betaFeedback

WorksPhysics and Astronomy

Three-Wave Forced Hasegawa–Wakatani Model: Reducing Chaos with Zonal Flows

Abstract

We investigate the nonlinear dynamics of resistive drift-wave turbulence with a three-wave truncation of the Hasegawa–Wakatani equations, in which the background density gradient either is fixed or evolves self-consistently, fed by an external particle source PP (such as neutral beam injection) and eroded by the turbulent particle flux. Two triads that differ only by a small tilt of the long-wavelength mode are compared, in one this mode is zonal and cannot be driven by the gradient, in the other it is a weakly driven drift wave. Period maps in the plane of the adiabaticity parameter CC and the drive show that the zonal triad is almost always regular, whereas the non-zonal triad becomes chaotic. The feedback of the flux on the gradient amplifies this contrast: the fraction of aperiodic states grows considerably in the non-zonal triad and remains negligible in the zonal one. Along a line of constant CC at low adiabaticity, the non-zonal triad with a fixed gradient enters chaos through torus breakdown with intermittency and leaves it through an interior crisis and a period-doubling cascade. With the self-consistent gradient, a Neimark–Sacker bifurcation leads to a phase-locked torus that breaks down abruptly, with hysteresis, and chaos then persists up to the largest drive, with the gradient driven intermittently above the threshold of the long-wavelength mode. None of these routes to chaos exists when that mode is zonal.

Introduction

One of the key problems in toroidal magnetic confinement systems is the underlying mechanism behind the emergence and suppression of anomalous transport[1, 2, 3]. By understanding how turbulence, or its components, arises in magnetically confined plasmas, one can use methods to control it and obtain information about how the transport will behave[2]. Such turbulent processes often display a broad fluctuation spectrum, with maxima at the longest measured scales (small wave vectors) and low frequencies[1]. Because wave-like instabilities occur only over a finite range of the spectrum, nonlinear mechanisms such as three- or four-wave interactions are essential for explaining the broadband fluctuations observed in laboratory plasmas[4, 5, 6].

Among the main waves that affect such systems are the drift waves, which arise from the coupling between oscillations of the electrostatic potential and the density in the presence of a background density gradient, and which play an important role in transport physics[1, 7]. In high-temperature plasmas, sufficiently steep density gradients at the plasma edge can give rise to fully developed drift-wave turbulence, which is considered a likely candidate for explaining the observed anomalous transport rates[2, 8].

Many models have been proposed to describe the emergence and evolution of this type of turbulence in magnetically confined plasmas. One of the first and most influential is the Hasegawa–Mima (HM) model[9, 10], which captures the nonlinear dynamics of the electrostatic potential in a magnetized plasma. It rests on simple assumptions, quasineutrality, a uniform and constant magnetic field, cold ions responding through the polarization drift, and Boltzmann (adiabatic) electrons. Even so, it is a robust model, widely used to describe drift-wave propagation and the formation of zonal flows[11, 12, 13].

The main limitation of the HM model is its absence of linear instability, where the density and potential fluctuations are exactly in phase, so there is no cross-field particle flux, and the system needs external parameters and forcing to display behaviors that are normally observed in confined plasmas[4, 14]. To overcome this, Hasegawa and Wakatani proposed a more complete model[15] (HW), in which the plasma density is incorporated into the dynamics and the potential and density are coupled through an adiabaticity parameter, which measures the strength of the parallel electron response, making each respond to changes in the other[15, 16]. This makes the HW model a minimal model of drift-wave turbulence, owing to the presence of linear instability, nonlinear resonant interactions, and zonal flows.

Zonal flows are axisymmetric sheared E×B\mathbf{E}\times\mathbf{B} flows, generated by the turbulence itself through nonlinear interactions, that regulate transport by shearing the turbulent eddies[8, 17, 18, 19, 20, 21]. In the HW system they have a distinctive property, since they have no variation along the drift direction (ky=0k_y=0), the density gradient cannot drive them, and they can only exchange energy with the drift waves through nonlinear coupling. They therefore act as a pure regulator of the turbulence, rather than as one more unstable wave[22, 23].

In the HW model the density is a dynamical field, but it evolves on top of a background density gradient that supplies the free energy for the instability. This gradient normally corresponds to a stationary profile that is kept constant over the plasma by external means. This is a strong simplification, since in a real plasma the profile is set by the balance between the particle sources and the turbulent flux, and the turbulence, in turn, relaxes the very gradient that drives it[24, 25]. This feedback is the basis of predator–prey descriptions of drift-wave turbulence, in which the gradient plays the role of the prey, fed by an external source, and the drift waves play the role of the predators, growing by consuming it, so that the two influence each other and give rise to regular or chaotic oscillations[19, 26, 27].

A natural way to study these ideas is through low-dimensional truncations. Keeping only a few Fourier modes that satisfy the resonance condition k1+k2+k3=0k_1+k_2+k_3=0 yields a triad, the elementary unit of nonlinear energy transfer, whose dynamics is described by a small set of ordinary differential equations[4, 5, 22]. Despite its simplicity, such a system allows tools that are not suitable for full turbulence simulations, such as bifurcation analysis, Lyapunov exponents, and period maps. In the adiabatic limit the HW triad reduces to the three-wave truncation of the HM equation, whose dynamics is well characterized. Keeping the first correction beyond this limit leads to a triad with growth and damping rates, which has been studied with these rates introduced phenomenologically. In the HW system, these rates follow directly from the resistive coupling and the density gradient, and the effect of letting the gradient evolve self-consistently, together with the zonal character of the long-wavelength mode, has not been addressed.

In this work we couple the HW triad to a time-dependent density gradient, which evolves self-consistently, fed by a particle source and eroded by the turbulent particle flux. To isolate the role of the zonal flow, we compare two triads that differ only by a small tilt of the long-wavelength mode. In one triad, it is purely zonal and cannot be driven by the density gradient, while in the other it is slightly tilted off the zonal axis and behaves as a weakly driven drift wave. Each triad is studied with both a fixed and a self-consistent gradient. Our main finding is that the zonal triad is almost always regular, whereas the non-zonal triad becomes chaotic over a sizeable region of the parameter space, and that the feedback of the flux on the gradient enlarges this region. We also identify the routes to chaos of the non-zonal triad, none of which exists when the long-wavelength mode is zonal, so the zonal flow acts as a mechanism that reduces chaos.

In Section II the Hasegawa–Wakatani model is presented and its wave expansion is obtained, together with the inclusion of the non-constant density background, and the linear stability of the modes is explored. Then, in Section III, the constants and parameters used are presented, together with the zonal and non-zonal triads and the numerical methods. Later, in Section IV, a reference case is defined and explored, with the background density gradient kept constant, comparing the two triads. In Section V the background density gradient is made non-stationary, which allows oscillations and chaotic and periodic regions intermixed in parameter space, and the comparison between the triads is repeated. Section VI then explores the routes to chaos of the non-zonal triad, which are absent when the long-wavelength mode is zonal. Finally, in Section VII, the conclusions of the work are presented.

Model

Hasegawa–Wakatani equations

The Hasegawa–Wakatani (HW) model[15, 16, 28] describes the edge turbulence of a plasma, where a steady background density in the radial direction is normally assumed, and the fluctuations lie in the plane xx–yy perpendicular to a uniform and constant magnetic field B=Bz^\mathbf{B}=B\hat{\mathbf{z}}. The spatial variable xx stands for the radial direction and yy for the poloidal direction. Lengths are normalized to the ion sound gyroradius ρs=cs/ωci\rho_s=c_s/\omega_{ci}, where cs=Te/mic_s=\sqrt{T_e/m_i} is the ion sound speed, TeT_e the electron temperature (in energy units) and mim_i the ion mass, and time tt is normalized to the inverse ion cyclotron frequency ωci−1\omega_{ci}^{-1}, with ωci=eB/mi\omega_{ci}=eB/m_i. It is an improvement on the Hasegawa–Mima equation: besides the fluctuation of the electrostatic potential ϕ\phi, it also includes the variation of the plasma density nn, which allows a better development of the turbulence and a more precise description of the interaction between the potential and the ions. The electrostatic potential ϕ\phi and the density fluctuation nn obey

∂∂t∇2ϕ+(z^×∇ϕ)⋅∇∇2ϕ=C(ϕ−n)+ν∇4ϕ(1)\frac{\partial}{\partial t}\nabla^2\phi+(\hat{\mathbf{z}}\times\nabla\phi)\cdot\nabla\nabla^2\phi=C(\phi-n)+\nu\nabla^4\phi \tag*{(1)}
∂∂tn+(z^×∇ϕ)⋅∇n+κ∂ϕ∂y=C(ϕ−n)+D∇2n(2)\frac{\partial}{\partial t}n+(\hat{\mathbf{z}}\times\nabla\phi)\cdot\nabla n+\kappa\frac{\partial\phi}{\partial y}=C(\phi-n)+D\nabla^2n \tag*{(2)}

Here CC is the adiabaticity parameter, defined below, and ν\nu and DD are the viscosity and the particle diffusion coefficients, which damp the vorticity and the density fluctuations, respectively. These equations display a similar structure to the HM equation, with ∇2ϕ\nabla^2\phi being the vorticity of the plasma, and the nonlinear interactions are given by (z^×∇ϕ)⋅∇(\hat{\mathbf{z}}\times\nabla\phi)\cdot\nabla operator, describing the advection of the multiplied quantity by the E×BE\times B drift velocity, with z^\hat{\mathbf{z}} being the unit vector parallel to the magnetic field. The background density gradient κ≡−dln⁡n0/dx\kappa\equiv-\mathrm{d}\ln n_0/\mathrm{d}x drives the instability through the term κ∂yϕ\kappa\partial_y\phi.

The coupling between the potential and the density are given by the adiabaticity parameter CC, that can be written as

C=Tek∥2e2n0η∥ωci(3)C=\frac{T_e k_\parallel^2}{e^2n_0\eta_\parallel\omega_{ci}} \tag*{(3)}

measures the parallel electron response, where k∥k_\parallel is the parallel wave number and η∥\eta_\parallel the parallel resistivity.

When C→0C \to0, the potential and the density decouple and the Navier–Stokes equation is recovered, being called the hydrodynamic regime. On the other hand, in the adiabatic limit, when C→∞C \to\infty, the coupling term dominates and ϕ~≃n~\tilde{\phi} \simeq\tilde{n} at the leading order, then recovering the Hasegawa-Mima equation. This case is normally associated with the electron parallel resistivity going to zero, causing the perturbation of the electrostatic potential to be in the phase with the fluctuation density. There, the system only presents drift waves propagating in the yy direction, and there is no linear instability to drive turbulence, since density and potential are in phase and the cross-field particle flux vanishes.

To achieve a more realistic construction of the model, we also include some small scale dissipation terms on the fluctuations through the viscosity coefficients ν\nu and DD, acting, respectively, on vorticity ∇2ϕ\nabla^2\phi and the density. These diffusion terms are mainly used for numerical purpose, to balance the exponential energy injection by the linear instability, and avoid the accumulation of energy which reaches the smallest scales by the direct enstrophy cascade.

Non constant density Gradient

The gradient κ\kappa is usually taken as a constant, which corresponds to a background profile maintained by external means. But to achieve a more realistic turbulence response, we can consider that the background density gradient is a function of time S(t)S(t), that is connected to already known predator–prey models in plasma turbulence[17, 19], considering it at the form

dSdt=P−λS(t)−Γk(4)\frac{dS}{dt}=P-\lambda S(t)-\Gamma_k \tag*{(4)}

where there is a pumping intensity PP that can be related to an external constant forcing, such as constant beam injection, or a constant depletion of ions into the wall. The rate λ\lambda describes the relaxation of the gradient by transport channels other than the turbulent flux, such as collisional diffusion. The last term is the turbulent radial particle flux averaged over the entire 2D2D space,

Γk=⟨nvx⟩x,y(5)\Gamma_k=\langle nv_x\rangle_{x,y} \tag*{(5)}

where vx=−∂yϕv_x=-\partial_y\phi is the radial E×B\mathbf{E}\times\mathbf{B} velocity. Without fluctuations, SS relaxes to S0=P/λS_0=P/\lambda, the gradient set by the source alone and ⟨⋅⟩x,y\langle\cdot\rangle_{x,y} denotes averaging over both xx and yy directions. Using the Fourier decomposition of both fields, and summing over one half of the wave-vector plane, the modes k\mathbf{k} and −k-\mathbf{k} contribute equally, we can re-write this as

Γk=−2∑kx,kykyIm⁡{nkϕk∗}(6)\Gamma_k=-2\sum_{k_x,k_y}k_y\operatorname{Im}\{n_k\phi_k^*\} \tag*{(6)}

Then, with the new background density, the set of equations for the Hasegawa–Wakatani model becomes

∂∂t∇2ϕ+(z^×∇ϕ)⋅∇∇2ϕ=C(ϕ−n)+ν∇4ϕ(7)\frac{\partial}{\partial t}\nabla^2\phi+(\hat{z}\times\nabla\phi)\cdot\nabla\nabla^2\phi=C(\phi-n)+\nu\nabla^4\phi \tag*{(7)}
∂∂tn+(z^×∇ϕ)⋅∇n+S(t)∂ϕ∂y=C(ϕ−n)+D∇2n(8)\frac{\partial}{\partial t}n+(\hat{z}\times\nabla\phi)\cdot\nabla n+S(t)\frac{\partial\phi}{\partial y}=C(\phi-n)+D\nabla^2 n \tag*{(8)}
dSdt=P−λS(t)+2∑kkyIm⁡{ϕk∗nk}(9)\frac{dS}{dt}=P-\lambda S(t)+2\sum_k k_y\operatorname{Im}\{\phi_k^*n_k\} \tag*{(9)}

Wave Modes

To study the nonlinear interactions and energy transfer within the system, we can expand both ϕ\phi and nn in Fourier wave modes k\mathbf{k}, thus allowing us to simplify the nonlinear operator (z^×∇ϕ)⋅∇(\hat{z}\times\nabla\phi)\cdot\nabla to some constant multiplication in the wave space.

ϕ=∑kϕk(t)eik⋅rn=∑knk(t)eik⋅r(10)\phi=\sum_{\mathbf{k}}\phi_k(t)e^{i\mathbf{k}\cdot\mathbf{r}} \qquad n=\sum_{\mathbf{k}}n_k(t)e^{i\mathbf{k}\cdot\mathbf{r}} \tag*{(10)}

with ϕ−k=ϕk∗\phi_{-\mathbf{k}}=\phi_{\mathbf{k}}^* and n−k=nk∗n_{-\mathbf{k}}=n_{\mathbf{k}}^*, the HW equations become

dϕkdt+(Ak−Bk)ϕk=Ck2nk(t)+αkp,q(11)\frac{d\phi_k}{dt}+(A_k-B_k)\phi_k=\frac{C}{k^2}n_k(t)+\alpha_k^{\mathbf{p},\mathbf{q}} \tag*{(11)}
dnkdt+(Ak+Bk)nk=(C−iS(t)ky)ϕk+γkp,q(12)\frac{dn_k}{dt}+(A_k+B_k)n_k=(C-iS(t)k_y)\phi_k+\gamma_k^{\mathbf{p},\mathbf{q}} \tag*{(12)}

with the nonlinear terms

αkp,q=12∑p+q=kq2−p2k2(p×q)⋅z^ϕqϕp(13)\alpha_k^{\mathbf{p},\mathbf{q}}=\frac{1}{2}\sum_{\mathbf{p}+\mathbf{q}=\mathbf{k}}\frac{q^2-p^2}{k^2}(\mathbf{p}\times\mathbf{q})\cdot\hat{z}\phi_q\phi_p \tag*{(13)}
γkp,q=∑p+q=k(p×q)⋅z^ϕpnq(14)\gamma_k^{\mathbf{p},\mathbf{q}}=\sum_{\mathbf{p}+\mathbf{q}=\mathbf{k}}(\mathbf{p}\times\mathbf{q})\cdot\hat{z}\phi_p n_q \tag*{(14)}

and the linear coefficients

Ak=12[Dk2+C+(Ck2+νk2)](15)A_k=\frac{1}{2}\left[ Dk^2+C+\left(\frac{C}{k^2}+\nu k^2\right)\right] \tag*{(15)}
Bk=12[Dk2+C−(Ck2+νk2)](16)B_k=\frac{1}{2}\left[ Dk^2+C-\left(\frac{C}{k^2}+\nu k^2\right)\right] \tag*{(16)}

To analyze the fundamental resonant interactions driving the turbulence, we truncate this infinite-dimensional system to consider only three coupled interacting waves that satisfy the triad resonance condition k1+k2+k3=0\mathbf{k}_1+\mathbf{k}_2+\mathbf{k}_3=\mathbf{0}. Defining the geometric interaction coefficient VV as

V=(k1×k2)⋅z^=(k2×k3)⋅z^=(k3×k1)⋅z^(17)V=(\mathbf{k}_1\times\mathbf{k}_2)\cdot\hat{z}=(\mathbf{k}_2\times\mathbf{k}_3)\cdot\hat{z}=(\mathbf{k}_3\times\mathbf{k}_1)\cdot\hat{z} \tag*{(17)}

The truncated system evaluates to the following set of ordinary differential equations

dϕ1dt+(A1−B1)ϕ1=Ck12n1+Vk12(k32−k22)ϕ2∗ϕ3∗(18)\frac{d\phi_1}{dt}+(A_1-B_1)\phi_1=\frac{C}{k_1^2}n_1+\frac{V}{k_1^2}(k_3^2-k_2^2)\phi_2^*\phi_3^* \tag*{(18)}
dϕ2dt+(A2−B2)ϕ2=Ck22n2+Vk22(k12−k32)ϕ3∗ϕ1∗(19)\frac{d\phi_2}{dt}+(A_2-B_2)\phi_2=\frac{C}{k_2^2}n_2+\frac{V}{k_2^2}(k_1^2-k_3^2)\phi_3^*\phi_1^* \tag*{(19)}
dϕ3dt+(A3−B3)ϕ3=Ck32n3+Vk32(k22−k12)ϕ1∗ϕ2∗(20)\frac{d\phi_3}{dt}+(A_3-B_3)\phi_3=\frac{C}{k_3^2}n_3+\frac{V}{k_3^2}(k_2^2-k_1^2)\phi_1^*\phi_2^* \tag*{(20)}
dn1dt+(A1+B1)n1=(C−iSk1y)ϕ1+V(ϕ2∗n3∗−ϕ3∗n2∗)(21)\frac{dn_1}{dt}+(A_1+B_1)n_1=(C-iSk_{1y})\phi_1+V(\phi_2^*n_3^*-\phi_3^*n_2^*) \tag*{(21)}
dn2dt+(A2+B2)n2=(C−iSk2y)ϕ2+V(ϕ3∗n1∗−ϕ1∗n3∗)(22)\frac{dn_2}{dt}+(A_2+B_2)n_2=(C-iSk_{2y})\phi_2+V(\phi_3^*n_1^*-\phi_1^*n_3^*) \tag*{(22)}
dn3dt+(A3+B3)n3=(C−iSk3y)ϕ3+V(ϕ1∗n2∗−ϕ2∗n1∗)(23)\frac{dn_3}{dt}+(A_3+B_3)n_3=(C-iSk_{3y})\phi_3+V(\phi_1^*n_2^*-\phi_2^*n_1^*) \tag*{(23)}
dSdt=P−λS+2∑j=13kjyIm⁡{ϕj∗nj}.(24)\frac{dS}{dt}=P-\lambda S+2\sum_{j=1}^{3}k_{jy}\operatorname{Im}\{\phi_j^*n_j\}. \tag*{(24)}

Within this three-wave approximation, the Fourier expansion of the electrostatic potential is given by

ϕ(x,t)=∑j=13[ϕj(t)eikj⋅x+c.c.](25)\phi(x,t)=\sum_{j=1}^{3}\left[\phi_j(t)e^{i\mathbf{k}_j\cdot\mathbf{x}}+\mathrm{c.c.}\right] \tag*{(25)}
n(x,t)=∑j=13[nj(t)eikj⋅x+c.c.](26)n(x,t)=\sum_{j=1}^{3}\left[n_j(t)e^{i\mathbf{k}_j\cdot\mathbf{x}}+\mathrm{c.c.}\right] \tag*{(26)}

Equations (18)–(24) define a 13-dimensional real system: three complex amplitudes ϕj\phi_j, three complex amplitudes njn_j, and the real gradient SS. The dissipation removes any conserved quantity. The system is, however, invariant under the phase rotations

ϕj→ϕjeiθj,nj→njeiθj,θ1+θ2+θ3=0,(27)\phi_j\to\phi_j e^{i\theta_j},\qquad n_j\to n_j e^{i\theta_j},\qquad\theta_1+\theta_2+\theta_3=0, \tag*{(27)}

which leave the quadratic couplings and the flux unchanged. This two-parameter symmetry reduces the dynamics to eleven phase-invariant variables: the six moduli ∣ϕj∣|\phi_j| and ∣nj∣|n_j|, the three phase differences arg⁡nj−arg⁡ϕj\arg n_j-\arg\phi_j, the total phase Θ=arg⁡ϕ1+arg⁡ϕ2+arg⁡ϕ3\Theta=\arg\phi_1+\arg\phi_2+\arg\phi_3, and SS.

This truncated three-wave system forms the theoretical basis for our dynamical analysis. By systematically varying the parameter space for the coupling strength CC and the energy pump drive PP, we can simulate how the NBI (Neutral Beam Injection) energy pump impacts mode stability and induces structural bifurcations within the plasma edge.

Linear Stability Analysis

The truncated system admits the trivial equilibrium

ϕj=nj=0,S=S0≡Pλ,j=1,2,3,(28)\phi_j=n_j=0,\qquad S=S_0\equiv\frac{P}{\lambda},\qquad j=1,2,3, \tag*{(28)}

in which the gradient is set only by the balance between the source PP and the relaxation λS\lambda S. To study its stability we add small perturbations ϕj\phi_j, njn_j and S=S0+δSS=S_0+\delta S, and keep only the terms that are linear in them. The triad couplings and the particle flux are products of two fluctuations, so they are of second order and drop out. The linearized system is

dδSdt=−λδS,(29)\frac{d\delta S}{dt}=-\lambda\delta S, \tag*{(29)}
dϕjdt=−(Aj−Bj)ϕj+Ckj2nj,(30)\frac{d\phi_j}{dt}=-(A_j-B_j)\phi_j+\frac{C}{k_j^2}n_j, \tag*{(30)}
dnjdt=(C−iS0kjy)ϕj−(Aj+Bj)nj.(31)\frac{dn_j}{dt}=(C-iS_0k_{jy})\phi_j-(A_j+B_j)n_j. \tag*{(31)}

The perturbation of the gradient always decays at rate λ\lambda, and the three modes evolve independently of each other. At linear order they only share the value of S0S_0, and the coupling between modes is purely nonlinear. The stability of the fixed point is therefore decided mode by mode, by the pair (ϕj,nj)(\phi_j,n_j) and the diagonal coefficients are given by

Aj−Bj=Ckj2+νkj2,Aj+Bj=C+Dkj2,(32)A_j-B_j=\frac{C}{k_j^2}+\nu k_j^2,\qquad A_j+B_j=C+Dk_j^2, \tag*{(32)}

are the damping rates of ϕj\phi_j and njn_j when the two fields are not coupled, so AjA_j is the mean of the two damping rates and BjB_j is half their difference. In the rest of this section we drop the index jj.

Equations (30) and (31) form a linear system with constant coefficients,

ddt(ϕn)=M(ϕn),M=(−(A−B)C/k2C−iS0ky−(A+B)).(33)\frac{d}{dt}\begin{pmatrix}\phi\\n\end{pmatrix} = M\begin{pmatrix}\phi\\n\end{pmatrix}, \qquad M= \begin{pmatrix} -(A-B) & C/k^2\\ C-iS_0k_y & -(A+B) \end{pmatrix}. \tag*{(33)}

Substituting (ϕ,n)=(ϕ^,n^)eσt(\phi,n)=(\hat{\phi},\hat{n})e^{\sigma t}, with ϕ^\hat{\phi} and n^\hat{n} constant, gives M(ϕ^,n^)T=σ(ϕ^,n^)TM(\hat{\phi},\hat{n})^{\mathrm{T}}=\sigma(\hat{\phi},\hat{n})^{\mathrm{T}}, so σ\sigma is an eigenvalue of MM. A nonzero solution exists only if det⁡(M−σI)=0\det(M-\sigma I)=0, that is,

(σ+A−B)(σ+A+B)−Ck2(C−iS0ky)=0.(34)(\sigma+A-B)(\sigma+A+B)-\frac{C}{k^2}(C-iS_0k_y)=0. \tag*{(34)}

The first product is (σ+A)2−B2(\sigma+A)^2-B^2, so this quadratic equation has the two roots

σ=−A±Δ,Δ=B2+C2k2−iCS0kyk2.(35)\sigma=-A\pm\sqrt{\Delta},\qquad\Delta=B^2+\frac{C^2}{k^2}-i\frac{CS_0k_y}{k^2}. \tag*{(35)}

Because of the drive term, MM is complex and so is σ\sigma. We write σ=γ−iω\sigma=\gamma-i\omega, so that the perturbation behaves as eγte−iωte^{\gamma t}e^{-i\omega t}. Together with the spatial factor eik⋅xe^{i\mathbf{k}\cdot\mathbf{x}}, each mode is a wave eγtei(k⋅x−ωt)e^{\gamma t}e^{i(\mathbf{k}\cdot\mathbf{x}-\omega t)} with frequency ω\omega and growth rate γ\gamma, and it is unstable when γ>0\gamma>0.

To separate γ\gamma and ω\omega we split Δ\Delta into its real and imaginary parts. From the definitions of AA and BB,

A2−B2=(A−B)(A+B)=C2k2+G,(36)A^2-B^2=(A-B)(A+B)=\frac{C^2}{k^2}+G, \tag*{(36)}
G≡CD+νCk2+νDk4≥0,(37)G\equiv CD+\nu Ck^2+\nu Dk^4\geq0, \tag*{(37)}

where GG collects all the dissipative terms. Then Δ=X−iY\Delta= X-iY, where

X=A2−G,Y=CS0kyk2.(38)X=A^2-G,\qquad Y=\frac{CS_0k_y}{k^2}. \tag*{(38)}

are real. XX does not depend on the gradient, while YY is proportional to it.

The growth rate and the frequency are the real and imaginary parts of σ=−A±Δ\sigma=-A\pm\sqrt{\Delta}, so we need the square root of the complex number Δ\Delta. We write it as Δ=a−ib\sqrt{\Delta}=a-ib, with aa and bb real and a≥0a\geq0. Squaring and comparing with Δ=X−iY\Delta=X-iY gives

a2−b2=X,2ab=Y.(39)a^2-b^2=X,\qquad2ab=Y. \tag*{(39)}

Substituting b=Y/2ab=Y/2a in the first equation gives a quadratic equation for aa, a4−Xa2−Y2/4=0a^4-Xa^2-Y^2/4=0, whose positive solution is

a=X+X2+Y22,b=Y2a.(40)a=\sqrt{\frac{X+\sqrt{X^2+Y^2}}{2}},\qquad b=\frac{Y}{2a}. \tag*{(40)}

The two roots are therefore σ=−A+a−ib\sigma=-A+a-ib and σ=−A−a+ib\sigma=-A-a+ib. The second one has γ=−A−a<0\gamma=-A-a<0 and is always damped. The first one is the drift wave, with growth rate and frequency

γ=−A+a,ω=b=Y2a.(41)\gamma=-A+a,\qquad\omega=b=\frac{Y}{2a}. \tag*{(41)}

Since a>0a>0, the frequency has the sign of YY, that is, of S0kyS_0k_y.

The mode grows when a>Aa>A. Without a gradient, S0=0S_0=0, we have Y=0Y=0 and a=Xa=\sqrt{X}. Since X=A2−GX=A^2-G and G≥0G\geq0, this gives a≤Aa\leq A, so every mode is damped, as expected in the absence of an energy source. As the gradient increases, ∣Y∣|Y| increases, and so does aa, since YY enters only through X2+Y2\sqrt{X^2+Y^2}. The growth rate therefore increases with the drive ∣S0ky∣|S_0k_y|, and the mode becomes unstable when aa reaches AA. Setting a=Aa=A in the expression for aa gives

X+X2+Y2=2A2.(42)X+\sqrt{X^2+Y^2}=2A^2. \tag*{(42)}

Isolating the square root and squaring, the terms X2X^2 cancel, and

Y2=4A2(A2−X)=4A2G.(43)Y^2=4A^2\left(A^2-X\right)=4A^2G. \tag*{(43)}

With Y=CS0ky/k2Y=CS_0k_y/k^2, mode kk is unstable when ∣S0∣|S_0| exceeds the threshold

Sc(k)=2Ak2C∣ky∣G,G=CD+νCk2+νDk4.(44)S_c(k)=\frac{2Ak^2}{C|k_y|}\sqrt{G},\qquad G=CD+\nu Ck^2+\nu Dk^4. \tag*{(44)}

The threshold is the ratio between the dissipation, contained in GG, and the efficiency C∣ky∣/k2C|k_y|/k^2 with which the gradient drives the mode. In the ideal limit D=ν=0D=\nu=0 we have G=0G=0 and Sc=0S_c=0, which recovers the classical result that any gradient destabilizes the drift wave.

The factor CC in the denominator shows that the drive requires the coupling between density and potential, and for C=0C=0 no mode can be destabilized. Modes with ky=0k_y=0 have Sc→∞S_c\to\infty and are never driven, since the drive term S∂ϕ/∂yS\partial\phi/\partial y vanishes. This also holds in the modified model used for the zonal mode (§ 4), where the coupling CC is removed for ky=0k_y=0. In that case MM is diagonal, with entries −νk2-\nu k^2 and −Dk2-Dk^2, and the mode is simply damped.

At the threshold, a=Aa=A and S0=ScS_0=S_c, and Eq. (41) gives the frequency of the marginal wave,

ωc=CScky2Ak2.(45)\omega_c=\frac{CS_ck_y}{2Ak^2}. \tag*{(45)}

In the triad, each mode has its own threshold, Sc(k1)S_c(k_1), Sc(k2)S_c(k_2) and Sc(k3)S_c(k_3), and the fixed point of Eq. (28) becomes unstable as soon as one of them is exceeded. We call the lowest of the three thresholds SthS_{\mathrm{th}}, the threshold of the triad. Which mode sets it depends on the triad geometry and on CC (§ 4). Since S0=P/λS_0=P/\lambda, the fixed point loses stability when the source exceeds

Pc=λSth.(46)P_c=\lambda S_{\mathrm{th}}. \tag*{(46)}

Both SthS_{\mathrm{th}} and PcP_c depend on CC through AA and GG, and the curve Pc(C)P_c(C) is the linear stability boundary in the (C,P)(C,P) plane.

Figure 1 shows these results for the parameters of § 4. Panels (a) and (b) give the growth rate γ\gamma of each mode as a function of CC at a fixed gradient, S0=1.5S_0=1.5. A mode is unstable where its curve lies above zero and stable where it lies below. Each drift wave is unstable only within a window of CC. For small CC the drive, which is proportional to CC, vanishes and the dissipation dominates. For large CC the fields approach the adiabatic response n≈ϕn\approx\phi, the flux that sustains the instability vanishes, and the mode is again damped. In between, γ\gamma has a maximum, near C≈0.4C\approx0.4 for modes 1 and 2. In the non-zonal triad, mode 1 is unstable for 0.038<C<2.080.038<C<2.08 and mode 2 for 0.044<C<1.600.044<C<1.60. In the zonal triad, mode 1 is the same, while mode 2 is unstable for 0.083<C<1.280.083<C<1.28, because k2yk_{2y} differs. Mode 3 behaves differently in the two triads. In the zonal triad it is damped for every CC, at the rate νk32=0.025\nu k_3^2=0.025. In the non-zonal triad it is weakly driven and is unstable only for C<0.144C<0.144 at this S0S_0.

Figure 1. Linear stability of the trivial equilibrium for D=ν=0.1D=\nu=0.1. Growth rate γj=Re⁡σj\gamma_j=\operatorname{Re}\sigma_j of the drift-wave root of Eq. (35) for each mode jj as a function of CC, at fixed S0=1.5S_0=1.5, for (a) the non-zonal triad and (b) the zonal triad. The shaded band marks γ>0\gamma>0 (instability). In the zonal triad, mode 3 has ky=0k_y=0 and the adiabatic coupling is removed for it, so it is damped at the rate νk32\nu k_3^2 for every CC. (c) Marginal stability curves Sc(kj)S_c(k_j), Eq. (44), for the modes of the non-zonal triad. A mode is unstable above its curve, and the shaded region is unstable for the triad. The dotted horizontal line marks S0=1.5S_0=1.5, the value used in (a) and (b). The threshold of the zonal triad is that of mode 1, and it coincides with the threshold of the non-zonal triad for C≳0.065C\gtrsim0.065.

Panel (c) shows the same information in the (C,S0)(C,S_0) plane, through the marginal curves Sc(kj)S_c(k_j), of Eq. (44). A mode is unstable for S0S_0 above its curve, so the unstable region of the triad lies above the lowest curve, Sth(C)S_{\mathrm{th}}(C). The horizontal line S0=1.5S_0=1.5 crosses the curves at the edges of the windows in panels (a) and (b). The threshold has a minimum, Sc≈0.85S_c\approx0.85 at C≈0.23C\approx0.23, set by mode 1, so for S0S_0 below this value the fixed point is stable for every CC. The thresholds of modes 1 and 3 cross at C≈0.065C\approx0.065. For smaller CC the long-wavelength mode of the non-zonal triad is the first to become unstable, and its threshold falls below that of mode 1, which sets the threshold of the zonal triad.

We now compute the particle flux carried by a linear mode, which enters the equation for SS. In a linear mode, nn and ϕ\phi are not independent. Setting dϕ/dt=σϕ\mathrm{d}\phi/\mathrm{d}t=\sigma\phi in Eq. (30) gives

n=k2C(σ+A−B)ϕ(47)n=\frac{k^2}{C}(\sigma+A-B)\phi \tag*{(47)}

so the density is the potential multiplied by a complex factor, which sets the phase shift between the two fields. Multiplying by ϕ∗\phi^* and taking the imaginary part, only Im⁡σ=−ω\operatorname{Im}\sigma=-\omega survives, since A−BA-B is real, and

Im⁡{ϕ∗n}=−k2Cω∣ϕ∣2.(48)\operatorname{Im}\{\phi^*n\}=-\frac{k^2}{C}\omega\lvert\phi\rvert^2. \tag*{(48)}

The contribution of mode jj to the flux in Eq. (4) is then

Γj=−2kjyIm⁡{ϕj∗nj}=2kjykj2Cωj∣ϕj∣2.(49)\Gamma_j=-2k_{jy}\operatorname{Im}\{\phi_j^*n_j\}=\frac{2k_{jy}k_j^2}{C}\omega_j\lvert\phi_j\rvert^2. \tag*{(49)}

Since ωj\omega_j has the sign of S0kjyS_0k_{jy}, the product kjyωjk_{jy}\omega_j is positive for S0>0S_0>0, and so is Γj\Gamma_j. The flux therefore always reduces SS. The unstable wave carries particles down the gradient and weakens the gradient that drives it, which is the predator–prey mechanism of the model. This flux exists only because nn and ϕ\phi are out of phase. For the adiabatic response n=ϕn=\phi, Im⁡{ϕ∗n}=0\operatorname{Im}\{\phi^*n\}=0 and there is no transport.

This feedback also determines what happens just above the threshold. Consider a state in which only one mode jj is excited. All triad couplings vanish, since each of them is a product of two different modes, so the mode remains alone for all times. It obeys Eqs. (30) and (31) with S(t)S(t) in place of S0S_0, while SS obeys Eq. (4) with the flux Γj\Gamma_j.

The gradient is then pushed towards the threshold of the mode from both sides. If S>Sc(kj)S>S_c(k_j), the mode grows, the flux increases and SS decreases. If S<Sc(kj)S<S_c(k_j), the mode decays, the flux decreases and the source increases SS. The system reaches a steady state with

S=Sc(kj),(50)S=S_c(k_j), \tag*{(50)}

in which the mode neither grows nor decays, and rotates at the frequency ωc\omega_c with constant amplitude. The amplitude follows from dS/dt=0\mathrm{d}S/\mathrm{d}t=0 in Eq. (4), that is, P=λSc=ΓjP=\lambda S_c=\Gamma_j. With ωj=ωc\omega_j=\omega_c, Eqs. (45) and (49) give Γj=kjy2Sc∣ϕj∣2/Aj\Gamma_j=k_{jy}^2S_c\lvert\phi_j\rvert^2/A_j, and therefore

∣ϕj∣2=Aj[P−λSc(kj)]kjy2Sc(kj).(51)\lvert\phi_j\rvert^2=\frac{A_j[P-\lambda S_c(k_j)]}{k_{jy}^2S_c(k_j)}. \tag*{(51)}

This result has three consequences. First, for the mode with the lowest threshold, λSc(kj)=Pc\lambda S_c(k_j)=P_c, the amplitude is zero at P=PcP=P_c and grows as P−Pc\sqrt{P-P_c}. Second, in this single-mode state only the phases rotate, as allowed by the phase symmetry of Eq. (27), while all amplitudes and SS are constant. In the period maps, which are based on amplitudes, this orbit appears as an equilibrium. Third, above threshold the gradient no longer increases with PP. It remains clamped at ScS_c, and the additional source is removed by the flux, through the growth of ∣ϕj∣\lvert\phi_j\rvert. This state exists only with a self-consistent gradient. For a fixed SS, nothing limits the growth of the unstable mode except the triad coupling. The stability of the single-mode state, and the dynamics that appears when it is lost, are discussed in § 6.

Parameters and Numerical Methods

In order to explore the effects of the change from a constant to a time dependent background density gradient, we need to set up some numerical parameters.

All quantities are in the standard Hasegawa–Wakatani normalization, with lengths in units of the ion sound gyroradius ρs\rho_s and time in units of the inverse ion cyclotron frequency ωci−1\omega_{ci}^{-1}. In these units, k∼1k \sim1 corresponds to the drift-wave scale kρs∼1k\rho_s \sim1, where the linear drive is strongest.

The dissipation coefficients are fixed at D=ν=0.1D=\nu=0.1. Equal values keep the two fields on the same dissipative range, so that differences between the ϕ\phi and nn dynamics come only from the adiabatic coupling and the gradient drive.

The adiabaticity parameter CC sets the strength of the parallel electron response, which couples ϕ\phi and nn through the term C(ϕ−n)C(\phi-n). It is varied over C∈[0.03,5]C \in[0.03,5] with logarithmic spacing. The lower end lies on the hydrodynamic side, where density and potential are weakly coupled and the dynamics approaches two-dimensional fluid turbulence. The upper end approaches the adiabatic regime, where n≈ϕn \approx\phi and the fluctuations behave as nearly drift waves.

The strict limit C→∞C \to\infty is not considered. In this limit n→ϕn \to\phi, the linear growth rate vanishes as C−1C^{-1}, the flux ∝Im⁡{ϕj∗nj}\propto\operatorname{Im}\{\phi_j^*n_j\} vanishes, and the fixed and self-consistent gradient cases become identical. The triad then reduces to the three-wave truncation of the Hasegawa–Mima equation, whose dynamics has already been characterized. Retaining the first correction in C−1C^{-1} gives the same triad with growth and damping rates γj\gamma_j, which is the model studied by Batista et al.⁵, where these rates were introduced phenomenologically. In the present model they follow from CC, DD, ν\nu and SS. This correspondence requires C≳20C \gtrsim20 for the parameters used here, beyond the scanned range.

The study compares four configurations, obtained by combining two triad geometries with two treatments of the background gradient. The first comparison, between a zonal and a non-zonal triad, isolates the role of a long-wavelength mode that cannot be driven by the gradient. The second, between a fixed and a self-consistent gradient, isolates the role of the feedback of the turbulent flux on the drive.

In the self-consistent case, SS evolves according to Eq. (4) with relaxation rate λ=0.5\lambda=0.5. In the absence of fluctuations, this equation has the fixed point S0=P/λS_0=P/\lambda, so the source PP sets the gradient that the triad would experience without any turbulent flux. The source is scanned over P∈[0.25,3]P \in[0.25,3], which corresponds to S0∈[0.5,6]S_0 \in[0.5,6]. In the fixed case, S˙=0\dot{S}=0, PP and λ\lambda drop out, and SS is scanned directly over the same interval, S∈[0.5,6]S \in[0.5,6]. The two maps therefore cover the same region, with SS in one corresponding to S0S_0 in the other. They also share the same linear stability boundary, since the condition P>Pc=λSthP>P_c=\lambda S_{\mathrm{th}} derived previously is equivalent to S0>SthS_0>S_{\mathrm{th}}. Any difference between the fixed and self-consistent maps is therefore nonlinear. With SS fixed, the flux carried by the unstable modes has no effect on the drive, and the amplitudes saturate only through the triad transfer to damped modes. With SS dynamic, the flux erodes the gradient that drives it, pulling SS towards marginal stability.

The three wave vectors satisfy the resonance condition k1+k2+k3=0\mathbf{k}_1+\mathbf{k}_2+\mathbf{k}_3=0, which implies the cyclic identity k1×k2=k2×k3=k3×k1≡V\mathbf{k}_1\times\mathbf{k}_2=\mathbf{k}_2\times\mathbf{k}_3=\mathbf{k}_3\times\mathbf{k}_1\equiv V. The zonal triad is

k1=(0,1.125),k2=(−0.5,−1.125),k3=(0.5,0),(52)\mathbf{k}_1=(0,1.125),\qquad\mathbf{k}_2=(-0.5,-1.125),\qquad\mathbf{k}_3=(0.5,0), \tag*{(52)}

with ∣k1∣=1.125|\mathbf{k}_1|=1.125, ∣k2∣≈1.231|\mathbf{k}_2|\approx1.231, ∣k3∣=0.5|\mathbf{k}_3|=0.5 and V=0.5625V=0.5625. Modes 1 and 2 are drift waves with kyρs∼1k_y\rho_s\sim1, near the scale of maximal linear growth, while mode 3, with k3y=0k_{3y}=0 and a longer wavelength, plays the role of a zonal flow. Since the drive term −iSk3yϕ3-iS k_{3y}\phi_3 vanishes, mode 3 is linearly damped for every SS, and since its contribution to the flux is proportional to k3yIm⁡{ϕ3∗n3}k_{3y}\operatorname{Im}\{\phi_3^*n_3\}, it does not enter Eq. (4). It gains energy only through the triad transfer out of modes 1 and 2. The linear threshold is set by mode 1, with Sc≈1.74, 0.96, 1.13S_c\approx1.74,\,0.96,\,1.13 and 2.222.22 at C=0.03C=0.03, 0.10.1, 11 and 55. Following the modified Hasegawa–Wakatani model²⁴, the adiabatic coupling is removed for the zonal mode, since components with ky=0k_y=0 are constant on a flux surface and have no parallel electron response. Mode 3 is then damped only by νk32\nu k_3^2 and Dk32Dk_3^2, and ϕ3\phi_3 and n3n_3 decouple at linear order. Retaining CC for ky=0k_y=0 is known to artificially suppress zonal flows. Because mode 3 does not set the threshold, the stability boundary is the same in the standard and modified models.

The non-zonal triad is

k1=(0,1.125),k2=(−0.5,−1.0),k3=(0.5,−0.125),(53)\mathbf{k}_1=(0,1.125),\qquad\mathbf{k}_2=(-0.5,-1.0),\qquad\mathbf{k}_3=(0.5,-0.125), \tag*{(53)}

with ∣k1∣=1.125|\mathbf{k}_1|=1.125, ∣k2∣≈1.118|\mathbf{k}_2|\approx1.118, ∣k3∣≈0.515|\mathbf{k}_3|\approx0.515 and again V=0.5625V=0.5625. It is constructed to differ from the zonal triad as little as possible, keeping k1\mathbf{k}_1, the coupling coefficient, and nearly the length of k3\mathbf{k}_3, which is only tilted slightly off the zonal axis. All three modes are now linearly driven and all contribute to the flux. Mode 3 is weakly driven, with Sc(k3)S_c(k_3) increasing from 0.930.93 at C=0.03C=0.03 to 8.088.08 at C=5C=5. For C≳0.065C\gtrsim0.065 mode 1 remains the first to become unstable, and the stability boundary coincides with that of the zonal triad. For C≲0.065C\lesssim0.065, mode 3 becomes the most unstable mode and the threshold falls below the zonal one, down to Sc≈0.93S_c\approx0.93 at C=0.03C=0.03 against 1.741.74 for the zonal triad. This is the only region in which the two geometries differ already at the linear level. Elsewhere, differences between them are due to whether the long-wavelength mode can extract energy from the gradient and carry flux (k3y≠0k_{3y}\ne0) or can only receive energy through the triad transfer (k3y=0k_{3y}=0).

The combination of the non-zonal triad with a fixed gradient deserves a separate remark. When SS exceeds all three thresholds Sc(kj)S_c(k_j), which within the scanned range occurs for C≲2.7C\lesssim2.7, every mode of the triad is linearly unstable, and no damped mode is available to absorb the energy redistributed by the nonlinear coupling. The amplitudes can then grow without bound. This cannot occur in the zonal triad, where mode 3 is always damped, nor with a self-consistent gradient, where the flux reduces SS.

The wave modes for ϕ\phi and nn are complex numbers, producing then a 12-dimensional phase space, and the variation of SS add up to 13-dimensions. Because of the frictions present, there are no conserved quantities that can be used to reduce the number of degrees of freedom, and we expect very complex behavior to be observed in the numerical simulations.

For all the simulations, we fix the initial conditions at Re⁡{ϕj}=Re⁡{nj}=0.01\operatorname{Re}\{\phi_j\} = \operatorname{Re}\{n_j\} = 0.01 with Im⁡{ϕj}=Im⁡{nj}=0\operatorname{Im}\{\phi_j\} = \operatorname{Im}\{n_j\} = 0 and, in the self-consistent case, S(0)=1S(0) = 1. All the equations are integrated with an adaptive fifth-order Dormand–Prince scheme[29], with absolute and relative tolerances of 10−610^{-6}. The integration runs up to T=2.5×105T = 2.5 \times10^5.

To explore the parameter space, we compute period maps on a 1024×10241024 \times1024 grid. For a fixed gradient the grid covers the (S,C)(S,C) plane and for a self-consistent gradient the (P,C)(P,C) plane, with S∈[0.5,6]S \in[0.5,6] or P∈[0.25,3]P \in[0.25,3] on a linear scale and C∈[0.03,5]C \in[0.03,5] on a logarithmic one. All points start from the initial condition given above. The asymptotic state is classified from the local maxima of an observable in the last 10410^4 time units, using ∣ϕ1∣|\phi_1| for a fixed gradient and SS for a self-consistent one. A point is period-pp (p≤10p \le10) when both the peak heights and the intervals between peaks repeat with lag pp over eight consecutive cycles, within a relative tolerance of 10−210^{-2}. When the observable varies by less than 10−410^{-4} in the window, the point is classified as an equilibrium. Below the linear threshold of Eq. (44) this is the quiescent state, while above it the mode amplitudes are constant and only their phases rotate. All other points are aperiodic. Points where any amplitude exceeds 10610^6 are marked as divergent.

Constant Background Density Gradient

For comparison with a simpler case, we first consider a constant background gradient, dS/dt=0\mathrm{d}S/\mathrm{d}t = 0. The gradient is then an external control parameter. It sets the linear drive of each mode, but the particle flux carried by the fluctuations has no effect on it. Once a mode is linearly unstable, the only mechanism that can limit its growth is the nonlinear transfer of energy to the other modes of the triad. This case also isolates the first question of this work, namely how the dynamics changes when the long-wavelength mode of the triad is zonal (k3y=0k_{3y}=0) or non-zonal (k3y≠0k_{3y} \ne0).

Figure 2 summarizes the results. Panels (a) and (b) are the period maps of ∣ϕ1∣|\phi_1| in the (S,C)(S,C) plane for the zonal and non-zonal triads, and the dashed line is the linear threshold SthS_{\mathrm{th}}. We use mode 1 as the observable because it has the lowest threshold over most of the plane, so it is the first to become unstable, and the main channel through which energy enters the triad. Panels (c)–(f) show the amplitudes ∣ϕj∣|\phi_j| and ∣nj∣|n_j| of the three modes at C=0.08C = 0.08 and S=1.5S = 1.5, the point marked in the maps.

Figure 2. Dynamics with a fixed gradient SS. (a), (b) Period of ∣ϕ1∣|\phi_1| in the (S,C)(S, C) plane for the zonal and non-zonal triads, respectively. Div.: divergent; Eq.: equilibrium (constant amplitude); Ap.: aperiodic; numbers give the period. The dashed line is the linear threshold SthS_{\mathrm{th}} and the red dot marks (S,C)=(1.5,0.08)(S, C) = (1.5, 0.08). (c), (d) Time series of ∣ϕj∣|\phi_j| and ∣nj∣|n_j| for the zonal triad at the marked point. (e), (f) Same for the non-zonal triad. Solid dark line: j=1j = 1; purple dashed line: j=2j = 2; pink dotted line: j=3j = 3.

In the zonal triad, Figure 2(a), the threshold is followed by a band of period-1 motion, whose width in SS grows as CC decreases. For stronger drive the amplitudes become constant and the energy exchange between the modes stops, so these points fall in the equilibrium class, which covers most of the plane. Aperiodic motion appears only in a thin strip along the threshold, where the amplitudes are slowly modulated, and it occupies a very small part of the map. No extended aperiodic region is found for any value of CC.

The time series at the marked point, Figure 2(c) and (Figure 2d), show the structure of the period-1 orbit. The amplitude of mode 1 oscillates between 0.2 and 1.3. Modes 2 and 3 oscillate almost in phase with each other and in antiphase with mode 1, so when ∣ϕ1∣|\phi_1| reaches a minimum, ∣ϕ2∣|\phi_2| and ∣ϕ3∣|\phi_3| reach their maxima, and vice versa. This is the classical energy exchange of a resonant triad. The drift wave k1k_1, fed by the gradient, transfers energy to k2k_2 and to the zonal mode k3k_3, which then return it. The density follows the same pattern with larger amplitudes, ∣nj∣≈5∣ϕj∣|n_j| \approx5|\phi_j|, which shows that at this value of CC the density and the potential are far from the adiabatic response n≈ϕn \approx\phi.

This regularity is a direct consequence of the linear properties of the zonal mode. Since k3y=0k_{3y} = 0, the drive term −iSk3yϕ3-\mathrm{i}S k_{3y}\phi_3 vanishes, and mode 3 cannot extract energy from the gradient. With the adiabatic coupling removed for ky=0k_y = 0, it is damped only by νk32\nu k_3^2 and Dk32D k_3^2, and it does not contribute to the flux. Mode 3 therefore acts as a passive reservoir that can only receive energy from the drift waves and return it to them. The triad has a single source of energy, mode 1, and a well-defined path for its redistribution, and the exchange locks into a fixed cycle or into constant amplitudes.

The non-zonal triad behaves very differently, Figure 2(b). Above threshold no constant-amplitude states are found, and the triad oscillates with period 1 over most of the plane. Aperiodic motion occupies a strip along the threshold for all CC, and a broad region for C≲0.2C \lesssim0.2 and 1≲S≲31 \lesssim S \lesssim3, interleaved with windows of period 2 to 10. Over the whole map, a small part of the points are aperiodic. For C≲0.04C \lesssim0.04 the non-zonal triad diverges. In this hydrodynamic limit mode 3 has the lowest threshold, and all three modes become linearly unstable at moderate SS, so no damped mode is left to absorb the energy transferred by the nonlinear coupling. In the zonal triad mode 3 is always damped, and no divergence occurs.

At the marked point, Figure 2(e) and (Figure 2f), the regular antiphase exchange of the zonal case is lost. Modes 1 and 2 still oscillate on the fast time scale, but with irregular amplitudes and without a fixed phase relation. Mode 3 now evolves on a slower time scale and shows large, intermittent excursions, reaching ∣ϕ3∣≈2.8|\phi_3| \approx2.8, about twice the amplitude of the drift waves, before relaxing again.

The non-zonal triad was built to differ as little as possible from the zonal one. The vector k1\mathbf{k}_1 and the coupling coefficient VV are unchanged, the length of k3\mathbf{k}_3 changes by a small factor. The only qualitative difference is that k3\mathbf{k}_3 is tilted slightly off the zonal axis, which gives it a finite k3yk_{3y}. This small tilt changes the role of mode 3 in two ways. It now couples linearly to the gradient, so it has its own frequency and growth or damping rate set by SS, and it now carries part of the flux. Instead of a passive reservoir, the triad has a third active mode that competes with the drift waves for the energy of the gradient. The energy can then enter the triad through more than one mode, and the phase relation that organized the exchange in the zonal case is lost.

The onset of chaos is related to the linear thresholds of the individual modes. At C=0.08C = 0.08, mode 1 becomes unstable at S=1.03S = 1.03 in both triads. In the non-zonal triad, modes 2 and 3 follow at S=1.13S = 1.13 and S=1.20S = 1.20, and the chaotic region, 1.04≲S≲1.631.04 \lesssim S \lesssim1.63, starts as soon as a second mode approaches its own threshold. In the zonal triad mode 2 becomes unstable only at S=1.52S = 1.52 and mode 3 is never driven, and even with two unstable modes the motion remains regular.

Self-consistent density gradient

We now let the gradient evolve according to Eq. (4). SS is no longer a control parameter. It is fed by the source PP, relaxes at rate λ\lambda, and is eroded by the particle flux carried by the unstable modes. This closes the loop between the drive and the fluctuations and gives the model a predator–prey structure, in which the gradient is the prey and the drift waves are the predators. Reduced models of this kind are usually written with phenomenological equations for the turbulence intensity, the zonal-flow energy and the gradient[17, 26, 30]. In the present model the fluctuations obey the truncated Hasegawa–Wakatani equations, and the only added ingredient is the balance of Eq. (4). The control parameter is now the source PP, which sets the gradient S0=P/λS_0 = P/\lambda the plasma would have in the absence of turbulence. The top axis of Figure 3 shows S0S_0, so the maps of Figure 2 and Figure 3 cover the same region.

Figure 3. Dynamics with a self-consistent gradient, λ=0.5\lambda= 0.5. (a), (b) Period of SS in the (P,C)(P,C) plane for the zonal and non-zonal triads, respectively; the top axis gives S0=P/λS_0 = P/\lambda. Colors and dashed line as in Figure 2. The red dot marks (P,C)=(1.5,0.19)(P,C) = (1.5, 0.19). (c), (d) Time series of ∣ϕj∣|\phi_j| and ∣nj∣|n_j| for the zonal triad at the marked point. (e), (f) Same for the non-zonal triad.

Figure 3 shows the period maps of SS in the (P,C)(P,C) plane, and the amplitudes of the three modes at C=0.19C = 0.19 and P=1.5P = 1.5 (S0=3S_0 = 3). We use the line C=0.19C = 0.19 to describe the transitions in detail.

In both triads the fixed point loses stability at the threshold PcP_c through a supercritical Hopf bifurcation, since the amplitude of the single-mode state grows as P−Pc\sqrt{P-P_c}, Eq. (51). The threshold is followed by a thin band classified as equilibrium, in which the amplitudes are constant. It has no counterpart in the non-zonal triad with a fixed gradient. Along C=0.19C = 0.19, for PP just above Pc=λSc(k1)≈0.427P_c = \lambda S_c(k_1) \approx0.427, the gradient is clamped at S=Sc(k1)S = S_c(k_1) and mode 1 rotates with constant amplitude. This state is stable only in a narrow interval. Mode 3 is excited at P≈0.451P \approx0.451 in the zonal triad and at P≈0.434P \approx0.434 in the non-zonal one. The energy taken from the gradient is now shared with the long-wavelength mode, and a steeper gradient is required to carry the same flux. The amplitudes start to oscillate at a secondary Hopf bifurcation, at P≈0.553P \approx0.553 for the zonal triad and P≈0.486P \approx0.486 for the non-zonal one, which marks the right edge of this band.

In the zonal triad, Figure 3(a), period-1 oscillations of SS fill most part of the map. The feedback leaves the triad regular. Aperiodic motion is present in a small portion of the parameter plane, confined to a small island around C≈0.2C \approx0.2 and P≈1.55P \approx1.55, surrounded by windows of period 4 and 9. A second region, of period 2, appears for P≳2.7P \gtrsim2.7 and 0.1≲C≲0.40.1 \lesssim C \lesssim0.4, and no divergence is found.

Along C=0.19C = 0.19 the maxima of SS keep increasing with the source, from S≈1.06S \approx1.06 at P=0.56P = 0.56 to S≈1.75S \approx1.75 at P=3P = 3. They exceed the threshold of mode 2 for P≳0.72P \gtrsim0.72, but always remain far below S0S_0. The marked point lies inside the aperiodic island. The time series, Figure 3(c) and (Figure 3d), show that the predator–prey cycle is superposed on the triad exchange. As with a fixed gradient, mode 1 oscillates in antiphase with modes 2 and 3, and it is the most energetic mode, with ⟨∣ϕ1∣2⟩=0.16\langle|\phi_1|^2\rangle= 0.16 against 0.09 and 0.08 for modes 2 and 3. The gradient, however, oscillates between S≈0.75S \approx0.75 and 1.51. In each cycle it crosses the thresholds of both drift waves, so modes 1 and 2 are alternately driven and damped. The time average ⟨S⟩=1.08\langle S\rangle= 1.08 is about a third of S0=3S_0 = 3. Averaged over the attractor, ⟨Γ⟩=P−λ⟨S⟩≈0.96\langle\Gamma\rangle= P - \lambda\langle S\rangle\approx0.96, so roughly two thirds of the source is carried by the turbulent flux, entirely by the two drift waves, since the zonal mode carries none.

The non-zonal triad, Figure 3(b), is the most complex of the four configurations. Aperiodic motion covers a third of the map and periodic orbits of period 2 to 10 another third, while period-1 motion is reduced to half of this. The aperiodic points form two large regions. The first lies at moderate adiabaticity, 0.45≲C≲2.50.45 \lesssim C \lesssim2.5 and 0.65≲P≲2.10.65 \lesssim P \lesssim2.1, bounded by a thin layer of period 6 and surrounded by broad regions of period 2 and 3. The second lies on the hydrodynamic side, C≲0.3C \lesssim0.3, and extends to the upper edge of the drive. In contrast to the fixed gradient, no divergence is found anywhere in the map.

The finely interleaved colors in parts of the map, for instance around the marked point, do not come from the classifier. They reflect coexisting attractors, reached from the same initial condition at neighboring parameters. At C=0.19C = 0.19 and P=1.5P = 1.5, a period-1 orbit of period T≈9.91T \approx9.91 coexists with a second periodic orbit of period T≈47.6T \approx47.6, which has six maxima of SS and five of ∣ϕ1∣|\phi_1| per cycle and is not a period-multiple of the first. Their basins are intermingled and have comparable size, and which orbit is reached from a given initial condition is sensitive to perturbations of the order of the integration error.

Along C=0.19C = 0.19, after the primary instability, the gradient rises to the threshold of mode 2, which is reached at P≈0.48P \approx0.48, and then remains nearly pinned there. The profile is therefore much stiffer than in the zonal triad. The time series at the marked point, Figure 3(e) and (Figure 3f), are strictly periodic. In this orbit SS stays below Sc(k3)S_c(k_3), so the long-wavelength mode is linearly damped at all times. It carries less than 2% of the flux, but it holds most of the energy, ⟨∣ϕ3∣2⟩=0.57\langle|\phi_3|^2\rangle= 0.57, about four times that of each drift wave. As long as the gradient stays below Sc(k3)S_c(k_3), the feedback makes mode 3 behave as the zonal mode does, as a damped reservoir fed only by the triad transfer.

This changes at P≈1.8P \approx1.8. The maxima of SS first exceed Sc(k3)S_c(k_3) at P≈1.82P \approx1.82, and at almost the same point the motion becomes aperiodic. The maxima of SS then spread over a wide band, from S≈0.35S \approx0.35 to 2.7 at P=3P = 3, and the amplitude of mode 3 reaches ∣ϕ3∣≈4.5|\phi_3| \approx4.5, four times that of the drift waves. The mechanism is the one identified for the fixed gradient, since the triad becomes irregular when the long-wavelength mode is linearly driven. The difference is that this now happens intermittently. During each excursion of the gradient above Sc(k3)S_c(k_3), mode 3 is driven, grows, and releases a burst of flux that pulls SS back below its threshold. Because the flux always acts to reduce SS, the amplitudes stay bounded.

The feedback does not simply suppress or enhance chaos. It moves chaos to a different region of the parameter space. With a fixed gradient, the non-zonal triad is chaotic in a narrow region just above threshold and mostly period-1 elsewhere. With a self-consistent gradient, the region near threshold has constant amplitudes or is periodic, and chaos appears at stronger drive and over a much wider range of CC. At C=0.08C = 0.08, for instance, the fixed-gradient triad is aperiodic for $1.04 \lesssim S \lesssim 1.63$, whereas the self-consistent one is aperiodic for P≳1.22P \gtrsim1.22 (S0≳2.4S_0 \gtrsim2.4), a region that contains a locked torus and a metastable chaotic set before the chaotic attractor proper at $P \gtrsim 1.32 ({xref}Sec. VI`). Over the whole map, the fraction of aperiodic points in the non-zonal triad grows from 6% with a fixed gradient to 31% with a self-consistent one.

In the zonal triad it remains below 1% in both cases. The feedback therefore amplifies the difference between the two geometries. A self-consistent gradient is strongly chaotic when the long-wavelength mode can extract energy from it, and remains essentially regular when that mode is zonal.

Routes to Chaos

Besides the regions of the parameter plane where chaos is found, it is relevant to know how it emerges to know whether the transition is gradual or abrupt, whether it is reversible or hysteretic, and which mechanism of the model is responsible for it. In a transport problem this matters directly, because a gradual transition gives an early warning, the dynamics becomes irregular slowly, as a parameter is varied, while an abrupt one does not. We therefore follow the transitions of the non-zonal triad for

both treatments of the gradient, with the adiabaticity parameter kept at C=0.08C = 0.08. The control parameter, SS for the fixed gradient and PP for the self-consistent one, is increased in small steps, each value starting from the final state of the previous one. For the self-consistent gradient the scan is repeated with PP decreasing, to detect coexisting attractors.

Each step is characterized by the following quantities, which are complementary. The first is the density of the maxima of an observable, ∣ϕ1∣|\phi_1| for the fixed gradient, SS for the self-consistent one, which gives a bifurcation diagram, a period-nn orbit has nn distinct maxima, while a quasi-periodic or a chaotic orbit fills an interval, so this shows the structure of the attractor. The second is the power spectral density (PSD) of the same observable, which shows how many incommensurate frequencies the motion contains, a periodic orbit appears as a fundamental line with its harmonics, a quasi-periodic orbit as additional lines at combinations of two frequencies, and chaos as a continuous background. The third is the largest Lyapunov exponent λ1\lambda_1, the mean exponential rate at which nearby trajectories separate, which separates regular motion (λ1=0\lambda_1 = 0) from chaotic motion (λ1>0\lambda_1 > 0). The phase symmetry of Eq. (27) leaves two Lyapunov exponents identically zero, since rotating the phases of a solution yields another solution, so the spectrum of the full system is that of the eleven phase-invariant variables plus two zeros. The second exponent therefore carries no information and is not shown.

These quantities say where and how the orbit changes, but not what the dynamics looks like just after chaos sets in. For that we use the laminar time. Close to some transitions the motion alternates between long, nearly regular laminar phases, during which the trajectory stays near the regular orbit that was lost, and short chaotic bursts. This is intermittency[31]. The mean duration of the laminar phases, ⟨τ⟩\langle\tau\rangle, diverges as the control parameter (here SS) approaches its critical value S∗S^*, ⟨τ⟩∝S−β\langle\tau\rangle\propto S^{-\beta}, and the exponent β\beta depends on the mechanism that destroys the regular orbit. In type-I intermittency, which follows a saddle-node bifurcation, β=1/2\beta= 1/2, while for types II and III, associated with a Hopf-type instability and with a period-doubling-type instability respectively, β=1\beta= 1. We use this analysis for two reasons. First, β\beta is a measurable signature of the mechanism, which allows the transition to be classified quantitatively, instead of only by eye from the bifurcation diagram and the PSD. Second, it separates soft transitions, where ⟨τ⟩\langle\tau\rangle diverges and the chaotic bursts appear gradually, from hard ones, where no laminar phases exist and ⟨τ⟩\langle\tau\rangle shows no scaling. In the present model the duration of the laminar phases is physically relevant, since during the chaotic bursts the particle flux, and therefore the gradient, changes strongly. Where the transition is hard, the relevant quantity is instead the lifetime of the chaotic transient and the width of the hysteresis loop, which we also report.

The rest of the section is organized by gradient treatment. For each triad we first look at the bifurcation diagram, the PSD and λ1\lambda_1 over the whole range, then at the enlargements around each transition, where, when the transition is soft, we also use the laminar time.

Fixed gradient

The zonal triad has no route to chaos. Its Lyapunov exponent is zero within numerical accuracy (λ1<3×10−3\lambda_1 < 3 \times10^{-3}) for 1.03≤S≤1.81.03 \le S \le1.8, and after a slowly modulated state close to threshold the motion is period-1 from S≈1.075S \approx1.075.

The non-zonal triad becomes chaotic through the breakdown of quasi-periodicity, as can be seen in Figure 4. The period-1 orbit undergoes a Neimark–Sacker bifurcation at S=1.1186S = 1.1186, which introduces a second, incommensurate frequency. In the period map the orbit becomes a small closed invariant curve around the fixed point. This is visible in the PSD, Figure 4(f): for S<1.1186S < 1.1186 the spectrum contains only the fundamental and its harmonics, and at the bifurcation a new family of lines appears at combinations of the two frequencies, growing in amplitude with SS. We identify the bifurcation by this appearance of a second frequency, together with the closed invariant curve in the period map, and locate it at the value of SS where the new lines first rise above the background. The invariant curve is destroyed at S∗=1.1197S^* = 1.1197. At this point the lines merge into a broadband background, λ1\lambda_1 leaves zero, and the maxima fill an interval, Figure 4(b). The transition is soft. Long laminar phases, during which the trajectory stays close to the former quasi-periodic orbit, are interrupted by chaotic bursts, with a mean laminar time ⟨τ⟩∝(S−S∗)−β\langle\tau\rangle\propto(S-S^*)^{-\beta} and β=1.04±0.05\beta= 1.04 \pm0.05. The exponent is close to the value 1 of type-II intermittency, which is associated with a Hopf-type instability[31].

Figure 4. Route to chaos with a fixed gradient at C=0.08C = 0.08, scanned in SS. (a), (b) Density of the maxima of ∣ϕ1∣\lvert\phi_1\rvert in the non-zonal triad, with each column normalized to unit weight and shown on a logarithmic grey scale. (c), (d) Largest Lyapunov exponent λ1\lambda_1 of the non-zonal (black) and zonal (dashed grey) triads. (e), (f) Power spectral density of ∣ϕ1∣(t)\lvert\phi_1\rvert(t), in log⁡10\log_{10} scale, as a function of frequency ff (Hann window, mean removed, maximum taken over frequency bins). Left: full range; right: enlargement around the quasi-periodic regime. The dotted lines in the enlargements mark the Neimark–Sacker bifurcation (NS) and the breakdown of quasi-periodicity (S∗S^*).

We define the laminar time as follows. For six values of SS above S∗S^*, logarithmically spaced in ε=S−S∗\varepsilon= S-S^* from 8×10−58 \times10^{-5} to 3×10−33 \times10^{-3}, we integrate the triad for 10710^7 time units after discarding a transient of 2×1042 \times10^4, and record every local maximum znz_n of ∣ϕ1∣|\phi_1| together with the time at which it occurs. During a laminar phase the trajectory follows the former quasi-periodic orbit, which is almost periodic, so successive maxima are nearly equal. We therefore call a maximum laminar if it differs from the next one by less than 0.02 and lies inside the band 0.28≤zn≤10.28 \le z_n \le1, which excludes the large excursions of the bursts. A laminar phase is a maximal run of consecutive laminar maxima, ended by the first maximum that fails this criterion, and only runs with at least 20 maxima are kept. Its duration τ\tau is the time between the first and the last maximum of the run, and phases cut by the beginning or the end of the record are discarded. The mean laminar time ⟨τ⟩\langle\tau\rangle at each SS is the average of τ\tau over all phases found, and only values of SS with at least 20 phases are used. The exponent β\beta is obtained from a straight-line least-squares fit of log⁡⟨τ⟩\log\langle\tau\rangle against log⁡ε\log\varepsilon, with S∗=1.11972S^* = 1.11972 kept fixed.

Chaos then persists up to S=1.6014S = 1.6014, with λ1\lambda_1 increasing from 0.03 to 0.13 and interrupted only by narrow periodic windows, which appear in the density of maxima as gaps and in the PSD as a return to sharp lines. The widest one, of period 3, opens abruptly at S=1.4690S = 1.4690 and coexists with the chaotic attractor at its lower edge. At Scr=1.6014S_{\mathrm{cr}} = 1.6014 the attractor collapses in an interior crisis into a narrow chaotic band (λ1≈0.08\lambda_1 \approx0.08). Just below ScrS_{\mathrm{cr}} the trajectory lingers in the band and escapes in bursts, as in crisis-induced intermittency[32, 33, 34]. The band is removed by a period-doubling cascade, which with decreasing SS reads 1→21 \to2 at S=1.7587S = 1.7587, 2→42 \to4 at 1.686, 4→84 \to8 at 1.6545 and 8→168 \to16 at 1.6495, and the subharmonic lines at f1/2nf_1/2^n are seen in the PSD. For S>1.7587S > 1.7587 the motion is period-1.

Self-consistent gradient

At the same CC the zonal triad is again regular. Over 0.45≤P≤30.45 \le P \le3 we find λ1<5×10−4\lambda_1 < 5 \times10^{-4} and a periodic oscillation of SS, with no bifurcation to chaos.

In the non-zonal triad the route has four stages, Figure 5. Apart from the first bifurcation to a two-frequency state, they do not resemble the fixed-gradient route.

Figure 5. Route to chaos with a self-consistent gradient at C=0.08C = 0.08, scanned in PP. (a), (b) Density of the maxima of SS in the non-zonal triad; the red density is the scan with PP decreasing and is shown only where it differs from the scan with PP increasing, that is, for P<P∗P < P^*. (c), (d) Largest Lyapunov exponent λ1\lambda_1, with PP increasing (black) and decreasing (red) for the non-zonal triad, and the zonal triad in dashed grey. (e), (f) Power spectral density of S(t)S(t), as in Figure 4. Left: full range; right: enlargement around the transition, with a step of 5×10−55 \times10^{-5} in PP. The dotted lines mark the Neimark–Sacker bifurcation (NS) and the breakdown of the locked quasi-periodic regime (P∗P^*).

The primary Hopf bifurcation occurs at Pc=0.515P_c = 0.515, and the amplitudes start to oscillate at P≈0.557P \approx0.557. From there up to P≈1.2P \approx1.2 the motion is period-1 and SS stays between about 1.07 and 1.19. The maxima of SS form a single thin branch, and the PSD contains only the fundamental and its harmonics. The long-wavelength mode is damped at all times and the triad behaves as the zonal one. The period-1 orbit loses stability at PNS≈1.234P_{\mathrm{NS}} \approx1.234, where the period-1 branch ends for increasing PP. As for the fixed gradient, we identify this as a Neimark–Sacker bifurcation from the appearance of a second frequency in the PSD and of a closed invariant curve in the period map, and we locate it at the end of the period-1 branch in the scan with step 10−410^{-4} in PP. The rotation number lies close to 2/52/5, and the motion locks into a period-5 orbit that exists for 1.249<P<1.3121.249 < P < 1.312, interrupted by narrow quasi-periodic intervals. In the PSD, Figure 5(f), this stage is a comb of sharp lines at multiples of f1/5f_1/5, with no continuous background. The period-5 orbit persists with decreasing PP down to P=1.21P = 1.21.

At P∗=1.3175P^* = 1.3175 the locked regime is replaced by a large chaotic attractor. Within a small increase in PP, λ1\lambda_1 jumps from below 10−310^{-3} to 0.050.05 and the maxima of SS spread in the interval. The PSD changes from a line spectrum to a broadband one at the same value of PP, with no intermediate stage in which the lines broaden or the background grows. Unlike the fixed gradient, the transition is hard. We measured the laminar time above P∗P^* as for the fixed gradient, since the regular orbit is now locked in period 5, a maximum is laminar if it differs from the one five maxima later by less than 0.020.02 and lies in the band 1.05≤zn≤1.261.05 \le z_n \le1.26, which contains the maxima of SS on the locked orbit, and only runs of at least 20 maxima are kept. Laminar phases are rare, about ten per record, and none lasts more than about 276 time units. Their mean duration, ⟨τ⟩≃2.2×102\langle\tau\rangle\simeq2.2 \times10^2, does not depend on P−P∗P - P^* (β=0.00±0.01\beta= 0.00 \pm0.01), whereas for the fixed gradient ⟨τ⟩\langle\tau\rangle grows by a factor of about 50 over a comparable range of ϵ\epsilon. The laminar time therefore does not diverge at P∗P^*. The transition is also hysteretic. When PP is decreased the chaotic state persists down to P=1.270P = 1.270, as shown by the red branch in Figure 5(b) and (d), where it coexists with the locked period-5 orbit reached with increasing PP, and from the map initial condition it is still found at P=1.2P = 1.2, where it is a long chaotic transient. The lower edge of the hysteresis is therefore not sharp, and depends on the step, because the chaotic set is metastable. The chaotic set becomes permanent only when the regular orbits disappear at P∗P^*. This accounts for the aperiodic region that the period map of Figure 3 shows from P≈1.22P \approx1.22 at this CC, which therefore contains a quasi-periodic orbit and a metastable chaotic set before the chaotic attractor proper. For 1.3175<P≤31.3175 < P \le3 the motion remains chaotic, with λ1\lambda_1 increasing slowly from 0.05 to 0.085 and only narrow periodic windows, for instance of period 6 near P=1.795P = 1.795.

In both models the non-zonal triad reaches chaos through a Neimark–Sacker bifurcation of the period-1 orbit and the zonal triad does not. The three diagnostics agree on where each stage occurs, but they differ in what they reveal. With a fixed gradient the two-frequency regime is a visible window, 1.1186<S<1.11971.1186 < S < 1.1197, in which new spectral lines grow before the lines merge into a continuum at S∗S^*, and the breakdown is soft, through intermittency. Chaos is then a range of SS that closes through a crisis and period doubling. With a self-consistent gradient the two-frequency regime is largely locked in a period-5 orbit, and its breakdown at P∗P^* is abrupt and hysteretic, with no spectral precursor and no divergence of the laminar time. Chaos then has no exit within the range of drive considered.

Conclusions

We have studied a resonant triad of Hasegawa–Wakatani modes in which the background density gradient, usually imposed from outside, evolves self-consistently, fed by a particle source and eroded by the turbulent flux, so that the predator–prey loop between drive and fluctuations is closed inside the model itself. To isolate the role of the zonal flow, we compared a triad whose long-wavelength mode is zonal with one that differs only by a slight tilt of that mode, which turns it into a weakly driven drift wave, each with a fixed and with a self-consistent gradient. The linear analysis yields a closed expression for the threshold of each mode and shows that the flux of an unstable wave always lowers the gradient that feeds it. The primary instability is therefore a supercritical Hopf bifurcation, in which the gradient is clamped at the marginal stability of the first unstable mode, and the strong-adiabaticity limit recovers the triad with growth and damping rates previously introduced by hand.

The central result is the contrast between the two geometries. A zonal mode cannot extract energy from the gradient and carries no flux, so it acts as a passive reservoir, and the triad remains essentially regular for both treatments of the gradient. When the same mode is tilted off the zonal axis it competes with the drift waves for the free energy of the gradient, the orderly exchange is lost, and chaos fills a sizeable part of the parameter space. The self-consistent gradient does not simply enhance or suppress chaos. It bounds the amplitudes that diverge with a fixed gradient, moves chaos from a narrow strip near threshold to stronger drive and a wider range of adiabaticity, and greatly enlarges the fraction of aperiodic states in the non-zonal triad, so the feedback amplifies the difference between the geometries. In both cases the trigger is the same, namely the long-wavelength mode being linearly driven, which with a self-consistent gradient happens intermittently and produces bursts of flux that return the gradient to a stiff profile.

The routes to chaos of the non-zonal triad also depend on the treatment of the gradient, and none of them exists when the long-wavelength mode is zonal. With a fixed gradient, a Neimark–Sacker bifurcation is followed by a soft breakdown of the torus through intermittency, and chaos is left through an interior crisis and a period-doubling cascade. With a self-consistent gradient the torus becomes phase-locked and breaks down abruptly, with hysteresis, long metastable transients and no exit from chaos within the range of drive explored. This matters for transport, since a soft transition gives an early warning whereas an abrupt and hysteretic one does not, and it is the feedback of the flux on the gradient that makes the transition abrupt.

Taken together, these results show that a zonal mode can reduce chaos structurally, not only by shearing eddies, because it removes the competition between active modes for the free energy of the gradient, and that this protection survives when the gradient responds to the turbulence. Since the analysis is restricted to a single triad and to specific geometries, the conclusions identify a mechanism rather than a quantitative prediction for a turbulent plasma. The natural continuation is a network of coupled triads, to test whether the reduction of chaos by zonal flows and its amplification by the self-consistent gradient persist when a broad spectrum of modes interacts.

References

  1. [1]¹W. Horton, “Drift waves and transport,” Reviews of Modern Physics 71, 735–778 (1999).DOI
  2. [2]²R. Balescu, Aspects of Anomalous Transport in Plasmas (CRC Press, Boca Raton, 2005).DOI
  3. [3]³W. Horton, Turbulent Transport in Magnetized Plasmas, 2nd ed. (WORLD SCIENTIFIC, 2017).
  4. [4]⁴P. Terry and W. Horton, “Stochasticity and the random phase approximation for three electron drift waves,” The Physics of Fluids 25, 491–501 (1982).DOI
  5. [5]⁵A. M. Batista, I. L. Caldas, S. R. Lopes, R. L. Viana, W. Horton, and P. J. Morrison, “Nonlinear three-mode interaction and drift-wave turbulence in a tokamak edge plasma,” Physics of Plasmas 13, 042510 (2006).DOI
  6. [6]J. S. Kim, R. J. Fonck, R. D. Durst, E. Fernandez, P. W. Terry, S. F. Paul, and M. C. Zarnstorff, “Measurements of Nonlinear Energy Transfer in Turbulence in the Tokamak Fusion Test Reactor,” Physical Review Letters 79, 841–844 (1997).DOI
  7. [7]W. Horton, “Nonlinear drift waves and transport in magnetized plasma,” Physics Reports 192, 1–177 (1990).DOI
  8. [8]P. H. Diamond, S.-I. Itoh, K. Itoh, and T. S. Hahm, “Zonal flows in plasma—a review,” Plasma Physics and Controlled Fusion 47, R35 (2005).DOI
  9. [9]A. Hasegawa and K. Mima, “Pseudo-three-dimensional turbulence in magnetized nonuniform plasma,” The Physics of Fluids 21, 87–92 (1978).DOI
  10. [10]A. Hasegawa and K. Mima, “Stationary Spectrum of Strong Turbulence in Magnetized Nonuniform Plasma,” Physical Review Letters 39, 205–208 (1977).DOI
  11. [11]C. Chandre, P. J. Morrison, and E. Tassi, “Hamiltonian formulation of the modified Hasegawa–Mima equation,” Physics Letters A 378, 956–959 (2014).DOI
  12. [12]N. Sato and M. Yamada, “A generalized Hasegawa–Mima equation in curved magnetic fields,” Journal of Plasma Physics 88, 905880319 (2022).DOI
  13. [13]N. Sato and M. Yamada, “Hamiltonian structure and nonlinear stability of steady solutions of the generalized Hasegawa–Mima equation for drift wave turbulence in curved magnetic fields,” Physica D: Nonlinear Phenomena 459, 134031 (2024).DOI
  14. [14]W. Horton and A. Hasegawa, “Quasi-two-dimensional dynamics of plasmas and fluids,” Chaos: An Interdisciplinary Journal of Nonlinear Science 4, 227–251 (1994).pubmed.ncbi.nlm.nih.gov/12780102
  15. [15]A. Hasegawa and M. Wakatani, “Plasma Edge Turbulence,” Physical Review Letters 50, 682–686 (1983).DOI
  16. [16]M. Wakatani and A. Hasegawa, “A collisional drift wave description of plasma edge turbulence,” The Physics of Fluids 27, 611–618 (1984).DOI
  17. [17]E.-j. Kim and P. H. Diamond, “Zonal Flows and Transient Dynamics of the $L\mathop{\mathrm{H}}$ Transition,” Physical Review Letters 90, 185006 (2003).pubmed.ncbi.nlm.nih.gov/12786016
  18. [18]B. D. Scott, “Energetics of the interaction between electromagnetic ExB turbulence and zonal flows,” New Journal of Physics 7, 92 (2005).DOI
  19. [19]L. Schmitz, L. Zeng, T. L. Rhodes, J. C. Hillesheim, E. J. Doyle, R. J. Groebner, W. A. Peebles, K. H. Burrell, and G. Wang, “Role of Zonal Flow Predator-Prey Oscillations in Triggering the Transition to H-Mode Confinement,” Physical Review Letters 108, 155002 (2012).pubmed.ncbi.nlm.nih.gov/22587261
  20. [20]G. Xu, V. Naulin, W. Fundamenski, C. Hidalgo, J. Alonso, C. Silva, B. Goncalves, A. Nielsen, J. Juul Rasmussen, S. Krasheninnikov, B. Wan, M. Stamp, and J. E. Contributors, “Blob/hole formation and zonal-flow generation in the edge plasma of the JET tokamak,” Nuclear Fusion 49, 092002 (2009).
  21. [21]Y. Zhang and S. I. Krasheninnikov, “Influence of zonal flow and density on resistive drift wave turbulent transport,” Physics of Plasmas 27, 122303 (2020).DOI
  22. [22]Ö. D. Gürcan, J. Anderson, S. Moradi, A. Biancalani, and P. Morel, “Phase and amplitude evolution in the network of triadic interactions of the Hasegawa–Wakatani system,” Physics of Plasmas 29, 052306 (2022).DOI
  23. [23]P. L. Guillon and Ö. D. Gürcan, “Phase transition from turbulence to zonal flows in the Hasegawa–Wakatani system,” Physics of Plasmas 32, 012306 (2025).DOI
  24. [24]R. Numata, R. Ball, and R. L. Dewar, “Bifurcation in electrostatic resistive drift wave turbulence,” Physics of Plasmas 14, 102312 (2007).DOI
  25. [25]A. E. Koniges, J. A. Crotinger, and P. H. Diamond, “Structure formation and transport in dissipative drift-wave turbulence,” Physics of Fluids B: Plasma Physics 4, 2785–2793 (1992).DOI
  26. [26]P. H. Diamond, Y.-M. Liang, B. A. Carreras, and P. W. Terry, “Self-Regulating Shear Flow Turbulence: A Paradigm for the L to H Transition,” Physical Review Letters 72, 2565–2568 (1994).pubmed.ncbi.nlm.nih.gov/10055916
  27. [27]D. Reiser, N. Oho, H. Tanaka, and L. Vela, “A plasma source driven predator-prey like mechanism as a potential cause of spiraling intermittencies in linear plasma devices,” Physics of Plasmas 21, 032302 (2014).DOI
  28. [28]A. Hasegawa and M. Wakatani, “Self-organization of electrostatic turbulence in a cylindrical plasma,” Physical Review Letters 59, 1581–1584 (1987).pubmed.ncbi.nlm.nih.gov/10035273
  29. [29]J. R. Dormand and P. J. Prince, “A family of embedded Runge-Kutta formulae,” Journal of Computational and Applied Mathematics 6, 19–26 (1980).DOI
  30. [30]M. A. Malkov and P. H. Diamond, “Weak hysteresis in a simplified model of the L-H transition,” Physics of Plasmas 16, 012504 (2009).DOI
  31. [31]Y. Pomeau and P. Manneville, “Intermittent transition to turbulence in dissipative dynamical systems,” Communications in Mathematical Physics 74, 189–197 (1980).
  32. [32]C. Grebogi, E. Ott, F. Romeiras, and J. A. Yorke, “Critical exponents for crisis-induced intermittency,” Physical Review A 36, 5365–5380 (1987).
  33. [33]C. Grebogi, E. Ott, and J. A. Yorke, “Chaotic Attractors in Crisis,” Physical Review Letters 48, 1507–1510 (1982).
  34. [34]C. Grebogi, E. Ott, and J. A. Yorke, “Crises, sudden changes in chaotic attractors, and transient chaos,” Physica D: Nonlinear Phenomena 7, 181–200 (1983).

Paper details

vlnc:W_4kb86q8k
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents