WorksMathematics
Full-horizon magnetic control of the Vlasov–Poisson system for plasma confinement
Abstract
We study the full-horizon control of a two-dimensional Vlasov–Poisson system through an externally imposed magnetic field. The control is represented by time-dependent coefficients associated with smooth, spatially localized basis functions, and is designed to steer selected position and velocity moments of the distribution function while limiting particle dispersion and the control effort. Starting from the kinetic constrained optimization problem, we derive the continuous first-order optimality conditions, including a backward Vlasov equation, an elliptic adjoint equation, and the variational inequality associated with bounded controls. We then project this system onto a particle–grid representation and identify the resulting particle adjoints and control gradient. The full-horizon field is then computed on a reduced particle ensemble, initialized by an instantaneous control, and subsequently applied to a larger set of particles. Numerical experiments indicate that the instantaneous control field already provides an effective baseline, while the full-horizon correction can further reduce the relative thermal energy near the boundaries and improve confinement over the prescribed time interval.
Introduction
Plasma dynamics is governed by the collective motion of charged particles and by their interaction with both externally imposed and self-consistent electromagnetic fields [6, 35, 50]. A particularly important application is magnetic confinement fusion, whose objective is to confine a high-temperature plasma for sufficiently long times to sustain thermonuclear reactions while limiting energy and particle losses [35, 50, 36]. Depending on the physical regime, plasmas may be thermal or non-thermal and may exhibit phenomena spanning a wide range of spatial and temporal scales. This multiscale character has led to different levels of description, including fluid, kinetic, and hybrid models [19, 32, 25, 8, 24]. Fluid descriptions, such as magnetohydrodynamics, provide an effective representation of the macroscopic evolution of the plasma and play a central role in the analysis of equilibrium and large-scale stability in magnetic-fusion devices [46, 7]. However, several processes relevant to tokamaks and stellarators, including non-equilibrium transport, energetic-particle dynamics, velocity-space anisotropies, and losses near the plasma boundary, require a kinetic description. Kinetic models describe the evolution of particle distribution functions in phase space and can therefore represent effects that cannot be adequately captured through a finite set of macroscopic quantities [15, 45, 16, 18]. Among kinetic descriptions, the Vlasov equation constitutes a fundamental model for collisionless plasmas. It describes the transport of the particle distribution under the action of electromagnetic forces, generally coupled to the distribution itself through Maxwell’s or Poisson’s equations. The high dimensionality of phase space, together with the nonlinear coupling between particles and fields, makes the numerical approximation of Vlasov-type systems particularly demanding. These challenges have motivated the development of a broad range of numerical techniques, including semi-Lagrangian, conservative, multiscale, and asymptotic-preserving methods [12, 17, 20, 23, 48, 33, 24, 25, 30, 31, 34, 52, 47, 51]. In this context, particle-in-cell (PIC) methods are especially attractive for high-dimensional problems. They approximate the distribution function through a finite ensemble of numerical particles, whose trajectories are advanced in a Lagrangian fashion, while the electromagnetic fields are computed on a spatial grid [10, 39, 38, 22]. This combination considerably reduces the computational cost associated with a direct discretization of the full phase space while retaining the essential kinetic features of the model.
The control of plasma dynamics is a central issue in magnetic confinement fusion. In both tokamaks and stellarators, externally generated magnetic fields can be used to modify particle trajectories, improve confinement, reduce cross-field transport, and limit the flux of energetic particles toward boundaries [21, 28, 36, 37]. The design of effective controls is nevertheless challenging because the magnetic field acts on an evolving particle distribution and must account for the nonlinear coupling between the plasma and its self-consistent fields. Furthermore, a successful strategy should not only reduce instantaneous losses but also anticipate the subsequent evolution of the plasma and preserve confinement over relevant time scales. Both the analytical foundations and the numerical realization of kinetic plasma control have therefore attracted increasing attention in recent years [9, 5, 42, 49, 27, 26, 11, 44, 43, 29]. In particular, instantaneous magnetic control strategies for deterministic and uncertain Vlasov–Poisson models were developed in [2, 3]. These approaches replace optimization over the entire time horizon with a sequence of local-in-time control problems. As a result, they are computationally inexpensive and can be readily coupled with kinetic solvers. Their local character, however, also represents an important limitation: the control is determined using only information available at the current time and does not explicitly account for the future evolution of the distribution function, the self-consistent field, or the resulting plasma position. This lack of predictive capability may become particularly relevant in magnetic confinement applications, where the plasma response to external actuation may be delayed and the primary objective is to maintain confinement and prevent particle losses over extended time intervals.
Motivated by the limitations of local in time control strategies, in this work we consider a deterministic, collisionless kinetic model and seek an optimal magnetic control over a prescribed time horizon. We describe the evolution of a population of negatively charged particles, identified with electrons, while the ions are treated as a fixed neutralizing background.
In dimensionless variables, the governing Vlasov–Poisson system reads
supplemented with the initial condition . The electron density is obtained as the velocity marginal of the distribution function
and the ion density is assumed to be homogeneous and stationary. Under this assumption, the net charge density on the right-hand side of the Poisson equation reduces to , where is the spatial mean of the electron density, which is conserved in time by the Vlasov equation and plays the role of the neutralizing ion background. With the sign convention adopted here, the effective Lorentz force acting on the electrons is given by
where denotes the externally applied magnetic field, which plays the role of the control variable, while is the self-consistent electric field determined through the Poisson equation in (1).
Our starting point is the instantaneous magnetic control framework introduced in [2, 3]. In the deterministic setting of [2], the control was obtained through a sequence of local in time optimization problems and represented by a spatially piecewise constant, time-dependent magnetic field. This spatial localization reflects the physical constraint that arbitrarily localized magnetic configurations cannot be generated in practice. The approach was subsequently extended in [3] to systems affected by uncertainty and collisions. Although these strategies are computationally efficient, their instantaneous nature prevents them from explicitly anticipating the future evolution of the plasma.
The aim of the present work is consequently to move beyond this instantaneous formulation and determine an open loop optimal magnetic control over the entire time interval . Full-horizon magnetic control of the Vlasov–Poisson system has been studied analytically in [42], for fields generated by fixed coils, and numerically in [5], where state and adjoint equations are solved by a particle-in-cell method. With respect to these works, we consider moment-based tracking and dispersion objectives in a 2D–2V geometry with reflecting walls, derive a particle–grid optimality system consistent with the discretize-then-optimize approach, initialize the optimization with the instantaneous control of [2], thus allowing a direct comparison between the two strategies, and compute the control on a reduced particle ensemble before validating it on a larger one.
To this end, we study the associated first-order optimality conditions through a coupled forward–backward system [44, 5, 42]. Accounting for the full time horizon allows the control to anticipate the subsequent evolution of the particle distribution and to balance immediate confinement with longer-term objectives. This predictive capability comes however at a substantially higher computational cost, since the state equation must be coupled with a backward adjoint equation and an optimality condition for the control. More precisely, we consider the constrained optimization problem
Here, denotes the set of admissible magnetic controls. At this stage, we introduce the cost functional in the general form
where . The running cost measures the deviation of selected position- and velocity-space observables from their prescribed targets and controls their dispersion. The regularization term penalizes the control effort, whereas promotes the desired plasma configuration at the final time . The precise definitions of these terms are introduced in § 2, after specifying the spatial parameterization of the external magnetic field. In the following, the first-order optimality conditions are derived through a Lagrangian (adjoint) approach. Solving the complete optimality system directly at the kinetic scale would, however, be prohibitively expensive. We therefore synthesize the control using a reduced particle representation of the plasma dynamics. The reduced system is employed to compute the open-loop optimal field, which is subsequently reconstructed and applied to a larger particle simulation. In this way, the expensive optimization is performed on a representative particle ensemble, while the effectiveness and transferability of the resulting control are assessed on the higher-dimensional dynamics.
The remainder of the paper is organized as follows. § 2 introduces the model and the optimal control objective. § 3 derives the continuous optimality system and its particle–grid representation. § 4 describes the numerical approximation and the optimization method. § 5 presents the computational experiments. A final section summarizes the main conclusions and discusses possible future developments. The appendices provide the discretize-then-optimize derivation and the instantaneous control strategy employed to initialize the full-horizon optimization.
Problem setting
We consider a reduced setting, with particle positions and velocities . The dynamics is studied in the plane orthogonal to the dominant component of the externally applied magnetic field. To obtain a physically and computationally realizable control, we partition the spatial domain into cells, with
Denoting the cells by , , the partition satisfies
We then associate with each cell a time-dependent control coefficient and introduce the control vector
For notational convenience, we set
where denotes the center of . The scalar component of the external magnetic field is reconstructed from the control coefficients as
Each kernel is assumed to be nonnegative and provides a smooth approximation of the indicator function of the corresponding cell. More precisely, it is approximately constant inside and smoothly decays to zero outside it. Consequently, the resulting magnetic field is nearly piecewise constant over the partition while remaining globally smooth across cell interfaces. The corresponding three-dimensional external magnetic field is defined as
The scalar component is subject to the pointwise amplitude constraint
where denotes the maximum admissible magnetic-field intensity. With this convention, the planar component of the magnetic force is given by
The precise form of the kernel, the choice of the number of control cells, and a sensitivity analysis with respect to this number will be discussed in the numerical experiments section. The corresponding two-dimensional Vlasov–Poisson system reads
Here, is the prescribed initial distribution, whereas denotes the electron distribution function in phase space. The quantity
is conserved in time by the Vlasov equation and plays the role of the neutralizing ion background. In both equations, periodic boundary conditions are imposed in the -direction, while reflective boundary conditions are prescribed in the -direction to model the presence of material walls. The control objective is to steer the plasma by means of the external magnetic field and, in particular, to reduce particle transport toward and interaction with these walls. We now reformulate the continuous optimal control problem (3) in terms of the finite set of control coefficients. Let
The set of admissible controls is defined as
The reduced optimal control problem is therefore
We first specify the regularization of the control coefficients. The regularization of the reconstructed magnetic field would naturally lead to
Since the kernels are smooth approximations of disjoint indicator functions, they satisfy
except within the narrow transition regions near cell interfaces. Motivated by this localization property, we neglect the overlap terms and adopt the diagonal approximation
This approximation penalizes the control coefficients independently and avoids coupling between controls associated with different cells. It will be used throughout the remainder of the paper. The cost functional is defined by
where the running cost is
with . The first contribution measures the deviation from the prescribed target configuration, whereas the second penalizes the magnitude of the localized control coefficients. For each , the tracking term is defined as
where . The functions represent the selected physical- or velocity-space observables, while denotes the corresponding prescribed target. The first term in (15) promotes the concentration of the distribution around the target value . The second term penalizes the dispersion of the observable around its instantaneous mean. More precisely, the mean of the observable is defined by
Finally, the terminal cost is defined as
It has the same structure as the running cost, but is evaluated at the final time . Its purpose is to promote the desired plasma configuration at the end of the control horizon.
Derivation of the optimality system
The aim of this section is to introduce and compare two formally equivalent strategies for solving the minimization problem (9). We first derive the continuous optimality system through a Lagrangian formulation and then project the resulting forward-backward equations onto a particle approximation of the distribution function. This procedure yields a particle-level optimality system involving the characteristic variables, their corresponding adjoint variables, and the gradient with respect to the control coefficients. We subsequently relate this construction to the alternative procedure in which the particle-grid approximation is introduced before deriving the optimality conditions. When the same kernel interpolation, quadrature rules, and particle discretization are employed, the two approaches formally lead to the same particle-level optimality system, up to the errors introduced by the discretization.
Continuous first-order optimality conditions
We derive here the continuous first-order optimality conditions associated with the constrained minimization problem (9) through a Lagrangian approach. Let and denote the adjoint variables associated with the Vlasov equation and the Poisson equation in (7), respectively. For a given control vector , we recall the notation
To distinguish the Lagrangian functional from the running cost introduced in (14), we denote the former by . It is defined as
where . The running cost , the tracking terms , the diagonal control regularization, and the terminal cost are defined in (14), (15), (12), and (17), respectively. In deriving the optimality conditions, we assume that all functions are sufficiently regular and that the boundary terms arising from integration by parts vanish under the prescribed state boundary conditions and the corresponding compatible adjoint boundary conditions. Let denote the optimal state-adjoint-control. Stationarity, together with the variational optimality condition for the constrained control, gives
where all derivatives are evaluated at the optimal value. If lies in the interior of , the last condition reduces to . The first two conditions in (19) recover the forward
Vlasov–Poisson system
where, for brevity, the dependence on has been omitted and the forward system is supplemented with the initial condition . The variations with respect to and yield the backward system
where the source term is the derivative of the running cost with respect to . It is given by
The backward Vlasov equation is supplemented with a terminal condition obtained from
Therefore,
where
The multiplier , on the other hand, does not satisfy a terminal condition, since it is determined at each time by the elliptic adjoint Poisson equation. In addition, periodic boundary conditions are imposed on , while reflective boundary conditions are applied to . Finally, we compute the derivative with respect to each control coefficient. For any , the -th component of the control gradient is
where, for two planar vectors and , we denote by the scalar (out-of-plane) component of their cross product.
The control optimality condition can thus be written as
For an interior optimum, this condition reduces to for every and almost every . Altogether, equations (20), (21), (23), and (26) define the continuous forward–backward optimality system. The state equations evolve forward from the prescribed initial distribution, whereas the adjoint Vlasov equation evolves backward from the terminal condition. The elliptic state and adjoint Poisson equations are solved at each time, and the control coefficients are characterized by the variational inequality (26).
From the continuous to the particle–grid optimality system
We now derive a particle–grid counterpart of the continuous optimality system (3.3)–(3.9). Throughout this subsection, continues to denote the number of control cells, whereas denotes the number of nodes of the spatial grid used by the Particle-in-Cell (PIC) discretization [14, 30, 31]. Let be the number of particles and consider the regularized empirical measure
where the particle weights satisfy , , and is a normalized, compactly supported shape function of width [40, 14, 30]. The particle positions and velocities are denoted by , with initial values chosen consistently with .
We introduce spatial grid nodes , quadrature weights , a discrete Laplacian , and a discrete gradient . We then have
while the discrete Poisson problem becomes
The grid electric field and its interpolation to the particle positions are defined by
The particle characteristics associated with the Vlasov equation are
for while . To project the continuous adjoint equation onto the particle trajectories, set
Taking the spatial and velocity gradients of the first equation in (21) yields
Defining the particle adjoints by
and applying the chain rule along (31), we obtain
Now we set
and
The gradients of the particle approximation of the source in (22) become then
with . The discretized elliptic adjoint equation becomes
where
and . The adjoints and use the same quadrature-weighted grid inner product used in (29)–(30), meaning that on a uniform grid one has and . Thus, the particle adjoint equations in backward form are
The terminal condition follows from (23):
where is obtained from the tracking part of , evaluated at ; in particular, it contains no control-regularization term. Finally, inserting (27) into the continuous gradient (25) gives, for ,
If has approximation order , then, for every sufficiently smooth function ,
Hence the particle control gradient is
The discrete control optimality condition is therefore
where is defined in (A.13). In the unconstrained case, or when the discrete amplitude constraint is inactive, this reduces formally to . Equations (31), (29), (38), (36), (39), (42) form the particle–grid optimality system. If the same shape function and quadrature are used, and if the discrete differential operators are paired with their adjoints, this projected optimize-then-discretize system agrees formally with the system obtained by discretizing the particle–grid dynamics first and then differentiating it, up to particle, grid, and time-discretization errors. A detailed comparison is provided in Appendix § 7.
Numerical approximation and optimization method
In this section, we describe the fully discrete strategy used to solve the particle–grid optimality system (31), (29), (38), and (36), with terminal conditions (39) and control optimality condition (42). Optimisation is performed using a projected-gradient method with Armijo backtracking, following standard approaches for particle and mean-field control problems [1, 4].
Let
denote the reduced, fully discrete cost functional, where is the particle approximation obtained by solving the discrete state equations for the control . The time integrals in are approximated using the same time quadrature used to define the discrete gradient below. We initialize the optimization with
where the coefficients are obtained from the instantaneous control described in Appendix § 8. In optimisation iteration , the current coefficients are kept fixed while the particle–grid state system is integrated forward in time. This produces . The adjoint equations are then solved backward along the same trajectory, producing , and the gradient is evaluated from (41). In particular, the magnetic field acting on particle is
After time discretization, let denote the discrete-time counterpart of , namely
We denote by the orthogonal projection onto this closed convex set with respect to the discrete inner product. For a trial step size , define
The accepted step size is
where is the smallest nonnegative integer for which
is satisfied, where
denotes the discrete time inner product approximating the continuous inner product via a Riemann sum, with denoting the gradient and control arrays at time respectively. The control update is then
Each trial evaluation in the line search requires a new forward solve, but the adjoint system and the gradient are computed only once at the beginning of iteration [13, 41]. The iteration stops either when the maximum number of iterations is reached, or when the cost functional stagnates, that is if
holds for consecutive iterations.
To reduce the computational cost, the state, adjoint, and gradient are evaluated on a representative subset , with . Assuming equal particle weights and relabeling the particles in the reduced ensemble by , we assign the rescaled weights
Equivalently, if the full ensemble has weights , then . Omitting for simplicity in the following the dependence on optimization iteration , we partition into intervals of length . The controls are then represented as piecewise constant functions in time, with while the Poisson problem is solved according to
The controlled magnetic field at the particle position becomes
The state characteristics in the reduced system () are advanced using a first-order semi-implicit scheme [30]:
In the same way, the particle adjoints are initialized by
while for , we employ a semi-implicit backward step for the particle adjoint equations:
with
and
The discrete gradient with respect to the control coefficients is finally computed as
We emphasize that (55) is obtained by time-discretizing the semidiscrete particle adjoint system. Its relation with the gradient obtained by differentiating the fully discrete state map is discussed in Appendix A.
While the first-order scheme (50) yields a simple and efficient formulation for the iterative optimization phase, it exhibits numerical dissipation over long times. To avoid artificial damping when computing physical quantities, all validation tests on the full ensemble are performed using an energy-preserving second-order semi-implicit scheme [30]. This reads
where the apex (1) denotes the first stage, while the second stage reads
where the electric field is computed as in (48) by solving the Poisson problem, whereas the magnetic field is computed as in (49) by replacing with . Finally, the numerical solution at time reads,
Numerical experiments
In this section, we assess the performance of the proposed control strategy on a set of standard benchmark problems. For the Armijo line search in (45)–(46), we set the parameters to , , and . The stopping tolerance for the optimization algorithm is set to over a stall limit of iterations. A full ensemble of particles is sampled from the initial distribution, whereas the forward–backward optimality system and the corresponding gradient are evaluated using a fixed representative random subset of particles at each optimization iteration. The rest of the parameters are specified separately for each test case. As described in § 2, the control is represented on a coarse rectangular partition consisting of cells, with and cells along the - and -directions, respectively. We denote the corresponding cell centers by . The control-cell widths are denoted by and . They are chosen substantially larger than the spacings and of the finer grid used in the Particle-in-Cell reconstruction, namely , . Using a double-index representation of the control coefficients introduced in § 2, the magnetic field evaluated at the -th particle position is given by
The two-dimensional control kernels are defined as tensor products,
where
with and . Once the optimal coefficients have been computed over the time interval using the reduced ensemble of particles, the resulting control is validated by evolving the full ensemble of particles using the second order scheme in (56)–(57)–(58).
To assess the effectiveness of the control strategy near the material walls, we introduce a boundary strip
where denotes the thickness of the strip. Unless otherwise specified, we set . Let denote the nodes of the fine spatial grid used for the PIC reconstruction, and define
The particle mass contained in the boundary strip is approximated by
where is the two-dimensional tensor-product quadratic B-spline shape function scaled by the spatial resolution parameter . The corresponding mean velocity is
We then define the thermal energy contained in the boundary strip as
For comparison, the total reconstructed mass and the global mean velocity are
The total thermal energy is consequently defined as
Finally, the percentage of the total thermal energy localized within the boundary strip is also measured in the sequel, and denoted by .
Two-dimensional double Sod test
We consider a two-dimensional extension of the double Sod shock-tube test. The spatial and velocity variables are denoted by and , respectively. The spatial domain is , with periodic boundary conditions in the -direction and reflective boundary conditions in the -direction. The initial distribution is given by
where the initial density and temperature are defined as
Thus, the initial configuration consists of a dense, high-temperature central layer surrounded by two low-density, low-temperature regions. Although the problem is posed in a two-dimensional domain, the initial data are homogeneous in the -direction.
The dynamics is evolved up to the final time , using the time step . The density and the self-consistent electric field are reconstructed on a uniform PIC grid with nodes, so that . For both the instantaneous and optimal controls, we set
with targets and . We first consider a reference configuration in which no optimization is performed and the magnetic field is prescribed to be uniform in space and constant in time: . Figure 1 shows the percentage of the total thermal energy contained in the boundary strip, together with three snapshots of the density profiles. The increase observed during the simulation indicates that an increasing fraction of the thermal energy becomes localized near the upper and lower boundaries.

Figure 1. Two-dimensional double Sod test with the prescribed uniform magnetic field . Snapshots of the uncontrolled dynamics at , , and and evolution of the percentage of total thermal energy contained in the boundary strip, .
We next consider the controlled dynamics, with the control computed using the optimization procedure described in the previous sections and initialized by the instantaneous control. We set and consider . Consequently, each control cell extends over the entire -direction and the reconstructed magnetic field depends only on and .
To describe the convergence of the optimization algorithm, we introduce the reduced cost functional
where denotes the solution of the forward problem associated with the control . Figure 2 reports the value of and the Euclidean norm of its discrete gradient as functions of the optimization iteration, for the different values of . The results indicate that the instantaneous control provides an effective initial guess and is already close to a stationary point of the discrete optimization problem: the initial gradient norm is relatively small and the variation of the cost functional over the iterations is limited. Nevertheless, the optimization produces a non-negligible correction to the instantaneous control and, in most cases, further reduces the fraction of thermal energy localized near the boundaries. This comparison is presented in Figure 3, which shows for the instantaneous and optimal controls. The three panels correspond to , , and , respectively. Overall, the optimal control reduces the concentration of thermal energy in the boundary strip compared with the instantaneous strategy. The first row of Figure 4 shows instead three snapshots of the

Figure 2. Two-dimensional double Sod test. Left: evolution of the reduced cost functional over the optimization iterations. Right: evolution of the discrete gradient norm . The curves correspond to .

Figure 3. Two-dimensional double Sod test. Comparison of the percentage of thermal energy contained in the boundary strip, , obtained with the instantaneous and optimal controls. From left to right: , , and .

Figure 4. Two-dimensional double Sod test. First row: snapshots of the controlled dynamics at , , and for . Second row: optimal control coefficients . The columns in the second row correspond, from left to right, to , , and .
controlled dynamics at , , and for . Similar qualitative behavior is observed for and and they are not reported. The second row displays the optimal control coefficients , for .
Two-stream plasma test
We consider the spatial domain
with periodic boundary conditions in the -direction and reflective boundary conditions in the -direction. The initial distribution is defined by
where
The transverse mean velocity is prescribed as
while the two density profiles are given by
with
With this convention, is centered at and at . The regions and are assigned opposite transverse velocities, thus generating counter-propagating particle streams. The dynamics is evolved up to the final time , using the time step . The density and the electric field are reconstructed on a uniform PIC grid with nodes, so that . For both the instantaneous and optimal controls, we set
with targets . We first consider a reference configuration in which no optimization is performed and the magnetic field is prescribed to be uniform in space and constant in time: . Figure 5 shows three density snapshots at , , and . Owing to the opposite initial transverse velocities, particles are transported toward the upper and lower boundaries and are subsequently reflected. The lower panel on the right reports the corresponding percentage of thermal energy contained in the boundary strip, .

Figure 5. Two-stream plasma test with the prescribed uniform magnetic field . Density snapshots at , , and and evolution of the percentage of thermal energy contained in the boundary strip, .
We next compute the controlled dynamics. We set and consider . Hence, the control cells extend over the entire -direction and the reconstructed magnetic field depends only on and . Figure 6 shows the convergence history of the optimization procedure. The two panels report the reduced cost functional and the Euclidean norm of its discrete gradient . As in the previous test, the iterative procedure produces a decrease of the reduced cost functional. Figure 7 shows three density snapshots of the controlled dynamics at , , and for . The control substantially limits transport toward the upper and lower boundaries and retains a larger fraction of the density in the central region. Similar qualitative behavior is observed for and . The second row displays the optimal control coefficients over time.

Figure 6. Two-stream plasma test. Left: evolution of the reduced cost functional . Right: discrete gradient norm . The curves correspond to and .

Figure 7. Two-stream plasma test. First row: density snapshots of the controlled dynamics at , , and for . Second row: optimal control coefficients . The columns in the second row correspond, from left to right, to , , and .
Finally, Figure 8 provides a direct comparison between the instantaneous and optimal controls. The optimized strategy yields a further reduction in the fraction of thermal energy localized near the boundaries.

Figure 8. Two-stream plasma test. Comparison of obtained with the instantaneous and optimal controls. From left to right: , , and .
Kelvin–Helmholtz instability
We next consider the development and control of a Kelvin–Helmholtz instability. The spatial domain is
with periodic boundary conditions in the -direction and reflective boundary conditions in the -direction. The initial spatial density is prescribed as
where
The initial phase-space distribution is
with
where
The dynamics is evolved up to the final time , using . The density and the electric field are reconstructed on a uniform PIC grid with nodes, so that . For both control strategies, we set
with targets . We first consider the reference dynamics generated by the uniform and time-independent magnetic field . Figure 9 shows density snapshots at , , and . The deformation of the initially localized layer illustrates the development of the instability. The lower panel on the right reports the corresponding percentage of thermal energy localized within the boundary strip, .

Figure 9. Kelvin–Helmholtz instability with the prescribed uniform magnetic field . Density snapshots at , , and and evolution of the percentage of thermal energy contained in the boundary strip, .
For the controlled simulations, we set and consider . Consequently, the control cells extend over the entire -direction and the reconstructed magnetic field depends only on and . Figure 10 reports the convergence history of the optimization procedure. The panels show the reduced cost functional and the discrete gradient norm . Figure 11 shows three density snapshots of the controlled dynamics at , , and for . Similar qualitative results are obtained for and . The control attenuates the development of the instability and substantially reduces transport toward the upper and lower boundaries. The lower panel show the optimal controls coefficients for . Finally, Figure 12 provides a direct comparison in terms of relative thermal energy between the instantaneous and optimal controls. As for the previous cases, in most of the cases, the optimized strategy yields a further reduction in the fraction of thermal energy localized near the boundaries.

Figure 10. Kelvin–Helmholtz instability. Left: evolution of the reduced cost functional . Right: discrete gradient norm . The curves correspond to and .

Figure 11. Kelvin–Helmholtz instability. Top row: density snapshots of the controlled dynamics at , , and for . Bottom row: optimal control coefficients . The columns in the second row correspond, from left to right, to , , and .

Figure 12. Kelvin–Helmholtz instability. Comparison of obtained with the instantaneous and optimal controls. From left to right: , , and .
Conclusions
We have developed a full-horizon, open-loop strategy for controlling a 2D–2D Vlasov–Poisson system by means of an externally imposed magnetic field. The field is parameterized by time-dependent coefficients associated with smooth, spatially localized kernels, thereby reducing the infinite-dimensional control problem to the determination of a finite family of control functions. The objective combines the tracking and dispersion of selected position and velocity observables with a penalty on the control effort, while a pointwise bound limits the intensity of the reconstructed magnetic field. Starting from the kinetic optimization problem, we derived the continuous first-order optimality conditions. We then projected this system onto a particle–grid representation and obtained particle adjoint equations. The discretize-then-optimize calculation in Appendix A shows that the same semidiscrete system is recovered when the particle deposition and interpolation operators, the grid quadrature, and the adjoint discrete operators are paired consistently, up to the localization error associated with the particle shape function. The resulting optimization problem is solved by a projected gradient method with Armijo backtracking on a reduced particle ensemble. An instantaneous feedback control, adapted to the same smooth spatial parameterization, provides the initial guess for the full-horizon iteration.
The numerical experiments illustrate the behavior of the method. In the tests considered, the instantaneous control already gives an effective initialization and is often close to a stationary point of the discrete full-horizon problem. Nevertheless, the full-horizon optimization produces a further correction that decreases the reduced objective and generally lowers the normalized thermal-energy diagnostic in the boundary strips.
Several extensions are natural such as the use of collisional or electromagnetic kinetic models, more realistic geometries and coil-generated fields. Moreover, objectives based directly on particle or energy fluxes at material boundaries would bring the framework closer to magnetic confinement applications.
Discretize-then-optimize derivation of the particle–grid optimality system
We provide a discretize-then-optimize derivation of the semidiscrete particle–grid optimality system introduced in § 3. Time remains continuous throughout this appendix. Let
the density and the electric field interpolated to particle are
The particle dynamics is therefore
Set
For , define
and
The cost functional at the particle level becomes
with and . Let denote the particle contribution
Thus
and the choices
give precisely the expressions in (35). The particle–grid Lagrangian is
Here denotes the grid vector whose entries are all equal to and . Stationarity with respect to , , and recovers (A.1)–(A.2). To compute the variation with respect to the grid potential, define
Using the grid inner product and the adjoint operators and , stationarity with respect to gives
which coincides with (36).
Integrating the terms containing and by parts in time, and using that the initial particle states are fixed, stationarity with respect to and yields
Moreover,
so that (A.10) is exactly (38). The terminal conditions are
where contains only the terminal tracking and dispersion terms, and no control regularization. Finally, the derivative of (A.7) with respect to is
At the particle–grid level, we enforce the field bound at the spatial grid nodes and set
For a bounded control, stationarity is then expressed by the variational inequality
Thus, the discretize-then-optimize calculation recovers (38), (36), (39), and (41). For the particle objective (A.5), the calculation is exact at the particle–grid semidiscrete level when the same discretization are employed. Its comparison with the projected continuous system additionally involves replacing convolutions against by evaluations along the particle trajectories. For a shape function of order , these localization discrepancies are of order ; in particular, this accounts for the remainder displayed in (41).
Relation with the fully discrete adjoint
The derivation above is semidiscrete, with time kept continuous. Indeed, in Section 4, the gradient (55) is obtained by discretizing in time the semidiscrete optimality system. We now compare it with the gradient obtained by differentiating the fully discrete state problem. The semi-implicit forward scheme (50) can then be written as
where
Together with the discrete Poisson problem, this defines the fully discrete state map
Using the same time quadrature as in Section 4, the corresponding discrete Lagrangian may be written as
Since
we introduce
which collects the contribution of the position and velocity adjoints through the semi-implicit velocity update. Thus, the discrete adjoint writes as
with terminal conditions
and where
Here denotes the contribution associated with the dependence of the electric field on the particle density, obtained from the same discrete Poisson-adjoint relation derived in (A.10) and evaluated at time . Finally, stationarity with respect to gives, for the diagonal regularization (12),
We observe that since
and the adjoint discretizations in (A.17) and (A.19) are such that
It follows that the gradient used in § 4 is a first-order consistent approximation of the fully discrete gradient, indeed we have
where is the gradient defined in (55).
Computation of the instantaneous control
In the following, we will briefly describe how to compute the baseline instantaneous feedback control , referring to [2, 3] for a complete description of the methodology. Accordingly, the initial guess for the open-loop control in the optimization procedure is chosen to be this instantaneous feedback law.
We consider the following minimization problem at the particle level:
with and defined as in (A.6) and (A.13) respectively, and a penalization coefficient.
We introduce an explicit first order discretization of the characteristic curves in (31) over the time interval , with step size as follows:
for any , and where
with denoting the spatial basis functions defined as in (60). Following [2, 3], we introduce the scaling
while and are kept fixed. Since the -component of the velocity update satisfies
we have
Using (B.4) and considering the limit with , the first-order optimality conditions yield the instantaneous feedback law
where is a diagonal matrix based on the assumption (11), and we have
with
Equation (B.5) provides the unconstrained instantaneous control, if the constraint in is active, the corresponding constrained quadratic problem is solved instead. Once determined through this first-order framework, the instantaneous control coefficients are applied within the second-order scheme to advance the particle trajectories.
Acknowledgement
This work has been written within the activities of GNCS and GNFM groups of INdAM (Italian National Institute of High Mathematics). The research of FF and LP was supported by the Italian Ministry of University and Research (MUR) through the Fondo Italiano per la Scienza 2022–2023 (Bando FIS 2), under Decree No. 1236 of 01/08/2023, within the FIS2023-01334 Advanced Grant “Tackling Complexity: Advanced Numerical Approaches for Multiscale Systems with Uncertainties” (ADAMUS). The work of LP was supported by the Royal Society Wolfson Fellowship “Uncertainty quantification, data-driven simulations and learning of multiscale complex systems governed by PDEs”. GA and GD acknowledge the support of the PRIN 2026 - 2026PYFEZ9, “Emergent Structures in Kinetic Theory: from Gases to Plasmas” funded with the contribution of MUR.
References
References
- [1]G. Albi, Y.-P. Choi, M. Fornasier, and D. Kalise. Mean field control hierarchy. Applied Mathematics & Optimization, 76(1):93–135, 2017.arxiv.org/abs/1608.01728
- [2]G. Albi, G. Dimarco, F. Ferrarese, and L. Pareschi. Instantaneous control strategies for magnetically confined fusion plasma. Journal of Computational Physics, 527:113804, 2025.arxiv.org/abs/2403.16254
- [3]G. Albi, G. Dimarco, F. Ferrarese, and L. Pareschi. Robust feedback control of collisional plasma dynamics in presence of uncertainties. Journal of Computational Physics, 546:114512, 2026.arxiv.org/abs/2505.19992
- [4]R. Bailo, M. Bongini, J. A. Carrillo, and D. Kalise. Optimal consensus control of the Cucker-Smale model. IFAC-PapersOnLine, 51(13):1–6, 2018.DOI
- [5]J. Bartsch, P. Knopf, S. Scheurer, and J. 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.
- [6]J. A. Bittencourt. Fundamentals of plasma physics. Springer Science & Business Media, 2013.
- [7]J. Blum. Numerical simulation and optimal control in plasma physics. Wiley/Gauthier-Villars Series in Modern Applied Mathematics. John Wiley & Sons, Ltd., Chichester; Gauthier-Villars, Montrouge, 1989.
- [8]R. Caflisch, C. Wang, G. Dimarco, B. Cohen, and A. Dimit s. A hybrid method for accelerated simulation of Coulomb collisions in a plasma. Multiscale Modeling & Simulation, 7(2):865–887, 2008.
- [9]S. Caprino, G. Cavallaro, and C. Marchioro. Time evolution of a Vlasov-Poisson plasma with magnetic confinement. Kinetic & Related Models, 5(4):729–742, 2012.DOI
- [10]E. Chacon-Golcher, S. A. Hirstoaga, and M. Lutz. Optimization of Particle-In-Cell simulations for Vlasov-Poisson system with strong magnetic field. ESAIM: Proceedings and Surveys, 53:177–190, 2016.
- [11]P. Chen, R. Jorge, Q. Li, and Y. Yue. Control of a Uniformly Magnetized Plasma with External Electric Fields. Journal of Statistical Physics, 193(40), 2026.arxiv.org/abs/2509.07988
- [12]C.-Z. Cheng and G. Knorr. The integration of the Vlasov equation in configuration space. Journal of Computational Physics, 22(3):330–351, 1976.
- [13]A. R. Conn, N. I. Gould, and P. L. Toint. Trust Region Methods, SIAM. MPS, 2000.DOI
- [14]G.-H. Cottet and P.-A. Raviart. Particle methods for the one-dimensional Vlasov-Poisson equations. SIAM Journal on Numerical Analysis, 21(1):52–76, 1984.DOI
- [15]A. Crestetto, N. Crouseilles, and M. Lemou. Kinetic/fluid micro-macro numerical schemes for Vlasov-Poisson-BGK equation using particles. Kinetic and Related Models, 5(4):787–816, 2012.DOI
- [16]N. Crouseilles, G. Dimarco, and M.-H. Vignal. Multiscale schemes for the BGK-Vlasov-Poisson system in the quasi-neutral and fluid limits. Stability analysis and first order schemes. Multiscale Model. Simul., 14(1):65–95, 2016.DOI
- [17]N. Crouseilles, M. Mehrenberger, and E. Sonnendrücker. Conservative semi-Lagrangian schemes for Vlasov equations. Journal of Computational Physics, 229(6):1927–1953, 2010.DOI
- [18]N. Crouseilles, T. Respaud, and E. Sonnendrücker. A forward semi-Lagrangian method for the numerical solution of the Vlasov equation. Comput. Phys. Comm., 180(10):1730–1745, 2009.DOI
- [19]P. Degond. Asymptotic-preserving schemes for fluid models of plasmas. In Numerical models for fusion, volume 39/40 of Panor. Synthèses, pages 1–90. Soc. Math. France, Paris, 2013.
- [20]P. Degond and F. Deluzet. Asymptotic-preserving methods and multiscale models for plasma physics. Journal of Computational Physics, 336:429–457, 2017.DOI
- [21]J. Degrave, F. Felici, J. Buchli, M. Neunert, B. Tracey, F. Carpanese, T. Ewalds, R. Hafner, A. Abdolmaleki, D. de las Casas, et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature, 602(7897):414–419, 2022.DOI
- [22]F. Deluzet, G. Fubiani, L. Garrigues, C. Guillet, and J. Narski. Efficient parallelization for 3d-3v sparse grid particle-in-cell: shared memory architectures. Journal of Computational Physics, 480:112022, 2023.DOI
- [23]G. Dimarco, Q. Li, L. Pareschi, and B. Yan. Numerical methods for plasma physics in collisional regimes. Journal of Plasma Physics, 81(1):305810106, 2015.DOI
- [24]G. Dimarco, L. Mieussens, and V. Rispoli. An asymptotic preserving automatic domain decomposition method for the Vlasov–Poisson–BGK system with applications to plasmas. Journal of Computational Physics, 274:122–139, 2014.DOI
- [25]G. Dimarco and L. Pareschi. Numerical methods for kinetic equations. Acta Numerica, 23:369–520, 2014.DOI
- [26]L. Einkemmer, Q. Li, C. Mouhot, and Y. Yue. Control of instability in a Vlasov-Poisson system through an external electric field. Journal of Computational Physics, 530:113904, 2025.arxiv.org/abs/2407.15008
- [27]L. Einkemmer, Q. Li, L. Wang, and Y. 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
- [28]A. Fasoli, S. Brunner, W. A. Cooper, J. P. Graves, P. Ricci, O. Sauter, and L. Villard. Computational challenges in magnetic-confinement fusion physics. Nature Physics, 12(5):411–423, 2016.DOI
- [29]F. Ferrarese. Control strategies for magnetized plasma: a polar coordinates framework. arXiv preprint arXiv:2511.10214, 2025.arxiv.org/abs/2511.10214
- [30]F. Filbet and L. M. Rodrigues. Asymptotically stable particle-in-cell methods for the Vlasov–Poisson system with a strong external magnetic field. SIAM Journal on Numerical Analysis, 54(2):1120–1146, 2016.
- [31]F. Filbet and L. M. Rodrigues. Asymptotically preserving particle-in-cell methods for inhomogeneous strongly magnetized plasmas. SIAM Journal on Numerical Analysis, 55(5):2416–2443, 2017.DOI
- [32]F. Filbet and E. Sonnendrücker. Comparison of Eulerian Vlasov solvers. Computer Physics Communications, 150(3):247–266, 2003.
- [33]F. Filbet, E. Sonnendrücker, and P. Bertrand. Conservative numerical schemes for the Vlasov equation. Journal of Computational Physics, 172(1):166–187, 2001.DOI
- [34]F. Filbet and C. Yang. Numerical Simulations to the Vlasov-Poisson System with a Strong Magnetic Field. arXiv:1805.10888 / hal-01800438, 2018.arxiv.org/abs/1805.10888
- [35]J. P. Freidberg. Plasma physics and fusion energy. Cambridge university press, 2008.DOI
- [36]P. R. Garabedian. Computational mathematics and physics of fusion reactors. Proceedings of the National Academy of Sciences, 100(24):13741–13745, 2003.pubmed.ncbi.nlm.nih.gov/14614129
- [37]V. Grandgirard and Y. Sarazin. Gyrokinetic simulations of magnetic fusion plasmas. Numerical models for fusion, 39:91–176, 2013.
- [38]A. Gu, Y. He, and Y. Sun. Hamiltonian particle-in-cell methods for Vlasov–Poisson equations. Journal of Computational Physics, 467:111472, 2022.DOI
- [39]E. Hairer and C. Lubich. Energy behaviour of the Boris method for charged-particle dynamics. BIT Numerical Mathematics, 58(4):969–979, 2018.
- [40]J. Hesthaven, C. Pagliantini, and N. Ripamonti. Adaptive symplectic model order reduction of parametric particle-based Vlasov–Poisson equation. Mathematics of Computation, 93(347):1153–1202, 2024.arxiv.org/abs/2201.05555
- [41]C. T. Kelley. Iterative methods for optimization. SIAM, 1999.DOI
- [42]P. Knopf and J. Weber. Optimal control of a Vlasov–Poisson plasma by fixed magnetic field coils. Applied Mathematics & Optimization, 81:961–988, 2020.
- [43]J. Lu, L. Wang, and J. Calder. Controlling instability in the Vlasov-Poisson system through moment-based optimization. arXiv preprint arXiv:2508.18412, 2025.arxiv.org/abs/2508.18412
- [44]J. Lu, L. Wang, and J. Calder. Dynamical feedback control with operator learning for the vlasov-poisson system. arXiv preprint arXiv:2509.23063, 2025.arxiv.org/abs/2509.23063
- [45]A. Medaglia, L. Pareschi, and M. Zanella. A DSMC-PIC coupling method for the Vlasov–Maxwell–Landau system. Journal of Computational Physics, page 115160, 2026.
- [46]S. Ravindran. Real-time computational algorithm for optimal control of an MHD flow system. SIAM Journal on Scientific Computing, 26(4):1369–1388, 2005.DOI
- [47]J. A. Rossmanith and D. C. Seal. A positivity-preserving high-order semi-Lagrangian discontinuous Galerkin scheme for the Vlasov-Poisson equations. Journal of Computational Physics, 230(16):6203–6232, 2011.DOI
- [48]E. Sonnendrücker, J. Roche, P. Bertrand, and A. Ghizzo. The semi-Lagrangian method for the numerical resolution of the Vlasov equation. Journal of Computational Physics, 149(2):201–220, 1999.DOI
- [49]J. Weber. Optimal control of the two-dimensional Vlasov-Maxwell system. ESAIM: Control, Optimisation and Calculus of Variations, 27:S19, 2021.DOI
- [50]J. Wesson and D. J. Campbell. Tokamaks, volume 149. Oxford university press, 2011.
- [51]C. Zhang and I. M. Gamba. A conservative scheme for Vlasov Poisson Landau modeling collisional plasmas. Journal of Computational Physics, 340:470–497, 2017.DOI
- [52]Y. Zhu and S. Jin. The Vlasov–Poisson–Fokker–Planck system with uncertainty and a one-dimensional asymptotic preserving method. Multiscale Modeling & Simulation, 15(4):1502–1529, 2017.DOI