Plasma and Fusion Research
Online ISSN : 1880-6821
ISSN-L : 1880-6821
Letters
Inference of Nonlinear Wave Generation in Plasma Boundary Layers Using a Physics-Informed Neural Network
Katsuya KONDO, Ryutaro MATSUI, Kenji IMADERA
著者情報
ジャーナル オープンアクセス HTML

2026 年 21 巻 論文ID: 21.1301053

詳細
Abstract

The nonlinear wave triggered at the boundary layer between laser-irradiated solid carbon and ambient hydrogen gas is analyzed using a physics-informed neural network (PINN)-based approach. We demonstrate that the boundary layer evolves through multiple phases characterized by (I) solid-plasma expansion while maintaining the contact surface, (II) formation of a mixing region, and (III) bifurcation of the gas surface into a nonlinear wave and a rarefaction front. The one-dimensional results show that the gas surface bifurcates simultaneously with the formation of the mixing region, in contrast to observations in two-dimensional cluster-gas systems. The PINN model quantitatively captures the dynamics across these transitions and verifies that the nonlinear wave launched through this process is governed by an advective Korteweg-de Vries (KdV)-type equation.

Nonlinear waves in plasmas are an important research topic. The nonlinearity becomes particularly significant in large-amplitude waves, triggering various phenomena such as shock formation and the excitation of wave trains [1, 2]. One possible origin of large-amplitude nonlinear waves is plasma boundary layers [3], where two qualitatively different plasmas come into contact. Plasma boundary layers are ubiquitous in both astrophysical and laboratory plasmas. In such boundary layers, a discontinuity triggers various nonlinear phenomena over a wide range of spatiotemporal scales.

Previous work has demonstrated the formation of nonlinear waves in the boundary layer formed by a solid carbon cluster and ambient hydrogen gas, produced by irradiation with a femtosecond-pulse laser [4]. The laser primarily heats the solid target, leading to formation of a plasma boundary layer consisting of the hot solid cluster and the cold ambient gas. The evolution of this boundary layer is characterized by the following distinct phases. First, the solid plasma expands while maintaining a contact surface, leading to shock acceleration of the ambient gas ions (phase 1). The accelerated gas ions accumulate near the contact surface, forming a compressed surface. After the formation of the compressed surface, the gas ions penetrate into the solid plasma, resulting in the formation of a mixing region (phase 2). Finally, the compressed surface bifurcates into a rarefaction front and a peak structure, which marks the onset of a nonlinear wave (phase 3). Since this nonlinear wave exhibits the characteristics of solitary waves [5], it suggests that the plasma boundary layer may be one possible origin of solitary wave formation. However, the universality of such nonlinear wave formation and the validity of its theoretical model have not been fully verified yet.

In this letter, we investigate the phase transition of the plasma boundary layer in a one-dimensional system using a particle simulation and a physics-informed neural network (PINN) [6], which can infer the governing equations by incorporating them into the learning process. The simulation results show that the compressed surface bifurcates simultaneously with the formation of the mixing region, in contrast to observations in two-dimensional cluster-gas systems. The PINN captures the dynamics across the transitions and quantitatively verifies that the nonlinear wave launched by the plasma boundary layer exhibits the characteristics of a soliton governed by the advective KdV equation.

The original simulation data are computed using the particle-based integrated code EPIC3D [7]. The initial plasma profiles are same as those used in the previous work [4]. In the present electromagnetic simulation, a quasi-one-dimensional system is modeled, in which the y-direction corresponds to the propagation direction of the laser, whereas the x-direction corresponds to the direction of the laser electric field Ex. The simulation domain is Lx×Ly = 0.04 × 64.00 μm, divided into 4 × 6,400 meshes; i.e., the spatial mesh size is Δx=Δy = 0.01 μm. Because a plane laser is assumed, the electromagnetic fields are nearly uniform in the x-direction and can be described using a small Lx with a periodic boundary condition. The target plasmas are assumed to be fully ionized. The solid slab consists of carbon C6+, whose electron density is neC = 227 nc, where nc = 1.70 × 1021 cm−3 is the cutoff density for the injected laser with wavelength λL = 0.810 μm. The slab occupies the region 6.75 ≤ y ≤ 7.00 μm. The ambient gas plasma consists of hydrogen H+, whose electron density is neH = 0.192 nc. The gas occupies the region 7.00 ≤ y ≤ 57.00 μm. The plane laser pulse is injected from an antenna located at y = 0.04 μm at t = 0 fs with electric field Ex and magnetic field Bz, and reaches its maximum amplitude at t = 40 fs. The peak reaches the front of the target at t = 62 fs. The pulse duration is τ0 = 33 fs, defined as the full width at half maximum (FWHM). The peak laser intensity is I = 1.0 × 1019 W/cm2, corresponding to the normalized intensity a0 = 2.19, where a0≡eE0/(mecωL). Here e and me denote the electron charge and mass, c is the speed of light, and ωL and E0 denote the laser frequency and peak value of the laser electric field, respectively. The boundary conditions for both fields and particles are transparent in the y-direction and periodic in the x- and z-directions. The transparent boundary condition for fields is implemented according to the Mur absorbing boundary condition. The transparent boundary condition for particles is implemented by fixing the particles that reach the boundary and setting their charge to zero.

First, we present the temporal evolution of the plasma boundary layer. Figure 1 shows the cross-sectional profiles of the ion and electron charge densities, together with the longitudinal electric field Ey. The electrons in the solid carbon are pushed into the +y direction by the ponderomotive force as indicated by the red dashed lines in Fig. 1(a). In contrast, the carbon ions cannot respond immediately to the ponderomotive force, resulting in charge separation and the formation of an ambipolar electric field at the rear surface of the solid carbon. The magnitude of the ambipolar field is on the order of Ey ∼ 1.0 TV/m. Once the ambipolar field is formed, the solid carbon begins to expand while maintaining the contact surface, which leads to shock acceleration of the ambient gas ions. The accelerated gas ions accumulate near the contact surface, resulting in the formation of a compressed surface [Fig. 1(a)]. This period corresponds to Phase 1, during which the expansion of the solid carbon is almost identical to free expansion into vacuum and is only weakly influenced by the ambient hydrogens.

Fig. 1.  Cross-sectional profiles of the ion density (red: solid carbon, blue: gas hydrogen) and the electric field Ey (green) at (a) t = 133.4 fs, (b) t = 166.8 fs, and (c) t = 266.8 fs. Only in (a), the red and blue dashed lines indicate the electron densities originating from the initial carbon and hydrogen species, respectively.

At t ∼ 133 fs, the phase transition occurs. Note that the electron density is omitted in Figs. 1(b)–(c) to clarify the dynamics of ions. At this time, the expansion front of the carbon begins to penetrate into the ambient gas, forming a mixing region between the expansion front of the carbon ions and the compressed surface of the ambient gas ions [Fig. 1(b)]. This period, characterized by the mixing region, corresponds to Phase 2. Simultaneously, the compressed surface of ambient gas bifurcates into a rarefaction front and a localized peak structure, indicating the phase transition from Phase 2 to Phase 3. The peak structure evolves into a nonlinear wave and propagates in the ambient hydrogen gas in the +y direction, eventually reaching a steady state [Fig. 1(c)]. Interestingly, the solid carbon develops a staircase structure, in which the carbon ion density exhibits a transition from a flat plateau to a steep gradient in its spatial profile, whereas such a structure does not appear in the case of free expansion into vacuum. The staircase structure is due to the negative component of the bipolar field generated by the compressed ambient gas.

Here, we discuss the difference between this 1D case and the 2D cluster case reported in previous work [4]. In the 1D case, the compressed surface bifurcates immediately after the mixing region is formed, whereas, in the 2D cluster case, the compressed surface remains for a finite period after the formation of the mixing region; thus, phase 2 is scarcely observed in the 1D case. After the bifurcation, the rarefaction front drifts monotonically in the +y direction in the 1D case, while in the 2D cluster case it propagates toward the center of the target. These differences arise from the presence or absence of pressure gradient in the region previously occupied by the solid carbon. In the 2D case, a lateral pressure gradient drives the ambient gas into the region previously occupied by the cluster, whereas no lateral pressure exists in the 1D slab case.

Next, we present the inferred results for the ambient gas obtained by the PINN model in Fig. 2. The model consists of 9 dense layers, each with 20 neurons. As the activation function, the hyperbolic tangent function is used in the hidden layers, and the linear function is used in the output layer. Training is performed for 10,000 epochs with a batch size of 256. The Adam optimizer is applied and the learning rate is set to 0.001. The loss function is defined as L=Ldata+keqLeq, where Ldata represents the squared deviation between the original simulation data and the PINN prediction, and Leq denotes the squared residual of the governing equation. The governing equation is assumed to take the form ∂tn+ax∂xn+axx∂x2n+axxx∂x3n+anxn∂xn = 0, where ∂t and ∂x denote partial differentiation with respect to t and x, respectively. This equation is introduced as a generic one-dimensional transport model that includes advection, dissipation, dispersion, and a nonlinear term. The coefficients ax,axx,axxx, and anx are treated as trainable parameters within the PINN framework and are inferred to best reproduce the observed spatiotemporal evolution of n⁡(x,t). In the case where only the dispersion term axxxnxxx and the nonlinear term anxnnx remain, the KdV equation can be considered to have been inferred. The details of PINN model and governing equation are discussed in Ref. [8]. The loss weight for the equation term is set to keq = 100.

Fig. 2.  Overview of the process of PINN training for the data at t = 200–300 steps (133–200 fs). (a) The evolution of the loss (purple: Ldata, orange: keqLeq). The loss is plotted every 200 epochs. (b) The evolution of the inferred coefficients (blue: ax, orange: axx, green: axxx, red: anx).

To generate the training dataset, we apply a sliding-window procedure in the temporal direction. Each window spans 100 time steps (66.7 fs) and is shifted by 50 time steps (33.35 fs). The spatial domain for each window is chosen such that the nonlinear wave is fully contained within the window. A separate training and inference run is performed for each window; for example, we train and infer on the data segments corresponding to t = 100–200 steps, t = 150–250 steps, and so on.

Figure 2 shows the training process for the data for 200–300 steps (133–200 fs). Figure 2(a) shows the evolution of Ldata and keqLeq. The ratio of two loss terms is Ldata/keqLeq ∼ 10, indicating that both terms contribute appropriately to the training process [8, 9]. Figure 2(b) shows the evolution of the coefficients ax,axx,axxx, and anx. All coefficients converge within 3,000 epochs.

To validate the inferred results, we show the solution of the inferred equation and the averaged values of the derivative terms. The solution is computed using the first time step of the training window as the initial condition. Figures 3(a1)–(a3) show the comparison between the original simulation data, the neural-network prediction, and the solution of the inferred equation for three typical periods: (i) 150–250 steps, (ii) 250–350 steps, and (iii) 300–400 steps. The inferred coefficients ax,axx,axxx and anx, as well as the advection velocity calculated from ax (vadvPINN/c) and that obtained from the propagation of the rarefaction front (vadvdata/c), are summarized in Table 1.

Fig. 3.  (a1)–(a3) Spatial profiles of the hydrogen ion density compared among the original data (red solid line), the neural network prediction (blue dashed line), and the solution of PINN-inferred equation (green solid line). (b1)–(b3) Time-averaged spatial profiles of each derivative term: nt (black), axnx (blue), axxnxx (orange), axxxnxxx (green) and anxnnx (red). The vertical dashed line indicates the position of the rarefaction front at each time.
Table 1. Comparison of the inferred coefficients by the PINN and the advection velocities vadv obtained from the PINN (vadvPINN/c) and from the rarefaction front propagation (vadvdata/c).

steps (i) 150–250 (ii) 250–350 (iii) 300–400
ax [arb.unit] 0.42340 0.28325 0.23376
axx [arb.unit] −0.00120 0.00075 −0.00096
axxx [arb.unit] −0.00524 0.00401 0.00926
anx [arb.unit] 0.11664 0.27441 0.30747
vadvPINN/c 0.021174 0.014165 0.011690
vadvdata/c 0.025005 0.014503 0.0095018

During period (i), which includes the phase transition, the numerical solution exhibits an excessively strong dispersive tail and therefore fails to accurately reproduce the original data. Although the peak position is nearly consistent with the original data, the amplitude of the nonlinear wave is reduced due to energy loss into the dispersive tail. This indicates that the coefficients inferred by the PINN do not correctly reproduce the balance between dispersion and nonlinearity. In this period, since the solid carbon strongly interacts with the ambient gas, such dynamics are not precisely captured using only the data from the ambient gas. After the phase transition, the numerical solution agrees well with the original data. In periods (ii) and (iii), the numerical solutions remain consistent with the original data. The deviation in amplitude becomes much smaller than that in period (i). Because the amplitude is highly sensitive to the balance between dispersion and nonlinearity, a slight deviation from the original data therefore remains.

Figures 3(b1)–(b3) show the time-averaged values of the derivative terms nt,axnx,axxnxx,axxxnxxx and anxnnx for each spatial grid point. In all cases, the derivative terms are computed properly in the region occupied by the ambient gas. The dissipation term axxnxx is below 1% of nt, indicating that dissipation is negligible. The presence of the dispersion term axxxnxxx and the nonlinear term anxnnx demonstrates that the PINN identifies a KdV-type equation as governing the nonlinear wave across all three periods. Remarkably, the propagation of the compressed surface and the rarefaction front indicates the advection of the ambient gas because the advection term axnx does not appear in the theoretical KdV soliton solution for a uniform background plasma [8]. In period (ii), the advection velocity calculated from ax is vadvPINN = 0.0142c, which is consistent with the velocity of the rarefaction front, vadvdata = 0.0145c. The relative error (vadvPINN−vadvdata)/vadvdata = 2.33% indicates that the PINN accurately infers the advection term associated with the propagation of the compressed surface. In the other periods, the inferred advection velocity vadvPINN remains on the same order as the velocity of the compressed surface vadvdata.

Finally, we investigate the temporal evolution of the spatially averaged derivative terms and PINN-inferred coefficients, as shown in Fig. 4. Both are evaluated at the midpoint of each time window. The dashed line shows the time at which the expansion front of the carbon overtakes the compressed surface of the gas, which corresponds to the phase transition. Before the phase transition, PINN cannot infer physically meaningful coefficients; the numerical solution based on inferred coefficients does not match the simulation data and diverges toward the end. During the phase-transition period, the error of the numerical solution still exceeds 0.08 in normalized units. However, as the nonlinear wave relaxes into steady propagation, the error decreases and converges to a smaller value of approximately 0.03. After the phase transition, the coefficients change gradually. The advection term and ax reach their peak at t = 167 fs and then decrease monotonically, corresponding to the deceleration of the rarefaction front. The dispersive term and axxx remain within nearly the same range, while the nonlinear term and anx increase gradually. This supports the scenario that the nonlinear wave formed in the early phase gradually relaxes into a steady KdV soliton.

Fig. 4.  The temporal evolution of (a) the spatially averaged values of each derivative term and (b) the inferred coefficients and the error between the original simulation data and the solution of PINN-inferred equation (black line). The derivative terms and coefficients at each time are inferred using the time window t ± 33.4 fs, which contains 100 steps.

In conclusion, we have investigated the dynamics of the nonlinear wave arising from the plasma boundary layer consisting of solid carbon and ambient hydrogen gas using a PINN-based method. The nonlinear wave is launched through three phases reported in the previous work. However, in contrast to the 2D cluster case, the compressed surface of the ambient gas bifurcates into a rarefaction front and a peak structure simultaneously with the formation of the mixing region. Additionally, the intensity of the electric field Ey is sustained for a longer time, and the compressed surface of gas does not return to −y direction. Therefore, from the viewpoint of ion acceleration, the 1D configuration provides more efficient acceleration of the ambient gas ions. The PINN has identified that the nonlinear wave exhibits KdV-type nonlinear and dispersive characteristics, as evidenced by the inferred coefficients and the time-averaged values of the derivative terms. In this work, the PINN fails to infer the coefficients before the transition phase due to the strong interaction between the solid carbon and the ambient gas. This suggests that a two-fluid model describing carbon and hydrogen may be required. Another possible extension is the introduction of an additional source term. Such extensions will be reported in future work.

Acknowledgement

The authors thank Dr. Yasuaki Kishimoto for fruitful discussion about the boundary layer. This work was supported by JSPS KAKENHI Grant Numbers 21H04452, 23K20038, 24K06992, and 26K00680. In this research work, we used the supercomputer of ACCMS, Kyoto University.

References
 
© 2026 The Japan Society of Plasma Science and Nuclear Fusion Research

This article is licensed under a Creative Commons [Attribution 4.0 International] license.
https://creativecommons.org/licenses/by/4.0/
feedback
Top