Navigation

Developments

O1 (WP1) – The theory of transport and machine learning architectures

Deliverable: A comprehensive turbulent transport model

Within this work package, we established the theoretical foundations for describing the dynamics of charged particles in fusion plasmas characterized by turbulence (Task 1.1), for the correct description of turbulent random fields (Task 1.2), and for the main machine-learning tools relevant to the purpose of this project (Task 1.3).

  • T1.1: Modeling the equations of motion for particles
  • We consider a population of charged particles, with mass $m$ and charge $q$, confined in a tokamak configuration dominated by a strong magnetic field $\mathbf{B}$. The usual particle coordinates in phase space are $(\mathbf{x},\mathbf{v})$. When the magnetic field is sufficiently strong compared to other electromagnetic components, the motion can be viewed as a superposition of a smooth large-scale motion and a fast Larmor gyration. This scale separation is the basis of Lie perturbation theory [1], which provides a rigorous mathematical framework for charged-particle dynamics in magnetized plasmas.

    In this approach, the true 6D phase space $(\mathbf{x},\mathbf{v})$ is replaced by the reduced gyrocentre coordinates $(\mathbf{X},v_{\parallel},\mu)$, which suppress the fast gyromotion through a perturbative expansion based on gyrokinetic ordering [2]. In this project we adopt the high-flow ordering [3], explicitly retaining the polarization drift [4].

    The gyrocentre dynamics satisfy

    $$ \frac{d\mathbf{X}}{dt} = v_{\parallel}\frac{\mathbf{B}^\star}{B_{\parallel}^\star} + \frac{\mathbf{E}^\star \times \mathbf{b}}{B_{\parallel}^\star} \qquad\text{(1)} $$

    $$ \frac{dv_{\parallel}}{dt} = \frac{q}{m}\frac{\mathbf{E}^\star\cdot\mathbf{B}^\star}{B_{\parallel}^\star} \qquad\text{(2)} $$

    $$ \frac{d\mu}{dt}=0 \qquad\text{(3)} $$

    where $$\mathbf{E}^\star = -\nabla\Phi^\star - \frac{\partial \mathbf{A}^\star}{\partial t}, \qquad \mathbf{B}^\star = \nabla\times\mathbf{A}^\star,$$ and

    $$ q\Phi^\star = \frac{m v_\parallel^2}{2} - \frac{m u^2}{2} + \mu B + q\phi_1^{\text{neo}} + q\phi_1^{\text{gc}} \qquad\text{(4)} $$

    $$ \mathbf{A}^\star = \mathbf{A}_0 + \frac{m}{q}(v_{\parallel}\mathbf{b}+u+v_E) + \mathbf{A}_1^{\text{gc}} \qquad\text{(5)} $$

    The zeroth-order magnetic potential $\mathbf{A}_0$ satisfies $\mathbf{B} = \nabla\times\mathbf{A}_0$. For axisymmetric equilibria, the field can be written as: $$ \mathbf{B} = F(\psi)\nabla\phi+\nabla\phi\times\nabla\psi, $$ where $\phi$ is the toroidal angle in cylindrical $(R,Z,\phi)$ or toroidal $(r,\theta,\phi)$ coordinates (COCOS = 2) [5].

    The local safety factor $q_\theta(\psi,\theta)$, its flux-surface average $\bar{q}(\psi)$, and the generalized poloidal angle $\chi(\psi,\theta)$ are defined by:

    $$ q_\theta(\psi,\theta) = \frac{\mathbf{B}\cdot\nabla\phi}{\mathbf{B}\cdot\nabla\theta} \qquad\text{(6)} $$

    $$ \bar{q}(\psi) = \frac{1}{2\pi}\int_0^{2\pi} q_\theta(\psi,\theta)\,d\theta \qquad\text{(7)} $$

    $$ \left.\frac{\partial\chi}{\partial\theta}\right|_{\psi} = \frac{q_\theta(\psi,\theta)}{\bar{q}(\psi)} \qquad\text{(8)} $$

    Zeroth-order MHD equilibrium is toroidally invariant and constant on magnetic surfaces, so $n=n(\psi)$, $P=P(\psi)$, $T_i=T_i(\psi)$, $T_e=T_e(\psi)$.

    A macroscopic radial electric field $E_0$ produces a toroidal rotation $$ \mathbf{u} = R^2\Omega_t(\psi)\,\nabla\phi, $$ where $\Omega_t(\psi)$ may have a gradient. The equations of motion are expressed in a frame rotating with $\mathbf{u}$ [3,7]. Poloidal flows are neglected [5], but a neoclassical poloidal potential $\phi_1^{\text{neo}}(\psi,\theta)$ appears:

    $$ \phi_1^{\text{neo}} = \frac{m\Omega_t(\psi)^2}{2|e|\,(1+T_i/T_e)} \left(R^2 - \langle R^2\rangle_\theta\right) \qquad\text{(9)} $$

    The first-order gyrocentre potentials $\phi_1^{\text{gc}}$ and $\mathbf{A}_1^{\text{gc}}$ include turbulent fluctuations and small external perturbations (RMPs). Gyro-averaging gives:

    $$ \tilde{\phi}_1^{\text{gc}}(k,t) = \tilde{\phi}_1(k,t)\, J_0(k_\perp \rho_L), $$ with $\rho_L=\sqrt{2m\mu}/(qB)$.

    The drift-wave turbulence relevant for ions is dominated by ITG/TEM modes [8].

    Circular geometry

    In circular approximation, $$ \mathbf{B} = B_0 R_0\,\nabla\phi + \frac{r b_\theta(r)}{R}\,\nabla\theta \qquad\text{(10)} $$ $$ F(\psi)=B_0 R_0 \qquad\text{(11)} $$ $$ \psi(r)=B_0R_0\int_0^r b_\theta(r')dr' \qquad\text{(12)} $$ $$ b_\theta(r)=\frac{r\,\bar{q}(r)}{R_0^2-r^2} \qquad\text{(13)} $$ $$ \chi(r,\theta)=2\tan^{-1} \left[\frac{R_0-r}{R_0+r}\tan\left(\frac{\theta}{2}\right)\right] \qquad\text{(14)} $$

    Solov’ev equilibria

    With linear profiles: $$ \mu_0\frac{dP}{d\psi}=A, \qquad \frac{dF^2}{d\psi}=\frac{2 A R_0^2 \gamma}{1+\alpha^2} \qquad\text{(15)} $$ The poloidal flux is $$ \psi=\frac{A}{2(1+\alpha^2)} \left[R^2 - \gamma R_0^2 - Z^2 + \frac{\alpha^2}{4}(R^2-R_0^2)^2\right] \qquad\text{(16)} $$ Minor radius: $$ a=\frac{R_0}{2\sqrt{2-\gamma-\sqrt{\gamma}}} \qquad\text{(17)} $$ Elongation: $$ \kappa=\frac{\alpha}{\sqrt{1-\gamma/(2-\gamma)}} \qquad\text{(18)} $$ Current: $$ F(\psi)=B_0R_0\left[1+\frac{\psi}{2A\gamma B_0(1+\alpha^2)}\right] \qquad\text{(19)} $$

    Experimentally reconstructed equilibria

    Experimental magnetic equilibria (G-EQDSK format) provide $\psi(R,Z)$ and profiles $F(\psi)$, $\bar{q}(\psi)$ on numerical grids. They are generated by reconstruction tools such as CHEASE [9], EFIT [10], PLEQUE [11], NICE [12], and EQUINOX [13]. T3ST uses these equilibria when analyzing specific discharges.

  • T1.2: Modeling the turbulence
  • Given the definitions of the magnetic equilibrium quantities $q_\theta(\psi,\theta)$, $\bar{q}(\psi)$, and $\chi(\psi,\theta)$ (Eqs. (6–8)), one can show that the Clebsch form of the magnetic field is valid:

    $$ \mathbf{B} = \nabla(\phi - \bar{q}\,\chi)\times\nabla\psi. $$

    For representing small-scale turbulent fluctuations, most gyrokinetic (GK) codes adopt field-aligned coordinates $(x,y,z)$ [14], which simplify the description of fluctuations elongated along magnetic field lines. These coordinates correspond to the radial, binormal, and parallel directions ($\mathbf{B}\propto\nabla y\times\nabla x\propto\partial \mathbf{r}/\partial z$):

    $$ x = C_x f(\psi) \qquad\text{(20)} $$

    $$ y = C_y\,( \phi - \bar{q}\,\chi ) \qquad\text{(21)} $$

    $$ z = C_z\,\chi \qquad\text{(22)} $$

    In practice, T3ST uses constant values $C_x=a$ (the tokamak minor radius), $C_y=r_0/\bar{q}(r_0)$ (with $r_0$ a reference radius), $C_z=1$, and defines $f(\psi)=\rho_t = \Phi_t(\psi)/\Phi_t(\psi_{\text{edge}})$, where $\rho_t\in[0,1]$ is the normalized effective radius. The toroidal magnetic flux is

    $$ \Phi_t(\psi)=\int_{\text{axis}}^{\psi} \mathbf{B}\cdot\mathbf{e}_\phi\, dS(\psi'), $$

    thus $\rho_t \approx r/a$ (in circular geometry).

    Instead of computing turbulent fields self-consistently, T3ST models them statistically as an ensemble of random fields $\{\phi_1, A_1\}$ with prescribed statistical properties, consistent with experimental ITG/TEM observations. Several moderate assumptions are used:

    Based on these assumptions, T3ST constructs an ensemble of random fields $\phi_1(x,t)$ from white-noise fields in Fourier space $\eta(\mathbf{k},\omega)$ satisfying $$ \langle \eta(\mathbf{k},\omega)\,\eta(\mathbf{k}',\omega')\rangle = \delta(\mathbf{k}+\mathbf{k}')\,\delta(\omega+\omega'). $$ The Fourier components of the potential are

    $$ \tilde{\phi}_1(\mathbf{k},\omega) = \sqrt{S(\mathbf{k},\omega)}\,\eta(\mathbf{k},\omega), $$

    ensuring Gaussianity and correct correlations.

    In toroidal geometry, the fields can be expressed through a double Fourier series:

    $$ \phi(x,t)=\sum_{n,m}\phi_{nm}(r,t)\, e^{i(n\phi+m\theta)} \qquad\text{(23)} $$

    which preserves natural periodicity. However, due to the anisotropic nature of turbulence (long correlation lengths along the field, short perpendicular), field-aligned representation is preferred. Thus T3ST evaluates $\phi_1$ in field-aligned coordinates, even though magnetic drifts are computed in cylindrical coordinates.

    The random field is represented as:

    $$ \phi_1(x,y,z,t) =\int dk\, d\omega\; \tilde{\phi}_1(\mathbf{k},\omega)\; e^{\,i(k_x x + k_y y + k_z z - [\omega^\star(\mathbf{k})+\omega] t)} \qquad\text{(24)} $$

    For $k_\parallel\ll k_\perp$, the correspondence with toroidal mode numbers is approximately:

    $$ k_y \approx \frac{n}{C_y},\qquad m = -\big[\bar{q}(x_0)\,n\big] + \Delta m,\qquad k_z = \frac{\Delta m + \{\bar{q}(x_0)n\}}{C_z}, \qquad\text{(25)} $$

    where $[\cdot]$ and $\{\cdot\}$ denote the integer and fractional parts, and $n,\Delta m\in\mathbb{Z}$.

    The spectrum $S(\mathbf{k},\omega)$ is modeled analytically to reproduce saturated drift-wave turbulence characteristics:

    $$ S = A_\phi^2 \, \frac{\tau_c\,\lambda_x\lambda_y\lambda_z}{(2\pi)^{5/2}}\, \frac{ \exp\!\left[-(k_x^2\lambda_x^2 + k_z^2\lambda_z^2)/2\right] }{ 1+\tau_c^2\omega^2 }\, \Big[ e^{-\frac{(k_y-k_0)^2\lambda_y^2}{2}} \;-\; e^{-\frac{(k_y+k_0)^2\lambda_y^2}{2}} \Big] \qquad\text{(26)} $$

    The parameters $\lambda_x,\lambda_y,\lambda_z$ are the correlation lengths in the field-aligned directions; $k_0$ selects the most unstable mode $$ k_{y,\text{max}} \approx \frac{k_0}{2} + \sqrt{\frac{k_0^2}{4} + \frac{2}{\lambda_y^2}}. $$ The correlation time $\tau_c$ measures deviations from the dispersion relation, and $A_\phi$ sets the turbulence amplitude. Typical parameters:

    $$ \lambda_x \approx 5\rho_i,\quad k_0 \approx 0.1/\rho_i,\quad \lambda_y \approx 5\rho_i,\quad \lambda_z \approx 1,\quad \tau_c \approx R_0/v_{\text{th}},\quad |e|A_\phi \approx 1\%\,T_i. $$

    Since $\lambda_z\approx 1$, we obtain $\lambda_\parallel\approx\bar{q}R_0$, consistent with experimental observations.

    Drift-wave turbulence in tokamaks is generally a superposition of ITG and TEM modes, weighted by coefficients $A_i$ and $A_e$:

    $$ \phi_1 = A_i\,\phi_1^{\text{ITG}} + A_e\,\phi_1^{\text{TEM}}, \qquad A_i + A_e = 1. $$

    It is well known that magnetic curvature, shear, internal transport barriers, ELM activity, and blob structures can induce non-Gaussian behavior and spatial variations of turbulence. Thus the assumptions of Gaussianity and homogeneity are only approximate. Nevertheless, numerical results (Sec. 4) show that particle transport is predominantly local, with particles moving only a few Larmor radii from their initial surface. Therefore, although not globally exact, these assumptions remain valid locally for the purposes of the present analysis.

    A possible extension recently implemented in the T3ST code is the ability to include inhomogeneities along the direction parallel to the magnetic field (“z”) through a simple prefactor:

    $$ \phi_1 \rightarrow \phi_1 \, g(z) $$

    $$ g(z) = e^{-\, z^2 / (2 \Lambda_z^2)} $$

    where \( \Lambda_z \) is a measure of the parallel correlation and of the “ballooning” character specific to electrostatic turbulence.

  • T1.3: Modeling the architecture of machine learning codes
  • O2 (WP2) – Developing and testing numerical codes

    Deliverable: 1 code for test particle simulations, 2 codes for machine learning

    O3 (WP3) – The construction of the database: extensive simulations

    Deliverable: Database, trained and tested machine learning tools

    ...
    O4 (WP4) – Predictions and physical mechanisms

    Deliverable: Characterization of transport mechanisms

    ...

    ← Return to the main page