2.1. Linear gyrokinetic equation

We use the standard gyrokinetic system of equations (Brizard & Hahm Reference Brizard and Hahm2007) to describe electrostatic fluctuations destabilised along a thin flux tube tracing a magnetic field line. The ballooning transform (Connor, Hastie & Taylor Reference Connor, Hastie and Taylor1978; Dewar & Glasser Reference Dewar and Glasser1983) is used to separate out the fast perpendicular (to the magnetic field) scale from the slow parallel scale. We assume Boltzmann-distributed (adiabatic) electrons, thus solving for the perturbed ion distribution $g_{i}({v_{\parallel }},{v_{\perp }},\ell ,t)$ , defined to be the non-adiabatic part of $\delta f_{i}$ ( $\delta f_{i}=f_{i}-f_{i0})$ , with $f_{i}$ the ion distribution function and $f_{i0}$ a Maxwellian. The electrostatic potential is $\phi ({\ell })$ , and $v_{\parallel }$ and $v_{\perp }$ are the particle velocities parallel and perpendicular to the magnetic field, respectively. The gyrokinetic equation reads

(2.1) \begin{align} i{v_{\parallel }} \frac {\partial g}{\partial \ell } + (\omega - \widetilde {\omega }_d)g = \varphi J_0(\omega - \widetilde {\omega }_*^{T})f_0 ,\end{align}

where $\omega$ is the mode frequency, $\widetilde {\omega }_*^{T} = (Tk_{\alpha }/q)\mathrm{d}\ln T/\mathrm{d}\psi (v^2/v_{\mathrm{T}}^2 - 3/2)$ is the diamagnetic frequency and $J_{0} = J_{0}(k_{\perp }(\ell )v_{\perp }/\varOmega (\ell ))$ is the Bessel function of zeroth order. The thermal velocity is $v_{\mathrm{T}} = \sqrt {2T/m}$ , the thermal ion Larmor radius is $\rho = v_{\mathrm{T}}/(\varOmega \sqrt {2})$ , $n$ and $T$ are the background ion density and temperature, $q$ is the ion charge, $\varphi = q\phi /T$ is the normalised electrostatic potential and $\varOmega =q B/m$ is the cyclotron frequency, with $B=|\boldsymbol{B}|$ . The magnetic drift frequency is $\widetilde {\omega }_d = (1/\varOmega )\boldsymbol{k_{\perp }} \boldsymbol{\cdot }[ ( \boldsymbol{b} \times \boldsymbol{\kappa }){v_{\parallel }}^2 + (1/B)(\boldsymbol{b} \times \boldsymbol{\nabla } B) {v_{\perp }}^2/2 ]$ , with $\boldsymbol{\kappa } = \boldsymbol{b} \boldsymbol{\cdot }\boldsymbol{\nabla }\ \boldsymbol{b}$ and $\boldsymbol{b}=\boldsymbol{B}/B$ . For simplicity in this analysis, we set $k_{\psi }=0$ . We hereafter focus on the curvature drift, $K_{d}(\ell ) \equiv a^{2}{\boldsymbol{\nabla }}\alpha \boldsymbol{\cdot }\boldsymbol{b} \times \boldsymbol{\kappa }$ , referring to $K_{d}(\ell )$ as the ‘drift curvature’ and to individual regions of bad curvature along the field line (where $K_{d}\gt 0$ ) as drift wells. The curvature drift is used since, at finite plasma $\beta$ , it experiences a shift relative to the grad-B drift proportional to $\boldsymbol{\nabla } P/B^{2}$ , with $P$ the plasma pressure and the sign of this term can make bad curvature worse. We define a radial coordinate $r=a\sqrt {\psi /\psi _{edge}}$ , with $a$ the minor radius corresponding to the flux surface at the edge, and $\psi _{edge}$ the toroidal flux at that location. The temperature-gradient scale length is measured relative to the minor radius, $a/L_{T}=-(a/T)\mathrm{d}T/\mathrm{d}r$ .

Finally, the gyrokinetic system is completed by quasineutrality

(2.2) \begin{equation} \int \text{d}^3\boldsymbol{v} J_{0} g = n(1 + \tau ) \varphi , \end{equation}

with $\tau =|q_{e}|T/(qT_{e})$ , $T_{e}$ the electron temperature and $q_{e}$ the electron charge.

2.2. Previous CG model

To start the discussion of CG models we quote an earlier expression from (Roberg-Clark et al. Reference Roberg-Clark, Plunk, Xanthopoulos, Nührenberg, Henneberg and Smith2023) that incorporates $\boldsymbol{\nabla } \alpha$ as well as the parallel connection length $L_{\parallel }$ of the geometry. This expression was obtained by fitting coefficients to a numerical solution of the linear dispersion relation for localised toroidal modes and reads

(2.3) \begin{equation} \frac {a}{L_{T,crit}} = \left ( \frac {1+\tau }{2} \right ) \frac {a}{R_{\mathrm{eff}}} \begin{cases} 2.58 + 4.926 \left ( \pi a|\boldsymbol{\nabla } \alpha | \dfrac{R_{\mathrm{eff}}}{L_{\parallel }} \right )^{2}, & \left (\pi a|\boldsymbol{\nabla } \alpha | \dfrac{R_{\mathrm{eff}}}{L_{\parallel }} \right ) \lt 0.78 \\[15pt] 7.5 \: \pi a|\boldsymbol{\nabla } \alpha | \dfrac{R_{\mathrm{eff}}}{L_{\parallel }}, & \left (\pi a|\boldsymbol{\nabla } \alpha | \dfrac {R_{\mathrm{eff}}}{L_{\parallel }} \right ) \geqslant 0.78. \end{cases} \end{equation}

Here, $R_{\mathrm{eff}}$ in this case was determined by doing a local quadratic fit to the profile of $K_{d}$ and selecting the peak value of the local fit $(1/R_{\mathrm{eff}}=K_{d,\text{peak}})$ , $\boldsymbol{\nabla } \alpha$ was chosen at the location of the peak fitted $K_{d}$ and $L_{\parallel }$ , the parallel connection length, was the distance between two points where the sign of $K_{d}$ flipped, defining the bad curvature region for the fitting. The quadratic fit was used as a form of coarse graining – an averaging procedure meant to draw out the underlying behaviour of a smooth background in the presence of oscillations. We found that, generally, for stellarators with a large number of field periods and a vacuum magnetic well, the short connection length (or weak curvature) limit $\pi a|\boldsymbol{\nabla } \alpha | R_{\mathrm{eff}}/L_{\parallel } \geqslant 0.78$ tends to be the most relevant and we therefore focus on this limit in what follows.

A basic issue with the previous method is that drift curvature wells in general do not take the form of a low-order polynomial function. It is often the case that the drift curvature at the outboard midplane, where ITG turbulence often peaks in gyrokinetic simulations, has a peculiar form for QI configurations. Instead of a single drift well that can be approximated by a sine function or a quadratic curve, a typical QI configuration instead shows a ‘split’, symmetric about the point mentioned above, into two distinct drift wells with a null at the central location. The null point is a consequence of having straight magnetic field lines in real space at the maximum of the magnetic field strength $B$ . While the previous method can perhaps accurately model $g^{\alpha \alpha }$ with its value at the centre of the drift well for quasisymmetric configurations, $g^{\alpha \alpha }$ can in fact vary significantly. Landreman et al. (Reference Landreman, Choi, Alves, Balaprakash, Churchill, Conlin and Roberg-Clark2025) found that $g^{\psi \psi }$ , when averaged over the drift well, was better correlated with heat fluxes than the value chosen at a single point, in this case the outboard midplane location where turbulence often causes the most transport. These results lend credence to the idea that drift-well-averaged quantities are the most relevant to transport and should lead to improvements over the old method.

We note that the approximate geometric situation described above is not limited to QI stellarators, since tokamaks with reverse triangularity often exhibit similar double-well behaviour, see figures 15(b) and 15(c) of Balestri et al. (Reference Balestri, Ball, Coda, Cruz-Zabala, Garcia-Munoz and Viezzer2024). We show a toy model geometry in figure 1 which has qualitative elements that can arise in both geometries, in that the drift curvature contains a null (or nearly zero) region of finite width and $g^{\alpha \alpha } = |\boldsymbol{\nabla }\alpha |^2$ is amplified over the region of bad curvature as a result of magnetic shear. Negative triangularity tokamaks have apparent ITG stability benefits such as relatively high CGs (Balestri et al. Reference Balestri, Ball, Coda, Cruz-Zabala, Garcia-Munoz and Viezzer2024), an effect which stellarators may emulate, despite belonging to a different geometry class. A CG model capable of handling such cases would thus also allow modelling of more complex stellarator and tokamak cases while avoiding possible issues related to evaluating $g^{\alpha \alpha }$ at a single point. Furthermore, such a model could be used in optimisation to force the parallel connection length of localised toroidal modes to lie within the now-separated drift wells, effectively doubling the CG (by halving $L_{\parallel }$ in expression (2.3)). We refer to this as ‘splitting the mode’ (§ 2.4) and now describe how we refine the CG model to handle such cases.

Figure 1. Left: simplified toy model geometry for the drift curvature $K_{d}$ and squared gradient of the binormal coordinate ( $g^{\alpha \alpha } = |\boldsymbol{\nabla } \alpha |^{2}$ ) depicted along a magnetic field line ( $\ell$ coordinate), with split drift curvature wells and secularly increasing $g^{\alpha \alpha }$ . A QI stellarator, or reverse triangularity tokamak with large magnetic shear and flat surfaces near the outboard midplane, could qualitatively achieve such geometry profiles. Right: magnetic flux surfaces and outer boundary for a QI configuration at zero toroidal angle as defined in the VMEC code (Hirschman & Whitson Reference Hirshman and Whitson1983) with a ‘reverse-D’ shape at the maximum of $B$ . Note the relative compression of the surfaces on the outboard. Roughly constant magnetic field (as the poloidal angle is traversed) is achieved by offsetting the $1/r$ scaling of $B$ with surface expansion on the inboard and compression on the outboard.

2.3. Updated CG model

We now imagine coarse graining of the geometry as a kind of field line average that takes place over a single region of bad curvature in which the sign of the bad curvature does not reverse. We update the previous model (Roberg-Clark et al. Reference Roberg-Clark, Plunk, Xanthopoulos, Nührenberg, Henneberg and Smith2023) by defining the root-mean-squared average

(2.4) \begin{equation} \langle (\cdots) \rangle _{CG} (\ell ) = \left [ \frac { \ \int \varTheta (K_{d}(\ell ^{\prime })) (\cdots)^{2}(\ell ^{\prime }) \varTheta _{w}(\ell ,\ell ^{\prime }) {\rm d} \ell ^{\prime } }{ \int \varTheta (K_{d}(\ell ^{\prime })) \varTheta _{w}(\ell ,\ell ^{\prime }) {\rm d} \ell ^{\prime }} \right ]^{1/2}, \end{equation}

to calculate quantities like the average of $|\boldsymbol{\nabla } \alpha |$ instead of choosing the central peak fitted value of $|\boldsymbol{\nabla } \alpha |$ as before. The ‘observation point’ $\ell$ effectively sets the mode location, as enforced by the Heaviside function $\varTheta _{w}(\ell ,\ell ^{\prime })$ , which is $1$ if $K_{d}$ has no sign changes between $\ell$ and $\ell ^{\prime }$ and zero otherwise. This restricts the integration to single drift wells, where the other Heaviside function, $\varTheta (K_{d}(\ell ^{\prime }))$ , is $1$ where $K_{d}$ is negative (destabilising) and zero otherwise. The parallel connection length is calculated by finding the width of this region, namely

(2.5) \begin{equation} L_{\parallel }(\ell )_{CG} = \int \varTheta (K_{d}(\ell ^{\prime })) \varTheta _{w}(\ell ,\ell ^{\prime }) {\rm d} \ell ^{\prime }, \end{equation}

while the average curvature drive at marginal stability is defined as

(2.6) \begin{equation} \left ( \frac {a}{R_{\mathrm{eff}}} \right )_{CG}(\ell ) = \frac {1}{L_{\parallel CG}(\ell )} \int K_{d}(\ell ^{\prime }) \varTheta (K_{d}(\ell ^{\prime })) \varTheta _{w}(\ell ,\ell ^{\prime }) {\rm d} \ell ^{\prime }. \end{equation}

Note that, in contrast to the previous model, the integration is carried out such that the endpoints (regions where the sign of $K_{d}$ flips between two grid points) are included with linear interpolation in the integral. Armed with more general methods for calculating averages, we express the new CG along a single field line $\alpha$ as

(2.7) \begin{equation} \frac {a}{L_{T,crit}}(\alpha ) = \text{Min}_{\ell } \left [ \pi \left ( \frac {1+\tau }{2} \right ) \left (3.75 \: \langle a|\boldsymbol{\nabla } \alpha | \rangle _{CG}(\ell ) \frac {a}{L_{\parallel CG}(\ell )} \right ) \right ], \end{equation}

which ignores any explicit dependence on $R_{\mathrm{eff}}$ (consistent with the weak curvature limit) and uses a coefficient of half the size compared with the old model for the $L_{\parallel }$ term. Implemented as an optimisation target, the method scans over all observation points $\ell$ present in the given flux tube and finds the location corresponding to the minimum CG. The lack of explicit dependence on $R_{\mathrm{eff}}$ means that during optimisation the size of curvature will be determined by properties other than the CG, like the aspect ratio, number of field periods, imposition of a vacuum magnetic well and addition of plasma $\beta$ , which modifies the drift curvature profile. In this finite $\beta$ limit we use the curvature drift, as defined through $K_{d}$ in § 2.1. The target is then defined as

(2.8) \begin{align} f_{CG}(r)=\left (\text{Min}_{\alpha }\left [\frac {a}{L_{T,crit}}(\alpha )\right ]-2.0\right )^{2} ,\end{align}

which can be used to improve the CG during optimisation on a specific surface $r$ . For all of the cases in this paper, we optimise at $r/a=0.5$ .

2.4. Splitting the mode

Figure 2. Splitting ITG modes – idealised field line geometry metrics at the typically most unstable location on the outboard midplane of a QI magnetic field configuration. (a) No attempt is made to split the mode geometrically across the standard gap at $\ell =0$ . Thus the mode can average the drift curvature across the gap. We surmise that $(a/R_{\mathrm{eff}})_{CG}$ will be similar to a case with no gap. The mode freely ‘tunnels’ across the gap to encompass an $L_{\parallel }$ of the entire well (achieving a low CG) with similar curvature drive to modes that localise on each side of the gap. (b) A barrier is imposed by the geometry through $g^{\alpha \alpha } = |\boldsymbol{\nabla }\alpha |^2$ , reducing the growth rate of the mode that tunnels across the gap. (c) Separating the two wells reduces the average curvature drive significantly for the tunnelling mode, as similarly portrayed in figure 1. Additional shear amplification of $g^{\alpha \alpha }$ further stabilises the localised modes.

Having found a CG formula that is compatible with well splitting, we can discuss how this effect is realised in practice. In particular, we imagine two different ways to ensure that the parallel connection length of toroidal ITG modes is halved by splitting the mode. The goal here, via manipulating the field line geometry, is to decouple the behaviour of modes that straddle both wells and those that localise to individual wells. The localised modes should then be further stabilised by the geometry.

To illustrate the common (un-optimised) situation for QI stellarators, we first show a case where no attempt is made to split the mode, aside from having a zero point of curvature (figure 2 a). In this case the mode can simply ‘tunnel’ across the gap at little cost to the average value of the curvature (the gap is small compared with the overall width of the region). The end result is a mode with a parallel connection length containing both drift wells that is not significantly affected, in terms of curvature drive or parallel dynamics, by a null point in the curvature.

A way to more effectively split the mode was already shown in figure 1 – to make the gap between the two resulting wells large, such that the gap is comparable to or larger than the width of the two wells, as also shown in figure 2(c). The modes which are strongly driven by curvature will be forced to localise in each separate well, and their stability will then be set by the local drift well behaviour, while the mode that spreads across the barrier will experience a reduced curvature drive $(a/R_{\mathrm{eff}})_{CG}$ as long as the curvature does not increase in proportion to the widening of the gap. Note this means that residual turbulence may grow at a CG below that of the more localised modes, since they have a larger $L_{\parallel }$ , which we believe will produce a ‘foot’ in plots of nonlinear heat flux versus temperature gradient (Zocco et al. Reference Zocco, Xanthopoulos, Doerk, Connor and Helander2018, Reference Zocco, Podavini, Garcìa-Regaña, Barnes, Parra and Mishchenko2022). A second way to split the well is to insert a barrier through which ITG modes with strong curvature drive cannot tunnel, in this case by increasing $\boldsymbol{\nabla } \alpha$ at the central location $\ell =0$ on the flux tube (Roberg-Clark, Plunk & Xanthopoulos Reference Roberg-Clark, Plunk and Xanthopoulos2021) (figure 2 b). The mode that tunnels suffers reduced growth rates as a result. We show stellarator configurations which utilise the first method later in the paper.

As already mentioned, the mode splitting (starting with a null point in the drift curvature), if it is present or nearly occurring in a vacuum magnetic equilibrium, can be altered by the shift in the curvature drift in the gyrokinetic equation brought about by a finite $\beta$ term proportional to $\boldsymbol{\nabla } p / B^{2}$ . This term can make bad curvature worse since on the outboard the temperature gradient and normal curvature vectors tend to be parallel. As such, the curvature drift may capture a lower CG from an increase in $L_{\parallel CG}$ in finite $\beta$ systems.

2.5. Ion-temperature-gradient instability above the threshold

Once the ITG modes are localised to individual drift wells, it remains to be seen how stable they will be. Growth rates for these modes depend on a variety of geometric factors, such as the magnitudes of $K_{d}$ , $|\boldsymbol{\nabla } \alpha |$ and $|\boldsymbol{\nabla } \psi |$ . Since, above the threshold, modes continue to localise in regions of bad curvature, we slightly modify the average in (2.4) by weighting it with the magnitude of curvature

(2.9) \begin{equation} \langle (\boldsymbol{\cdot }\boldsymbol{\cdot }\boldsymbol{\cdot }) \rangle _d (\ell ) = \left [ \frac { \int \: K_{d}(\ell ^{\prime }) \varTheta (K_{d}(\ell ^{\prime })) (\boldsymbol{\cdot }\boldsymbol{\cdot }\boldsymbol{\cdot })^{2}(\ell ^{\prime }) \varTheta _{w}(\ell ,\ell ^{\prime }){\rm d} \ell ^{\prime }}{\int K_{d}(\ell ^{\prime }) \varTheta (K_{d}(\ell ^{\prime }))\varTheta _{w}(\ell ,\ell ^{\prime }) {\rm d} \ell ^{\prime }} \right ]^{1/2} . \end{equation}

This factor will take into account surface compression (through relating $\boldsymbol{\nabla } \psi$ to $\boldsymbol{\nabla } \alpha$ through $\boldsymbol{B}$ ) and local and global shear effects that stabilise the localised ITG modes above the CG.

2.6. Kinetic electrons

An additional factor is the effect of kinetic electrons on ITG mode growth rates. The non-adiabatic response of electrons can significantly increase growth rates and nonlinear heat fluxes of ITG modes compared with the adiabatic electron case (see e.g. Proll et al. Reference Proll, Plunk, Faber, Görler, Helander, McKinney, Pueschel, Smith and Xanthopoulos2022). Several strategies to combat this enhanced drive from trapped particles have been developed in recent years with regards to trapped electron modes, including calculations of the available energy for turbulence (Mackenbach et al. Reference Mackenbach, Proll and Helander2022, Reference Mackenbach, Proll, Wakelkamp and Helander2023; Duff et al. Reference Duff, Faber, Hegna, Pueschel and Terry2025), modifying flux surface shaping in both tokamaks (Garbet et al. Reference Garbet, Donnel, Gianni, Qu, Melka, Sarazin, Grandgirard, Obrejan, Bourne and Dif-Pradalier2024) and stellarators (Gerard et al. Reference Gerard, Pueschel, Geiger, Mackenbach, Duff, Faber, Hegna and Terry2024) or directly targeting the overlap of trapped particles with drift curvature (Proll et al. Reference Proll, Mynick, Xanthopoulos, Lazerson and Faber2016; Gerard et al. Reference Gerard, Geiger, Pueschel, Bader, Hegna, Faber, Terry, Kumar and Schmitt2023). Recent theoretical work (Costello & Plunk, 2025a) focused on the modification of linear, toroidal ITG mode growth rates

(2.10) \begin{align} \omega = \pm \sqrt{-{\omega_{*i}}\int_{-\infty}^{\infty}\frac{d\ell}{B} \hat{\omega}_{di}(\ell)|\delta \phi|^{2} }\left( \tau \int_{-\infty}^{+\infty}|\delta \phi |^{2} \frac{dl}{B} - \frac{\tau}{2} \Sigma_{j} \int_{{1}/{B_{min}}}^{{1}/{B_{max}}} \tau_{B,j} |\overline{\delta \phi_{j}}|^{2} d\lambda\! \right)^{-({1}/{2})}\!, \end{align}

where the bounce average is

(2.11) \begin{equation} \overline {(\boldsymbol{\cdot }\boldsymbol{\cdot }\boldsymbol{\cdot })} = \frac {1}{\tau _{b}}\int ^{\ell _{2}}_{\ell _{1}} \frac {(\boldsymbol{\cdot }\boldsymbol{\cdot }\boldsymbol{\cdot })}{\sqrt {1-\lambda B}} {\rm d}\ell ,\end{equation}

and the bounce time is $\tau_{B} = \int^{{\ell}_{2}}_{{\ell}_{1}} (1-\lambda B)^{-({1}/{2})} d\ell$ . Several effects are at play in (2.10). The first is the integral along the field line of the drift curvature $\hat {\omega }_{di}(\ell ) \propto -K_{d}(\ell )$ , indicating stability based on the size and sign of the average curvature. Negative curvature in this notation indicates stability of the mode, and is a motivating factor for the max-J property (Proll et al. Reference Proll, Helander, Connor and Plunk2012; Rodríguez et al. Reference Rodríguez, Helander and Goodman2024) where J is the parallel adiabatic invariant of particles, and max-J refers to the case where J is a decreasing function of radius, which has led to successful QI designs with reduced turbulence (Goodman et al. Reference Goodman, Xanthopoulos, Plunk, Smith, Nührenberg, Beidler, Henneberg, Roberg-Clark, Drevlak and Helander2024; García-Regaña et al. Reference García-Regaña, Calvo, Sánchez, Thienpondt, Velasco and Capitán2025). The ITG modes can be completely stabilised in cases where the mode extent includes both regions of good and bad curvature, since the average over this region can be of the ‘good’ sign (figure 3 a). However, toroidal magnetic fields must have some regions of bad curvature, and we are exploring ITG modes that only exist in regions of bad curvature. The max-J property itself cannot stabilise such modes, since $|\delta \phi | =0$ in regions of good curvature, by definition (see figure 3 b). Thus the average drive of these modes will be unaffected by the max-J property once they are driven unstable above the CG.

Figure 3. Inverse mirror configurations sidestep mode inertia – idealised schematics for different QI magnetic field shapes along a magnetic field line with different, roughly constant ITG mode potentials but the same QI-like curvature profile $K_{d}$ . (a) A standard broad minimum and narrow maximum which leverages the max-J property. The spread-out mode potential achieves no net curvature drive and is stable. (b) The same geometry but with sufficient drive such that the CG for a localised ITG mode is exceeded, which sits only in the bad curvature well. The mode inertia effect now increases drive for the ITG mode roughly by the factor $(1-\sqrt {1-B_{\text{min}}/B_{\text{max}}} )^{-1}$ . (c) An ‘inverse mirror’ (IM) magnetic field with a narrow minimum and broad maximum avoids the mode inertia effect by removing trapped particles from the bad curvature region ( $\langle \epsilon (\ell ) \rangle \rightarrow 0$ ).

To reduce the growth rates of localised ITG modes in the kinetic electron regime, we are left with the options of reducing the magnitude of the bad curvature in the numerator or reducing the effect of non-adiabatic electrons in the denominator. As already stated, we would prefer to leave the former quantity alone during optimisation. The latter effect was referred to as ‘mode inertia’ (Costello & Plunk, Reference Costello and Plunk2025a ) and results from the presence of trapped particles capable of strongly destabilising ITG modes. Note that, in the limit of a square well magnetic field, for which $B=B_{\min}$ in some central region bounded by step functions with height $B=B_{\max}$ , the second term in the denominator includes the trapped particle fraction $\epsilon = \sqrt {1-B_{\min}/B_{\max}}$ . Since the overall denominator is positive–definite, it reduces in magnitude from the non-adiabatic response and leads to increased growth rates. To make an estimate for spatially varying magnetic fields with ITG modes localised to regions of bad curvature, we carry out the $\lambda$ integration in the second term of the denominator. To evaluate the factor $|\overline {\delta \phi }|^{2}$ , which is the modulus squared of the bounce-averaged potential, we use the Cauchy–Schwarz inequality

(2.12) \begin{equation} |\overline {\delta \phi }|^{2} \leqslant \overline {|\delta \phi |^{2}} \end{equation}

(Helander, Proll & Plunk Reference Helander, Proll and Plunk2013) for which the equality holds in the case of a constant potential. In fact, the case of equality is an upper bound on the destabilising effect of the denominator, so we can use it to estimate the strongest effect of trapped electrons on the mode growth. Assuming a constant potential but this time in the bad curvature region and doing the $\lambda$ integration over the region of finite potential $\delta \phi$ , we find the field line average of $\delta \phi$ squared over B with a spatially varying trapped particle fraction $\epsilon (\ell ) = \sqrt {1-B(\ell )/B_{\max}}$ , where contributions to the integral outside of the bad curvature region are ignored. Note that this is equivalent to using a trial function for the potential which is a step function located entirely in the bad curvature region. The result is

(2.13) \begin{align} \omega = \pm \sqrt {-\omega _{*i} \langle \hat {\omega }_{di} \rangle _{\delta \phi }/ \langle (1-\epsilon (\ell ) \rangle _{\delta \phi } }, \end{align}

where, as in Costello & Plunk (Reference Costello and Plunk2025a ), the volume average including the mode potential is

(2.14) \begin{align} \langle (\boldsymbol{\cdot }\boldsymbol{\cdot }\boldsymbol{\cdot }) \rangle _{\delta \phi } = \int _{-\infty }^{\infty } (\boldsymbol{\cdot }\boldsymbol{\cdot }\boldsymbol{\cdot }) |\delta \phi |^{2} \frac {{\rm d} \ell }{B}. \end{align}

For the purposes of optimisation, we can approximate this average with the definition $\langle (\boldsymbol{\cdot }\boldsymbol{\cdot }\boldsymbol{\cdot })\rangle _d$ used in (2.9), since this estimates the effect of mode localisation in the region of peak bad curvature. In doing so we ignore the factor $1/B$ in the $\ell$ integrals in both the numerator and denominator for simplicity.

Expression (2.13) suggests that we can shift trapped particles to a region of zero or good curvature in order to increase mode inertia (i.e. stabilise the mode), in line with work such as Proll et al. (Reference Proll, Mynick, Xanthopoulos, Lazerson and Faber2016) and Gerard et al. (Reference Gerard, Geiger, Pueschel, Bader, Hegna, Faber, Terry, Kumar and Schmitt2023). An idealised example of this would be the limit of a very narrow square well magnetic field where for the most part $B=B_{\max}$ with a small region of $B=B_{\min}$ , in which the curvature is either zero or stabilising. Such magnetic fields can be achieved by optimising QI configurations to have an IM shape as shown in figure 3(c). Modern optimisation of QI designs (Subbotin et al. Reference Subbotin2006; Sánchez et al. Reference Sánchez, Velasco, Calvo and Mulas2023; Goodman et al. Reference Goodman, Xanthopoulos, Plunk, Smith, Nührenberg, Beidler, Henneberg, Roberg-Clark, Drevlak and Helander2024) has tended to lead towards configurations with a broad, flat magnetic field minimum that traps a large number of particles in a region of small field line curvature. The magnetic field around the corresponding minimum of $B$ tends to be relatively straight and wide in real space while the maximum of $B$ tends to be narrow and exhibit bean-shaped cross-sections reminiscent of that seen in Wendelstein 7-X configurations. However, the IM magnetic field is another possible solution of the near-axis equations in QI geometry. While this type of configuration may struggle to fulfil the max-J property, it nonetheless has the tendency to push bad curvature close to the maximum of $B$ such that $\langle \epsilon (\ell ) \rangle \rightarrow 0$ . In this limit it may be possible to achieve similar levels of ITG stability without enforcing the maximum J property, as we shall later see. We emphasise that it is not required for a magnetic field to be in this extreme limit to achieve reasonable stabilisation, merely that it is one way to stabilise ITG modes via expression (2.13) at finite mirror ratio.

3.1. Small orbit widths

It has been argued by Plunk & Helander (Reference Plunk and Helander2024) and shown via optimisation (Goodman et al. Reference Goodman, Xanthopoulos, Plunk, Smith, Nührenberg, Beidler, Henneberg, Roberg-Clark, Drevlak and Helander2024) that QI configurations with low neoclassical transport tend to have robust zonal flow suppression of turbulence. Theoretically, it was argued that this effect comes from the small orbit width of particles in QI fields (compared with tokamaks or quasisymmetric stellarators), related to the relatively small distance between particle bounce points. By choosing QI configurations as candidates for turbulence optimisation, we seek to leverage this property in addition to the CG and growth rate strategies discussed earlier. Having an additional contour at the minimum of $B$ be poloidally straight (i.e. a contour of constant $\phi _{B}$ ) in the Boozer plane also seems advantageous from this perspective. This forces the magnetic surface to achieve low geodesic curvature as the poloidal derivative of $B$ is small in the vicinity of $B_{\text{min}}$ . Small geodesic curvature has been known to be favourable for zonal flow reduction of electrostatic turbulence (Nunami, Watanabe & Sugama Reference Nunami, Watanabe and Sugama2013; Nakata & Matsuoka Reference Nakata and Matsuoka2022; Nishimoto et al. Reference Nishimoto2024), tied to the damping of geodesic acoustic modes, and the following results suggest this approach indeed works.

Since targeting the minimum of $B$ to lie at constant toroidal angle seems distinct from previous QI optimisations, we would like to classify it differently. To delineate this case, we define an ‘I-number’ (I as in isodynamic) counting the number of poloidally straight contours, which is $I=1$ in the case of a standard QI stellarator (one contour at $B_{\text{max}}$ ) and $I = 2$ approximately in the case of the new optimised configurations presented here (an additional contour at $B_{\text{min}}$ ). Note that quasi-isodynamicity (Cary & Shasharina Reference Cary and Shasharina1997; Plunk et al. Reference Plunk2019) requires poloidally straight contours to come in pairs ( $\phi = \phi _{\text{min}} \pm \phi _{\text{str}}$ ) if they are not located at the minimum, with $\phi _{\min}$ the location of the minimum of $B$ . Omnigenous magnetic fields with no poloidally straight contours such as piecewise omnigenous configurations (Velasco et al. Reference Velasco, Calvo, Escoto, Sánchez, Thienpondt and Parra2024), the various configurations of Wendelstein 7-X, tokamaks, or quasisymmetric stellarators have $I=0$ .

3.2. Radial drift

The primary omnigeneity target minimises a version of the quantity $\epsilon _{\text{eff}}$ (Nemov et al. Reference Nemov, Kasilov, Kernbichler and Heyn1999) to reduce the net radial drift of particles in accordance with omnigeneity. Instead of taking an average over a very long distance of a single field line (in the limit of the distance $s$ going to infinity), we sample a large number of field lines ranging from $\phi =0$ to $\phi =2\pi /n_{fp}$ and assume that, over each field period, a single well in the magnetic field exists, such that additional bounce wells are not taken into the calculation. This is enforced with targets (3.3) and (3.2). The target reads

(3.4) \begin{align} f_{\overline {\epsilon }_{\text{eff}}} & = \sum _{s_{0}} \overline {\epsilon }_{\text{eff}}^{2}(s_{0}) \nonumber\\[-5pt] \overline {\epsilon }_{\text{eff}}^{3/2} & = \pi R_{0}^{2}\left ( \sum _{\alpha } \int _{\ell _{-}(\alpha )}^{\ell _{+}(\alpha )}\frac {{\rm d}\ell }{B} \right )\left ( \sum _{\alpha } \int _{\ell _{-}(\alpha )}^{\ell _{+}(\alpha )} \boldsymbol{\nabla } \psi \frac {{\rm d}\ell }{B}\right )^{-2}\sum _{\alpha } \int ^{B_{\max}(\alpha )}_{B_{\min}(\alpha )}db^{\prime } \varTheta _{B}(b^{\prime }-B) \frac {\hat {H}^{2}}{\hat {I}} \nonumber \\[5pt] \hat {H}& = \frac {1}{b^{\prime }}\int _{\ell _{-}(\alpha )}^{\ell _{+}(\alpha )}\sqrt {b^{\prime }-\frac {B}{B_{0}}}\left (4\frac {B_{0}}{B}-\frac {1}{b^{\prime }}\right )\boldsymbol{\nabla } \psi \kappa _{g} \frac {{\rm d}\ell }{B} \nonumber \\[5pt] \hat {I} & = \int _{\ell _{-}(\alpha )}^{\ell _{+}(\alpha )}\sqrt {1-\frac {B}{B_{0} b^{\prime }}}\frac {{\rm d}\ell }{B}, \end{align}

with $R_{0}$ a fixed major radius used to stop the optimiser from reducing the aspect ratio by deflating the magnitude of $\epsilon _{\text{eff}}$ , $b^{\prime }=B/B_{0}$ measures a specific value of B lying between the local $B_{\min}$ and $B_{\max}$ for a specified field line label $\alpha$ , which contains the toroidal interval $\{0,2\pi \}$ in PEST coordinates (for which the toroidal angle is the standard cylindrical angle, and the poloidal angle is defined to make the field lines straight), while $B_{0}$ is a reference value of $B$ . The step function $\varTheta _{B}$ is $1$ when $b^{\prime } \geqslant B$ and zero otherwise.

4.1. The QI configurations

The starting point for the two new optimised QI configurations was a five-field-period configuration with zero axis helicity. We try to optimise for a high CG using the aforementioned targets, ignoring for the moment growth rates above the threshold. We find that it is necessary to optimise by checking $32$ flux tubes spanning 2 toroidal field periods to catch the most unstable spot on the flux surface. The optimisation is carried out targeting volume-averaged $\beta$ to be $2\,\%$ using artificial, linear pressure profiles that peak at $r=0$ and are zero at $r=a$ . The edge toroidal flux is allowed to vary as an optimisation parameter to keep $\beta$ at this value. We note here that both new configurations possess a low value of $\epsilon _{\text{eff}}$ as well as small geodesic curvature in comparison with the HM configuration (figure 6).

Figure 7. Geometric quantities entering the gyrokinetic equation for the three configurations, for a flux tube centred at the outboard midplane, including drift curvature, normalised magnetic field strength and $\boldsymbol{\nabla } \alpha$ . Top left: for W7-X HM, the metrics somewhat resemble those of a quasisymmetric stellarator near the outboard midplane (no curvature gap at $\ell =0$ ). Top right: the same but for the QICG configuration with $a/L_{T,\text{crit}}=2.14$ , for which the ‘mode splitting’ is clearly visible in the curvature quantity $K_{d}$ , which shows two very separate bad curvature wells near the outboard midplane (near $\ell \sim \pm 4 a$ ). Bottom centre: the IM configuration, similar to the previous case but with less pronounced splitting. The bad curvature has noticeably shifted away from the minimum of $B$ .

The first new configuration, which we refer to as QI and Critical Gradient (QICG), was optimised for QI quality (see previous section), a vacuum magnetic well, an $\iota$ ranging from $1.35$ to $1.50$ and an aspect ratio of $14$ . The targets used were

(4.1) \begin{align} f_{QICG} & = f_{well} + f_{iota} + f_{CG} + f_{asp} \nonumber \\ & \quad + f_{mono} + f_{extrema} + f_{mirror} + f_{\beta } + f_{\overline {\epsilon }_{\text{eff}}}, \end{align}

with

(4.2) \begin{align} f_{asp}(A_{t}) = (A - 14)^{2}, \end{align}

( $A$ is the aspect ratio)

(4.3) \begin{align} f_{\iota }=\left [(\iota (s_{0}=0.0) - 1.35)^{2} + (\iota (s_{0}=1.0) - 1.49)^{2} \right ], \end{align}

and

(4.4) \begin{align} f_{well}=\sum _{r}\left [(1+V^{\prime \prime }(r)/0.01)\varTheta (1+V^{\prime \prime }(r)/0.01) \right ]^{2}, \end{align}

as in Roberg-Clark et al. (Reference Roberg-Clark, Plunk, Xanthopoulos, Nührenberg, Henneberg and Smith2023) with $r/a=\{0.0,0.25,0.95\}$ , and

(4.5) \begin{equation} f_{\beta }=(\beta - 0.02)^{2}, \end{equation}

with $\beta$ the volume-averaged value calculated from VMEC. The targeted surfaces $r/a$ for $\epsilon _{\text{eff}}$ were $r/a=(0.02,0.10,0.20,0.30,0.40,0.50,0.60,0.70,0.80,0.90)$ and varying weights were used for all of the terms over the course of the optimisation to prioritise at first QI quality and then the CG metric (2.8). The vacuum magnetic well (after removing pressure from the equilibrium) is roughly 0.25% (not shown), while at $2\,\%$ $\beta$ it has a low $\epsilon _{\text{eff}}$ of roughly $0.2\,\%$ on axis. Figure 5 suggests that transport is relatively low below $a/L_{T}=2.14$ and stays moderate above that point. The primary effect leading to the large CG of QICG is the mode splitting effect, which is clearly demonstrated in the top right-hand plot of figure 7, with an additional boost to $g^{\alpha \alpha }$ through a localised shear effect.

Figure 8. Nonlinear flux tube ITG simulations with kinetic electrons, a fixed temperature gradient $a/L_{T}=a/L_{T,i}=a/L_{T,e}=3$ and variable density gradient $a/L_{n}$ for both species at half-radius at the most unstable location. For W7-X, the flux tube is $\alpha =0$ and for IM, it is $\alpha =\pi /n_{fp}$ (see caption for figure 5). The heat flux for the IM case remains below or comparable to that of the W7-X HM configuration up to $a/L_{n}=2$ .

Since we have not taken into account stability above the threshold, initial nonlinear results (not shown) reveal prohibitively large heat fluxes, likely exacerbated by the mode inertia effect. We re-optimise the QICG configuration so that the ITG growth rates are not too large, yielding the configuration we call ‘IM’. To reduce linear ITG growth rates with kinetic electrons, we construct a simple target that combines the factors discussed in §§ 2.5 and 2.6, which reads

(4.6) \begin{equation} f_{Q} = (\left [ \langle |\boldsymbol{\nabla } \alpha | \rangle (1 - \langle \epsilon (\ell ) \rangle _d ) \right ]^{-1} - t_{Q})^{2}, \end{equation}

with the averages calculated as in expression (2.9). Using a new set of targets

(4.7) \begin{equation} f_{IM} = f_{QICG} + f_{Q}, \end{equation}

with $t_{Q}=2.0$ (QICG values exceeded 3) as well as a CG of 2.0 (2.7), we arrive at the configuration in figure 4, with a significant positive (tokamak-like) shear, a rotational transform ranging from 1.3 to 0.98 (near targeted values $1.25$ to $1.01$ ), an $\epsilon _{\text{eff}}$ of roughly $0.2\,\%$ on axis (figure 6 c), $A=13$ (as targeted) and a vacuum magnetic well of $0.17\,\%$ . In particular, the positive shear seems to be helpful for the mode inertia effect, as the bad curvature region appears to be shifted away from the minimum of $B$ . We see that the configuration has favourable heat fluxes compared with W7-X HM when kinetic electrons are included (figure 8).