WorksPhysics and Astronomy
Recovering hysteresis in reduced models of plasma turbulence: a Mori–Zwanzig-inspired closure
Abstract
Hysteretic transitions between turbulent and strongly zonal-flow-dominated states provide an important test for reduced models of plasma turbulence and a minimal setting for studying dynamics relevant to L-H transitions. The 12-mode truncation of the modified Hasegawa-Wakatani system developed by Guillon and Gürcan retains this transition but underestimates the hysteresis observed in direct numerical simulations (DNS). Guided by the Mori-Zwanzig framework, we construct a phenomenological closure for the influence of discarded modes. The discarded spectrum is coarse-grained into a reservoir, whose feedback on the retained modes is modeled together with a separate drift-wave damping term. Without increasing the number of retained modes, the closure recovers nearly all of the hysteresis width missing from the unclosed truncation and places both transition thresholds close to their DNS values. To test the model beyond hysteresis, we also examine particle flux. The closed model improves upon the unclosed truncation, agreeing more closely with DNS at low adiabaticity but underestimating transport at high adiabaticity. These results show that a low-dimensional representation of discarded-mode feedback can improve predictions of hysteretic turbulence-flow transitions.
Introduction
Transitions between turbulent and strongly sheared-flow states are central to confinement regulation in magnetized plasmas. Drift-wave (DW) turbulence generates zonal flows (ZFs) through nonlinear energy transfer, while the resulting shear suppresses turbulence and regulates cross-field transport[1, 2, 3, 4, 5]. This turbulence–flow feedback can support distinct dynamical states and hysteretic transitions, contributing to the physics of the low- to high-confinement (L–H) transition[6, 7, 8, 9, 10, 11]. In particular, experiments have identified turbulence-driven ZFs and flow shear as important components of the transition to H-mode confinement[8].
The modified Hasegawa–Wakatani (MHW) system provides a minimal setting for studying this nonlinear interaction[12, 13, 14]. The MHW system exhibits a transition between turbulent and ZF-dominated states[14], with different transition thresholds as the adiabaticity parameter is increased and decreased[15]. Related hysteretic behavior has been observed in a warm-ion gyrofluid extension[16]. Although the MHW system does not describe the full L–H transition, it allows the associated turbulence–flow feedback and hysteresis to be studied in a simplified setting.
Reduced-order models (ROMs) seek to retain these transition dynamics with fewer degrees of freedom[17]. Spectral truncation may retain the modes needed for a transition while removing interactions needed to reproduce its hysteresis. For the MHW system, Guillon and Gürcan identified a 12-mode Fourier truncation that preserves the transition between turbulent and ZF-dominated states[15]. In our comparison with direct numerical simulations (DNS), however, this unclosed truncation recovers only about 16% of the hysteresis width. Richer spectral truncations provide one route to improved fidelity[18, 19], while closure and data-driven approaches represent the effects of discarded scales[20]. Here we ask whether a low-dimensional closure for the discarded-mode feedback can recover the missing hysteresis without increasing the number of retained Fourier modes.
The Mori–Zwanzig (MZ) formalism provides a framework for expressing the influence of eliminated degrees of freedom on the retained dynamics[21, 22, 23]. Eliminating these degrees of freedom yields exact equations containing memory and fluctuating contributions, providing a systematic starting point for reduced-order closure[24, 25, 26]. This framework has motivated closures for nonlinear and turbulent systems[27, 28, 29, 30], including plasma turbulence, where Maeyama and Watanabe showed that eliminated scales can damp, phase-shift, and excite retained fluctuations[31]. Low-dimensional phenomenological models have shown that essential turbulence–flow transition dynamics can survive substantial reduction. Malkov and Diamond used three coupled variables representing DW turbulence, zonal flow, and the driving pressure gradient to reproduce oscillatory intermediate states, bistability, and hysteresis, including an asymmetry between the forward and reverse transitions[11]. Their approach illustrates how effective feedback laws can capture transition dynamics without resolving the underlying turbulent spectrum.
Guided by the Mori–Zwanzig framework, we construct a closure for the feedback of the discarded spectrum while preserving the original 12-mode triadic structure. We first coarse-grain the discarded spectrum into a reservoir and determine its dynamics. We then model how this coarse-grained description couples back to the retained modes. We test whether this recovers the hysteresis lost under spectral truncation and whether the resulting reduced-order model generalizes to particle flux. These comparisons assess how much of the discarded-spectrum influence can be represented without adding retained Fourier modes.
The remainder of the paper is organized as follows. Section II introduces the MHW system and the 12-mode reduction. Section III derives the resolved–unresolved energy balance and develops the closure. Section IV compares the closed model with full-field DNS. § 1.2 discusses the physical interpretation and limitations of the model. Section VI summarizes the main conclusions.
Modified Hasegawa–Wakatani model and 12-mode reduction
The MHW system provides the DNS reference, while the canonical 12-mode truncation of Guillon and Gürcan[15] defines the resolved state. We introduce the spectral equations and transformed components needed to construct the closure, then define the transition diagnostic.
MHW dynamics in Fourier space
The MHW system describes resistive DW turbulence in a plane perpendicular to a uniform magnetic field [12, 13, 14]. The coordinates and denote the radial and poloidal directions. The dynamical fields are the normalized electrostatic potential and density perturbation about a prescribed background density . The vorticity is , and the background density gradient supplies the free energy for the DW instability.
We decompose each field as , where is the zonal average. The modified system restricts the adiabatic coupling to non-zonal fluctuations[14]. Here dissipation is also restricted to these fluctuations, giving
Here represents advection. The adiabaticity parameter controls the coupling between density and potential fluctuations, and the same dissipation coefficient is used in both equations.
For periodic fields, we use , where with and . All spectral sums exclude the spatially uniform mode and include conjugate partners satisfying . In Fourier space, the non-zonal projection takes the form , where removes the zonal modes and leaves the non-zonal modes unchanged. Writing , the spectral equations take the form
where contains the adiabatic coupling, dissipation and background-gradient drive,
with and .
The nonlinear term describes quadratic advection,
where is the geometric coupling factor. For zonal modes, and , giving . Under the chosen dissipation prescription, these modes therefore evolve solely through nonlinear coupling. Two zonal inputs have and do not interact directly. Every nonzero advective contribution thus requires at least one non-zonal input, referred to here as a DW mode.
Adiabatic and nonadiabatic components
Diagonalizing the coupling–dissipation part of (Eq. (4) with ) separates each non-zonal fluctuation into a nonadiabatic component relaxed by the adiabatic coupling and an adiabatic component on which that coupling vanishes. Its eigenvectors are and , with eigenvalues and . Writing , the transformed state has components
The amplitude is proportional to and therefore measures the departure from the adiabatic density response. The amplitude is proportional to , the generalized-vorticity. It is unaffected by the adiabatic coupling, damped only by . The same transformation is used for zonal modes, although their adiabatic coupling vanishes and neither component is linearly relaxed.
Restoring the background gives the full linear operator
The background-gradient drive mixes the two components for , allowing DW growth despite the non-negative coupling–dissipation rates . For zonal modes, , consistent with their evolution through nonlinear interactions[14, 32].
The transformed spectral system is
Equation (9) provides the starting point for the resolved–unresolved energy balance and closure in Section III.
Canonical 12-mode reduction
We adopt the canonical 12-mode Fourier truncation of Guillon and Gürcan[15]. The retained modes define the resolved state, and the remaining modes form the unresolved spectrum. The retained wavevectors are specified by
where denotes . Here is the reference poloidal wavenumber at , and . This lattice remains fixed throughout the sweep, whereas the injection wavenumber entering the DNS dissipation prescription in Eq. (40) varies with .
The set contains twelve independent wavevectors, comprising two zonal and ten DW modes. Each wavevector carries both vorticity and density amplitudes. Including conjugate partners gives
where is the complete nonzero Fourier lattice associated with the retained wavevector spacings. The reality condition determines the amplitudes at from those at .
The four outer sidebands and act as buffer modes. They remain resolved and participate in the retained nonlinear triads, but receive the additional damping specified in Section III E. Figure 1 shows the retained lattice.

Figure 1. Canonical 12-mode Fourier truncation, showing the twelve independent retained wavevectors. Conjugate partners are supplied by the reality condition. Zonal modes have , while drift-wave modes have . The four outer sidebands remain dynamically active but receive additional damping as buffer modes.
The full MHW system exhibits a transition between turbulent and ZF-dominated states[14]. The 12-mode truncation retains this first-order-like transition, characterized by and the zonal fraction as an effective order parameter[15]. We hold fixed and sweep upward and downward to determine the transition thresholds and the hysteresis width.
We diagnose the transition using the zonal kinetic-energy fraction
where
The sums extend over the full Fourier spectrum for DNS and over for the ROM. Thus measures the zonal share of kinetic energy within each model’s spectrum. Agreement in this diagnostic does not by itself imply agreement in the energy distribution among the retained modes. Let and denote the transition thresholds on the increasing and decreasing sweeps. The hysteresis width is
The thresholds are estimated from the first crossing of on each sweep, using the averaging and interpolation procedures in Section III E. This definition characterizes hysteresis under the prescribed finite-duration sweep. The unclosed truncation retains nonlinear interactions within but omits feedback involving . Section III develops a closure for this missing feedback, and Section IV evaluates its effect on the transition thresholds and hysteresis width.
Methodology
The closure represents the nonlinear feedback of discarded modes on the retained system. Inspired by the Mori–Zwanzig framework, we first characterize their coupling to the retained dynamics. We then use the resulting structure to construct a low-dimensional phenomenological closure.
Resolved–unresolved partition
We partition the spectral state as using the mode sets defined in Section II. The MZ formalism expresses the influence of the unresolved state on the resolved dynamics through projected dynamics, memory and orthogonal dynamics [21, 22, 23, 24, 26]. For , the Liouville operator generates the evolution of observables along the full dynamics. Let project onto observables of the retained state, and let denote the complementary projection. Thus is the mode observable evaluated after evolution for time . The exact resolved evolution is
The terms on the right-hand side describe projected dynamics, orthogonal dynamics, and memory, respectively. The projected dynamics need not coincide with the unclosed truncation because they can contain averaged unresolved contributions. We use this decomposition only to guide the closure.
The linear operator acts independently at each wavevector, so coupling between and occurs only through the advective triads . Because the drift velocity is divergence free, the nonlinear advection term redistributes energy across scales. It cannot change the total energy. Therefore,
For later use, we write this energy as
where the conventional factor of is omitted.
In the transformed coordinates,
The modal energy is therefore
We define the transformed inner product and norm by
where the sum is restricted to the indicated mode set.
The same inner product gives the energy rate,
Energy transfer between resolved and unresolved modes
The energy separates into resolved and unresolved contributions, , where
Here is the exact energy of the unresolved spectrum. The reservoir introduced below provides a reduced representation of its activity.
We define the nonlinear transfer rate by
where denotes transfer into the unresolved spectrum. Conservation does not constrain the sign of this transfer.
Including the linear contributions gives the exact balances
For either mode set,
This contribution includes both background-gradient drive and dissipation and need not be negative. These balances guide the reservoir and feedback models below, but do not by themselves determine their functional forms.
Closure construction
The truncation retains interactions among the resolved modes but omits their coupling to the unresolved spectrum. We coarse-grain the unresolved energy into a reservoir , then model its evolution and feedback on the retained modes.
Reservoir dynamics
Equation (18) separates linear evolution from nonlinear energy transfer between the resolved and unresolved modes. To identify the interactions responsible for this transfer, we write the MHW dynamics in the two mode sets:
The three letters identify the two input mode sets and the output mode set; mixed channels include both input orderings.
The resolved zonal modes evolve through nonlinear coupling, while the resolved DW modes also experience linear drive and dissipation. Separating and gives
Linear drive and dissipation vary across the unresolved modes, so their net contribution depends on how is distributed among them. Restricting the energy-rate identity in Eq. (16) to the unresolved modes and their linear dynamics, we express the net linear power as an instantaneous rate multiplying the unresolved energy:
Positive denotes net linear loss and negative net linear drive. Under the reservoir coarse-graining , we approximate this composition-dependent rate by a constant :
We now consider the nonlinear contribution to the reservoir dynamics. Interactions entirely within redistribute energy among unresolved modes but do not change the total unresolved energy. Exchange between and therefore enters through the RRU and RUU channels.
The RRU channel contains two resolved inputs and forces the unresolved modes. Under a short-memory, or Markovian, approximation, we model the unresolved response as local in time, so that its contribution to the reservoir energy is proportional to the squared instantaneous RRU forcing33,34.
Since ZF–ZF interactions vanish by Eq. (5), the RRU forcing contains DW–DW and mixed ZF–DW contributions,
As a minimal source model, we retain the DW–DW contribution. Because this forcing is quadratic in the resolved DW amplitudes,
Absorbing the response coefficient and reservoir normalization into gives
This source represents reservoir excitation by resolved DW–DW interactions; modulation by mixed ZF–DW interactions is not included.
The RUU channel contains one resolved and one unresolved input. In the unresolved-energy rate, is paired with the unresolved state, so the contribution contains two unresolved amplitudes and is therefore proportional to at fixed resolved state. We represent the remaining state dependence by a coarse-grained rate . Under ,
where may have either sign.
Combining this exchange with the unresolved linear contribution gives the net relaxation rate
The reservoir dynamics are therefore
Thus, resolved DW–DW interactions supply the reservoir, while the combined unresolved linear dynamics and RUU exchange determine the relaxation time .
The reservoir therefore provides a coarse-grained representation of the energy in the unresolved spectrum. Having modeled the evolution of the reservoir, we next determine the coupling between the reservoir and the resolved modes, beginning with the zonal flow.
Zonal-flow feedback
The resolved zonal modes couple to the unresolved spectrum through RUR and UUR interactions in Eq. (20), with
In the saturated regime, we assume that the unresolved energy dominates the resolved energy,
In this regime, interactions involving two unresolved modes are expected to dominate those involving only one. We therefore retain UUR as the leading unresolved contribution to the zonal dynamics. From Eq. (20) and Eq. (24), we have
The instantaneous UUR feedback depends on the detailed configuration of the unresolved modes discarded by the truncation. Following the statistical-projection interpretation of Mori–Zwanzig reduction[25], we characterize their collective effect by averaging over unresolved configurations at fixed resolved state and unresolved energy,
where the subscript denotes the quantities, resolved state and unresolved energy respectively, held fixed in the ensemble average. The zonal dynamics are then approximated by
The contribution of to the zonal-energy rate is . We isolate this energy-changing component by projecting onto the zonal state using the real part of the energy inner product:
where is the total retained zonal energy. The remaining component makes no instantaneous contribution to the zonal-energy change and is neglected in the scalar closure.
The projection does not constrain the sign of . Separating the unresolved-energy factor from the remaining dependence on the resolved state and unresolved statistics, we write
Under the coarse-graining , we neglect the remaining state dependence and represent by a constant positive coefficient, consistent with the use of effective relaxation rates in Markovian turbulence closures[33, 34]. Absorbing the reservoir normalization into then gives .
The zonal closure is therefore
The zonal closure therefore represents the coarse-grained unresolved feedback through a reservoir-dependent damping of the retained zonal modes. We next construct the corresponding feedback on the DW modes, focusing on the ZF-dominated plateau of the hysteresis loop, where DW perturbations are weak.
Drift-wave feedback
On this plateau, the zonal field remains strong while the DW amplitudes are small,
The zonal field can therefore couple a resolved DW perturbation into the unresolved spectrum, whose response subsequently couples back to the resolved DW modes. We construct this RRU–RUR return by retaining terms linear in the DW perturbation. Since ZF–ZF interactions vanish and DW–DW forcing is quadratic in the perturbation, the leading RRU forcing contains one zonal and one DW amplitude:
Thus, at fixed zonal field, , the leading RRU forcing is linear in the DW perturbation, . The resulting unresolved perturbations evolve under their linear dynamics and further interactions with the zonal field, before returning to the resolved DW modes through the RUR channel. We denote this feedback by .
For example,
For a first-harmonic zonal field of fixed structure and complex amplitude , the two interactions in this direct return contribute one factor of and one of . Under the short-memory approximation, both are evaluated at the current time, giving .
We define the first-harmonic zonal energy by
including both conjugate partners. This leading return motivates a feedback proportional to . As a modeling choice, we apply this dependence to all retained DW modes and represent the feedback by instantaneous scalar damping:
Together with the reservoir dynamics and zonal feedback above, this completes the three closure terms.
The closure
The resulting closure combines reservoir excitation and relaxation with feedback on the retained zonal and DW modes. The closure takes the form
The reservoir equation can be solved explicitly as
The first term carries the contribution of the initial reservoir state. The integral contains an exponentially weighted history of the resolved DW activity: activity at the earlier time is weighted by . Substituting this expression into the zonal-feedback term in Eq. (32) gives
where the modeled memory kernel is
Thus, past DW activity modifies the present zonal damping, with older contributions progressively forgotten on the characteristic timescale .
For calibration and integration, it is convenient to combine the reservoir normalization and zonal-feedback strength into parameters that enter the resolved dynamics directly. We absorb the zonal-feedback coefficient into the reservoir coordinate,
The closed reduced-order model is
The change of variable writes the reservoir contribution to the zonal equation as . In this form, the modeled kernel is
and its integrated strength is
Hence, sets the overall strength of the reservoir-mediated memory, while sets its decay time.
Closure parameters
We fix the reservoir relaxation rate at and calibrate to the DNS hysteresis-loop area. The closure parameters are
Appendix B gives their selection procedure. Here sets the DW damping per unit first-harmonic zonal energy.
Evaluation
The ROM lattice and buffer-damping prescription follow Guillon and Gürcan 15. The lattice remains fixed at its geometry, with , so varying changes the linear rates without changing the retained wavevectors. Ordinary viscosity is set to zero in the ROM, and the four outer sidebands receive the additional damping
where is the maximum growth rate of the inviscid homogeneous MHW system at and . The quantity is the real frequency of the corresponding mode on the positive- branch. DNS uses the viscosity prescription in Eq. (40).
The control parameter is swept from to and then down to in steps of , giving 32 plateaus with the turning point counted once. Each plateau lasts , and its final state initializes the next plateau. The ROM uses fixed-step fourth-order Runge–Kutta integration with and an initial warmup of . DNS uses adaptive integration and a warmup of . All ROM runs use , and the reservoir is initialized with before the warmup.
For the hysteresis comparison, the ROM uses second-half plateau averages of and DNS uses full-plateau averages. The particle-flux comparison uses the common averaging window 100 for DNS and both reduced models, where is the start of the plateau.
For each realization, and are determined from the first strict crossing of along the corresponding sweep, with linear interpolation between adjacent plateau averages. The width is calculated for each realization with both crossings before ensemble averaging. Transition thresholds are averaged over realizations with the corresponding crossing. Because the reduced models use matching initial conditions, comparisons between them are paired realization by realization. Appendix B gives the numerical details, the parameter selection and the construction of the DNS diagnostics.
Results
We compare the unclosed 12-mode truncation and the closed ROM with full-field DNS. The two reduced models share the same 32 initial conditions, allowing paired comparisons. The DNS reference uses an independent ensemble of 32 realizations. The evaluation is defined in Section III E.
Recovery of the hysteretic transition
The closure recovers most of the hysteresis lost under spectral truncation, as shown in Figure 2 and Table I. The closure increases the hysteresis width from in the unclosed model to , compared with in DNS, a gain of , or about 98% of the truncation–DNS difference. The corresponding threshold shifts are reported in Table I. Fourteen of the 32 unclosed realizations produce inverted loops, with . No inverted loops occur in the closed ROM or DNS ensembles.
Figure 2. (a) Ensemble-averaged zonal kinetic-energy fraction on the increasing- and decreasing- sweeps for DNS, the 12-mode truncation, and the closed ROM. Each ensemble contains 32 realizations; the two reduced models use the same initial conditions. Shading is defined in Appendix B. (b) Transition thresholds obtained from the crossings. Error bars denote one standard error. The hysteresis widths are 0.0433 0.0016 for DNS, 0.0069 0.0036 for the truncation, and 0.0427 0.0024 for the closed ROM. The truncation width is based on the 31 realizations with both crossings.
| Model | Inverted loops | |||
| DNS (512 512) | 0.1210 0.0007 | 0.0777 0.0018 | 0.0433 0.0016 | 0 |
| 12-mode truncation | 0.1252 0.0008 | 0.1184 0.0038 | 0.0069 0.0036 | 14 |
| Closed ROM | 0.1227 0.0011 | 0.0800 0.0022 | 0.0427 0.0024 | 0 |
Table I. Mean transition thresholds and hysteresis widths obtained from the first crossing on each sweep. Widths are calculated for each realization before ensemble averaging. Uncertainties denote one standard error. Each ensemble contains 32 realizations. Threshold means use the realizations with the corresponding crossing: 32 in all cases except the truncation and , which use 31. The reduced models use the same 32 initial conditions, while DNS is independent. Inverted loops satisfy .
The closed-ROM thresholds lie slightly above DNS, by 0.0017 for and 0.0023 for . These offsets are of the same order as the reported standard errors. The hysteresis width differs from DNS by only 0.0006.
Particle flux
Particle flux was not used to set the closure parameters and therefore tests the model beyond hysteresis. We evaluate
where the brackets denote a spatial average of the product. Since throughout, equals the radial particle flux . Figure 3 compares the ensemble-mean flux on both sweep branches using the common averaging window of Section III E. The reduced models use the same 32 initial conditions, while DNS uses 32 independent realizations. To quantify the change in agreement with DNS, we define
Figure 3. Ensemble-averaged radial particle flux on the increasing- and decreasing- sweeps for DNS, the 12-mode truncation, and the closed ROM, together with the error difference . Across the 32 branch– conditions, the closed ROM is closer to DNS at 19, the truncation at 7, and the difference is unresolved at 6. Shading in the flux panels and error bars in the panels show 95% confidence intervals across realizations.
where the overbar denotes the ensemble mean of the plateau-averaged flux. Negative indicates a smaller discrepancy for the closure.
Across the 32 branch– conditions, the closed ROM is closer to DNS at 19, the truncation at 7, and the difference is statistically unresolved at 6. The closed ROM agrees more closely with DNS at lower . At larger , however, its particle flux falls toward zero while DNS retains finite transport.
Discussion
The results show that feedback from the discarded spectrum can recover much of the hysteresis lost under spectral truncation without increasing the number of retained modes. Particle flux, which was not used to set the parameters, shows that the improvement is not uniform across the sweep. We first interpret the two sweep directions, then assess the two working assumptions underlying the closure against DNS, and finally discuss the remaining limitations and their implications for reduced modeling.
Physical interpretation of the two sweep directions
DWs generate ZFs through nonlinear energy transfer, while ZFs shear the fluctuations and reorganize their spectrum[3, 35, 36]. The two sweep directions probe different stages of the same turbulence–flow interaction. On the increasing- branch, a zonal state forms from a DW-dominated state; on the decreasing- branch, DWs regrow about an established zonal state[14, 15, 16].
As increases from the turbulent state, strong DW activity supplies the reservoir and sustains the reservoir-dependent ZF damping, . This feedback opposes ZF growth and therefore acts directly on the formation of the zonal state.
On the decreasing sweep, the system starts in a strongly zonal state with weak DW activity. The reservoir is then weakly driven and relaxes, reducing the zonal damping. The zonal field remains strong, however, so the DW damping, which increases with first-harmonic zonal energy, continues to suppress perturbations and delays their regrowth.
DNS assessment of the closure assumptions
Figure 4 assesses the amplitude ordering and short-memory approximation used in the closure. Under the nearest-grid resolved–unresolved partition, the amplitude ratio has a median of 12.41 and a minimum of 4.78 across the sampled late-time states. These values support the assumed ordering for this partition. The energy ratio alone, however, does not establish the relative magnitudes of the feedback terms, which also depend on the interacting wavevectors and phase correlations.
Figure 4. (a) Unresolved-to-resolved amplitude ratio along the increasing- and decreasing- sweeps. Curves denote the median across realizations and shading the interquartile range. (b) Centered temporal correlations of the RRU forcing and unresolved energy at on the increasing sweep and on the decreasing sweep.
The nearest-grid mapping does not preserve all ROM triads, so the RRU correlation in Figure 4(b) is an approximate diagnostic of the ROM partition. The temporal correlations are consistent with a faster variation of the RRU forcing than of the unresolved energy. At on the increasing branch and on the decreasing branch, the forcing correlation is already close to zero at the first nonzero lag, . The corresponding unresolved-energy correlations remain 0.70 and 0.85. Because the forcing decay at shorter lags is unresolved, these measurements support a qualitative separation of timescales.
Limitations and outlook
The closure remains phenomenological. Its reservoir source retains only DW–DW excitation, the net relaxation is represented by a constant positive rate, and the resolved feedback is approximated by scalar damping. The exact resolved–unresolved energy balance guides the construction, while the source and damping terms are modeled separately. The reservoir is therefore an effective measure of unresolved activity.
Particle flux was not used to set the closure parameters and therefore provides a separate test beyond hysteresis. The closed ROM agrees more closely with DNS at lower , while at larger its particle flux falls toward zero and DNS retains finite transport. The DW damping is set by the instantaneous first-harmonic zonal energy , whereas discarded-mode fluctuations can also repopulate the DW sector. The remaining discrepancies therefore point to feedback missing from this scalar damping model.
More fundamentally, the closure assumes that the discarded-spectrum feedback can be represented in terms of the retained modes and the single reservoir variable . The remaining discrepancies leave open whether this reduced state contains enough information to distinguish the turbulent and zonal-flow-dominated dynamics. Motivated by the history-dependent zonal states and hysteretic dynamics reported in gyrofluid MHW models [16, 37], a natural next step is to reconstruct the corresponding turbulent and zonal-flow-dominated attractors and determine what additional coordinates, if any, are required to describe their dynamics. This tests the state representation itself, rather than introducing additional feedback laws within the existing state.
The MHW system does not contain the equilibrium, geometric and transport physics required for a complete description of the L–H transition. Its role here is instead to provide a minimal setting for testing reduced models of hysteretic turbulence–flow transitions. The results show that a spectral truncation can retain the modes required for the transition while losing feedback from the discarded spectrum that is important for its hysteresis. Representing that feedback through one additional coarse-grained variable recovers much of the lost transition dynamics without enlarging the retained spectral system. Extending this approach to more complete plasma models will require identifying both the reduced state and the discarded-scale feedbacks needed to reproduce transition thresholds and transport.
Conclusion
We developed a closure for the 12-mode MHW system to recover the hysteresis lost under spectral truncation without adding retained Fourier modes. Guided by the Mori–Zwanzig framework, the closure adds a reservoir that damps the retained zonal modes and a separate DW damping that increases with the first-harmonic zonal energy. The original 12-mode triads remain unchanged.
For the sweeps considered here, the closed ROM reproduces the DNS hysteresis width, and both transition thresholds lie close to the DNS values. Particle flux was not used to set the model parameters and therefore provides a separate test beyond hysteresis. The closed ROM recovers the hysteresis and improves particle-flux agreement over much of the sweep; at large , the finite transport retained by DNS points to additional physics in strongly zonal states.
The remaining discrepancies motivate direct tests of the reservoir source and feedback laws against DNS, together with an assessment of whether additional state variables are needed. The MHW system provides a minimal setting for these tests and for developing reduced descriptions of turbulence–flow feedback relevant to L–H transition physics. Taken together, these results show that a compact closure can recover the hysteretic transition lost under spectral truncation while preserving the original 12-mode triadic structure.
Acknowledgments
This work was funded by National Research Foundation Singapore (NRF) core funding “Fusion Science for Clean Energy”, under the framework of Singapore Alliance with France for Fusion Energy (SAFE).
Author declarations
Conflict of Interest
The authors declare no conflicts of interest.
Author Contributions
Ameir Shaa: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Kyungtak Lim: Conceptualization, Methodology, Resources, Supervision, Writing – review & editing. Claude Guet: Conceptualization, Project administration, Resources, Funding acquisition, Supervision, Writing – review & editing. Xavier Garbet: Conceptualization, Project administration, Resources, Funding acquisition, Supervision, Writing – review & editing.
Data availability
The data supporting the findings of this study are available from the corresponding author upon reasonable request.
Supporting derivations
Channel notation
With the indicator of mode set , the ordered coupling, expressed in the transformed coordinates, of inputs from and into the output set is
A channel label lists the two input sets followed by the output set, and a mixed channel contains both input orderings: and .
Reservoir-source restriction
Let denote the unresolved response generated from zero by the RRU forcing,
The unresolved state can then be decomposed exactly as
where contains the pre-existing unresolved state and all contributions not generated by the RRU forcing. The RRU contribution to the unresolved-energy rate, , therefore splits as
where is the work of the RRU forcing on its own accumulated response. The reservoir source is constructed from ; the adopted law retains its DW–DW contribution.
At fixed zonal amplitude, scaling the DW amplitude by gives , where and are the mixed ZF–DW and DW–DW forcings at unit DW amplitude.
Substituting into this response work orders it as
where is the work of forcing on the propagated response to , and propagates the unresolved linear response from to . The source of Eq. (22) represents , the DW–DW contribution to the exact input of Eq. (39).
Numerical reproducibility
Numerical details
The DNS viscosity in Eq. (1) is
following Guillon and Gürcan [15]. DNS uses a grid on a square periodic domain of side with dealiasing and , integrated by DOP853 with relative and absolute tolerances and and maximum step 0.1. Section III E specifies the sweep, plateau duration, warmup, continuation, ROM integration and crossing rule.
Each ROM realization starts with amplitude in every retained mode except , which starts with amplitude ; and receive independent phases drawn uniformly from . Each DNS realization starts from with , one uniformly distributed phase shared by both fields, and the spatially uniform mode removed.
DNS stores 201 samples per plateau and the ROM stores 101. In Figure 2(a), the DNS band is , where is the root-mean-square deviation of all full-plateau samples from their pooled mean; the reduced-model bands are standard errors of the realization means. Negative individual widths are retained in the width averages. The particle-flux averages use the 49 samples per plateau at . Pointwise 95% intervals for the flux and for are percentiles of 5000 whole-realization bootstrap replicates, with DNS resampled independently of the paired reduced-model realizations.
The reported ROM results use 32 realizations that completed all 32 plateaus with finite , selected by ranking the mean absolute deviation of their closed-ROM plateau averages from the DNS mean hysteresis curve; the truncation uses the same initial conditions. DNS uses an independent ensemble of 32 realizations.
Parameter selection
We fix and determine the reservoir coupling from the DNS hysteresis-loop area, with . We define
where denotes trapezoidal integration over the increasing- grid and interpolates the decreasing-branch mean curve onto the same grid.
The calibration used a grid in the reservoir-source and zonal-feedback coefficients. In the -form of the final ROM, these 12 settings correspond to nine distinct effective couplings,
The closest agreement with the DNS hysteresis-loop area occurs at
For the DW damping coefficient, refinement against the DNS transition thresholds selected and . We adopt their midpoint, . Particle flux was not used in parameter selection.
DNS diagnostics
For Figure 4, the resolved DNS partition is the set of nearest-grid images of all retained ROM modes, including the eight DW sidebands. Zonal modes map to with ; DW carriers to ; the sidebands to at those ; all remaining nonzero modes form .
Panel (a) uses 32 DNS realizations, 32 plateaus and 11 late-time frames per plateau, at stored frames 100, 110, . . . , 200, giving sampled states. For each realization and plateau, the plotted value is the median over the 11 sampled frames; the curves and shading show the median and interquartile range across realizations. The pooled median and minimum quoted in § 1.2 B refer to all 11264 sampled states.
Panel (b) uses the 101 consecutive stored frames 100–200 of DNS realization 0 at on the increasing sweep and realization 15 at on the decreasing sweep, with between adjacent frames. For a record , , with and whole-record mean , the lag- correlation is
For the RRU forcing, at frame with the energy inner product; for the unresolved energy, with the scalar product.
References
References
- [1]W. Horton, “Drift waves and transport,” Rev. Mod. Phys. 71, 735–778 (1999).DOI
- [2]H. Biglari, P. H. Diamond, and P. W. Terry, “Influence of sheared poloidal rotation on edge turbulence,” Phys. Fluids B 2, 1–4 (1990).DOI
- [3]P. H. Diamond, S.-I. Itoh, K. Itoh, and T. S. Hahm, “Zonal flows in plasma—a review,” Plasma Phys. Control. Fusion 47, R35 (2005).DOI
- [4]Z. Lin, T. S. Hahm, W. W. Lee, W. M. Tang, and R. B. White, “Turbulent transport reduction by zonal flows: Massively parallel simulations,” Science 281, 1835–1837 (1998).
- [5]A. Fujisawa, “A review of zonal flow experiments,” Nucl. Fusion 49, 013001 (2009).DOI
- [6]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,” Phys. Rev. Lett. 72, 2565–2568 (1994).pubmed.ncbi.nlm.nih.gov/10055916
- [7]E. Kim and P. H. Diamond, “Zonal flows and transient dynamics of the L-H transition,” Phys. Rev. Lett. 90, 185006 (2003).pubmed.ncbi.nlm.nih.gov/12786016
- [8]L. Schmitz et al., “Role of zonal flow predator-prey oscillations in triggering the transition to H-mode confinement,” Phys. Rev. Lett. 108, 185002 (2012).pubmed.ncbi.nlm.nih.gov/22587261
- [9]L. Schmitz, “The role of turbulence–flow interactions in L- to H-mode transition dynamics: recent progress,” Nucl. Fusion 57, 025003 (2017).DOI
- [10]K. Miki, P. H. Diamond, Ö. D. Gürcan, G. R. Tynan, T. Estrada, L. Schmitz, and G. S. Xu, “Spatio-temporal evolution of the L → I → H transition,” Phys. Plasmas 19, 092306 (2012).DOI
- [11]M. A. Malkov and P. H. Diamond, “Weak hysteresis in a simplified model of the L–H transition,” Phys. Plasmas 16, 012504 (2009).DOI
- [12]A. Hasegawa and M. Wakatani, “Plasma edge turbulence,” Phys. Rev. Lett. 50, 682 (1983).DOI
- [13]M. Wakatani and A. Hasegawa, “A collisional drift wave description of plasma edge turbulence,” Phys. Fluids 27, 611–618 (1984).DOI
- [14]R. Numata, R. Ball, and R. L. Dewar, “Bifurcation in electrostatic resistive drift wave turbulence,” Phys. Plasmas 14, 102312 (2007).DOI
- [15]P. L. Guillon and Ö. D. Gürcan, “Phase transition from turbulence to zonal flows in the Hasegawa–Wakatani system,” Phys. Plasmas 32, 012306 (2025).DOI
- [16]F. Grander, F. F. Locker, and A. Kendl, “Hysteresis in the gyrofluid resistive drift wave turbulence to zonal flow transition,” Phys. Plasmas 31, 052301 (2024).DOI
- [17]P. Benner, S. Gugercin, and K. Willcox, “A survey of projection-based model reduction methods for parametric dynamical systems,” SIAM Rev. 57, 483–531 (2015).DOI
- [18]P. L. Guillon, R. Angles, Y. Sarazin, and Ö. D. Gürcan, “Anisotropic truncation for turbulent transport and zonal flows in the Hasegawa–Wakatani system,” Plasma Phys. Control. Fusion 68, 085022 (2026).DOI
- [19]P. L. Guillon, Ö. D. Gürcan, G. Dif-Pradalier, Y. Sarazin, and N. Fedorczak, “Flux-driven turbulent transport using penalisation in the Hasegawa–Wakatani system,” J. Plasma Phys. 91, E145 (2025).DOI
- [20]K. Li, Y. Cho, X. Garbet, C. Wan, R. Varennes, K. Lim, V. Grandgirard, Z. S. Qu, and Y. S. Ong, “Accelerating edge turbulence simulations: A physics-informed data-driven closure model for the Hasegawa–Wakatani system,” Plasma Sci. Technol. (2026), 10.1088/2058-6272/ae936d, published online 31 July 2026.DOI
- [21]S. Nakajima, “On quantum theory of transport phenomena: steady diffusion,” Prog. Theor. Phys. 20, 948 (1958).DOI
- [22]R. Zwanzig, “Memory effects in irreversible thermodynamics,” Phys. Rev. 124, 983 (1961).
- [23]H. Mori, “Transport, collective motion and Brownian motion,” Prog. Theor. Phys. 33, 423 (1965).DOI
- [24]A. J. Chorin, O. H. Hald, and R. Kupferman, “Optimal prediction and the Mori–Zwanzig representation of irreversible processes,” Proc. Natl. Acad. Sci. U.S.A. 97, 2968–2973 (2000).
- [25]A. J. Chorin, O. H. Hald, and R. Kupferman, “Optimal prediction with memory,” Physica D 166, 239–257 (2002).DOI
- [26]D. Givon, R. Kupferman, and A. Stuart, “Extracting macroscopic dynamics: model problems and algorithms,” Nonlinearity 17, R55–R127 (2004).DOI
- [27]P. Stinis, “Renormalized Mori-Zwanzig-reduced models for systems without scale separation,” Proc. R. Soc. A 471, 20140446 (2015).pubmed.ncbi.nlm.nih.gov/27547070
- [28]E. J. Parish and K. Duraisamy, “Non-Markovian closure models for large eddy simulations using the Mori-Zwanzig formalism,” Phys. Rev. Fluids 2, 014604 (2017).DOI
- [29]E. J. Parish and K. Duraisamy, “A dynamic subgrid scale model for large eddy simulations based on the Mori-Zwanzig formalism,” J. Comput. Phys. 349, 154 (2017).DOI
- [30]K. K. Lin and F. Lu, “Data-driven model reduction, Wiener projections, and the Koopman-Mori-Zwanzig formalism,” J. Comput. Phys. 424, 109864 (2021).DOI
- [31]S. Maeyama and T.-H. Watanabe, “Extracting and Modeling the Effects of Small-Scale Fluctuations on Large-Scale Fluctuations by Mori-Zwanzig Projection Operator Method,” J. Phys. Soc. Jpn. 89, 024401 (2020).DOI
- [32]Ö. D. Gürcan and P. H. Diamond, “Zonal flows and pattern formation,” J. Phys. A: Math. Theor. 48, 293001 (2015).DOI
- [33]S. A. Orszag, “Analytical theories of turbulence,” J. Fluid Mech. 41, 363–386 (1970).DOI
- [34]J. C. Bowman, J. A. Krommes, and M. Ottaviani, “The realizable Markovian closure. I. General theory, with application to three-wave dynamics,” Phys. Fluids B 5, 3558–3589 (1993).DOI
- [35]R. L. Dewar and R. F. Abdullatif, “Zonal flow generation by modulational instability,” in Frontiers in Turbulence and Coherent Structures, World Scientific Lecture Notes in Complex Systems, Vol. 6, edited by J. Denier and J. S. Frederiksen (World Scientific, 2007) pp. 415–430.DOI
- [36]A. V. Pushkarev, W. J. T. Bos, and S. V. Nazarenko, “Zonal flow generation and its feedback on turbulence production in drift wave turbulence,” Phys. Plasmas 20, 042304 (2013).DOI
- [37]F. Grander, T. Gröfler, F. F. Locker, M. Rinner, and A. Kendl, “Merging of zonal flows in gyrofluid resistive drift-wave turbulence,” Physics of Plasmas 33, 052303 (2026).


