19]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020.arxiv.org/abs/2010.08895
- [20]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Andrew Stuart, Kaushik Bhattacharya, and Anima Anandkumar. Multipole graph neural operator for parametric partial differential equations. Advances in Neural Information Processing Systems, 33:6755–6766, 2020.arxiv.org/abs/2006.09535
- [21]Jingcheng Lu, Li Wang, and Jeff Calder. Controlling instability in the Vlasov–Poisson system through moment-based optimization. arXiv preprint arXiv:2508.18412, 2025.arxiv.org/abs/2508.18412
- [22]Jingcheng Lu, Li Wang, and Jeff Calder. Dynamical feedback control with operator learning for the Vlasov–Poisson system. arXiv preprint arXiv:2509.23063, 2025.arxiv.org/abs/2509.23063
- [23]Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218–229, 2021.
- [24]Jaemin Seo, SangKyeun Kim, Azarakhsh Jalalvand, Rory Conlin, Andrew Rothstein, Joseph Abbate, Keith Erickson, Josiah Wai, Ricardo Shouba, and Egemem Krolemen. Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature, 626(8000):746–751, 2024.DOI
- [25]Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of parametric partial differential equations with physics-informed deeponets. Science advances, 7(40):eabi8605, 2021.
- [26]Jörg Weber. Optimal control of the two-dimensional Vlasov–Maxwell system. ESAIM: Control, Optimisation and Calculus of Variations, 27:S19, 2021.DOI
- [27]Kane Yee. Numerical solution of initial boundary value problems involving maxwell’s equations in isotropic media. IEEE Transactions on antennas and propagation, 14(3):302–307, 1966.DOI
- [28]Hartmut Zohm. Magnetohydrodynamic stability of tokamaks. John Wiley & Sons, 2014.
