Spectrally accurate reverse-mode differentiable bounce-averaging operator and its applications
K. Unalmis \aff 1 \corresp, R. Gaur \aff 2, R. Conlin \aff 3, D. Panici \aff 2, E. Kolemen \aff 2,4 · 2024
Read Spectrally accurate reverse-mode differentiable bounce-averaging operator and its applications by K. Unalmis \aff 1 \corresp, R. Gaur \aff 2, R. Conlin \aff 3, D. Panici \aff 2, E. Kolemen \aff 2,4 (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-01724.txt.
Introduction
Kaya Unalmis and four co-authors build a fast, highly accurate algorithm for averaging particle motion in fusion devices, designed so its derivatives come free, in a paper from twenty twenty-four. Bounce-averaging simplifies the kinetic models used to design stellarators, machines that hold plasma in twisted magnetic fields. Optimization over many design knobs needs gradients that older tools supplied slowly or not at all.
Built into the DESC design suite, the algorithm reaches exponential accuracy and supports reverse-mode automatic differentiation. That makes gradient cost independent of how many knobs are tuned. The team used it to directly shrink a ripple transport measure in a working-pressure stellarator, a first. They also point to gains for particle confinement and turbulence studies.
Catalog record
| CREATOR | K. Unalmis \aff 1 \corresp, R. Gaur \aff 2, R. Conlin \aff 3, D. Panici \aff 2, E. Kolemen \aff 2,4 |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-01724
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Spectrally accurate reverse-mode differentiable bounce-averaging operator and its applications K. Unalmis \aff 1 \corresp, R. Gaur \aff 2, R. Conlin \aff 3, D. Panici \aff 2, E. Kolemen \aff 2,4 Abstract We present a spectrally accurate, automatically differentiable bounce-averaging operator implemented in the DESC stellarator optimization suite. Using this operator, we can perform efficient optimization of many objectives to improve stellarator performance, such as the ϵ eff 3 / 2 \epsilon_{\mathrm{eff}}^{3/2} proxy for the neoclassical transport coefficient in the 1 / ν 1/\nu regime. By employing this differentiable approximation, for the first time, we directly optimize a finite- β \beta stellarator to reduce neoclassical transport using reverse-mode differentiation, ensuring that the computational cost of determining the gradients is independent of the number of input parameters. Spectrally accurate differentiable bounce-averaging 1 Introduction Stellarators, first conceived by spitzer1958stellarator , represent a distinct approach to magnetic confinement fusion that offers unique advantages over tokamaks. These toroidal devices achieve plasma confinement through external magnetic fields rather than through plasma current, providing greater design flexibility and operational stability. The absence of a continuous toroidal symmetry allows for magnetic field optimization through boundary shaping, which helps minimize the net toroidal current and thereby avoid current-driven plasma instabilities and disruptions that plague tokamak operation ( helander2014theory ) . The design of optimal stellarator configurations is a complex optimization problem involving hundreds of degrees of freedom. Traditional optimization approaches have evolved significantly over the past decades. The VMEC (Variational Moments Equilibrium Code), developed by hirshman1983steepest_VMEC , has served as the foundation for numerous stellarator optimization efforts. Building upon VMEC , several optimization frameworks have emerged: STELLOPT ( stellopt ; spong1998stellopt ) , which implements a suite of physics-based optimization criteria; ROSE ( drevlak2018optimisation ) , which focuses on coil optimization and engineering constraints; and more recently, SIMSOPT ( landreman2021simsopt ) , which provides a flexible approach to stellarator optimization. However, in general, traditional approaches rely on finite-difference techniques to compute gradients. Such techniques may not yield an accurate estimate of the gradient which can affect the optimizer convergence. Furthermore, the objective function must be recomputed to estimate the gradient for each optimizable parameter. In DESC, unlike previous optimizers, it is not necessary to resolve the MHD force balance equation at each optimization step. Additional objectives that depend on equilibrium force balance can be optimized simultaneously on a single device while ensuring ideal MHD force balance. In this paper we present a fast, automatically-differentiable bounce-averaging operator which is used to simplify kinetic models such as drift and gyrokinetics. Specifically, we will focus on neoclassical transport in the 1/\nu regime by evaluating and optimizing for the effective ripple ( nemov1999evaluation ) . The bounce-averaging operators and drift kinetic model is explained in the following sections. This operator is implemented as a part of the DESC ( dudt2020desc ; panici2023desc ; conlin2023desc ) optimization suite. In section 2 , we will briefly explain how we solve the ideal MHD equation and calculate an equilibrium using the DESC suite. In section 3 , we will introduce the drift kinetic model and explain the derivation of neoclassical transport in the 1/\nu collisionality regime, as done similarly by Nemov. In section 4 , we explain our implementation of the reverse-mode differentiable proxy and demonstrate how it is faster than forward-mode computation. Section 5 describes the numerical methods for the optimization. In section 6 , we present comparison of the effective ripple with the NEO code from the STELLOPT suite, and we optimize against neoclassical transport for various stellarator equilibria in section 7 . In section 8 , we conclude this work and explain how it can be extended. 2 Ideal MHD equilibrium Our study only concerns solutions whose magnetic field lines lie on closed nested toroidal surfaces, known as flux surfaces. We label the flux surfaces with their enclosed toroidal flux \psi . On each flux surface, curves of constant \alpha correspond to magnetic field lines. A divergence-free magnetic field can be written in the Clebsch form ( d2012flux ) . \boldsymbol{B}=\boldsymbol{\nabla}\psi\times\boldsymbol{\nabla}\alpha. (1) In a set of specialized coordinates, known as flux coordinates, \alpha=\theta_{\rm{P}}-\iota\phi where \theta_{\rm{P}} is a PEST poloidal angle, \phi is the toroidal angle of the cylindrical coordinate system (R,\phi,Z) , and the quantity \iota=\frac{\boldsymbol{B}\cdot\boldsymbol{\nabla}\theta_{\rm{P}}}{\boldsymbol{B}\cdot\boldsymbol{\nabla}\phi}, (2) is the pitch of the field line on a flux surface, commonly referred to as the rotational transform. The definition of \boldsymbol{B} in ( 1 ) is consistent with ( 2 ). We solve the steady-state, ideal MHD (MagnetoHydroDynamic) equation \boldsymbol{\nabla}\left\lparen\mu_{0}p+\frac{\lvert B\rvert^{2}}{2}\right\rparen=\boldsymbol{B}\cdot\boldsymbol{\nabla}\boldsymbol{B} (3) which describes a balance between the plasma pressure p , magnetic field pressure \lvert B\rvert^{2} and the effect of field line curvature \boldsymbol{B}\cdot\boldsymbol{\nabla}\boldsymbol{B} . Unlike a tokamak, we cannot simplify ( 3 ) to a two-dimensional scalar Grad-Shafranov equation. Assuming the existence of nested flux surfaces, we can only reduce ( 3 ) to two coupled three-dimensional equations. The maps in the MHD equation are derived from Fourier-Zernike series. These orthogonal polynomials enable efficient computation of derivatives and ensures smoothness of the solution, in particular regularity near the magnetic axis ( dudt2020desc ; panici2023desc ) . This boundary value problem is then solved as a minimization problem using a trust-region method. Since DESC is written in JAX, the computation is typically accelerated on a GPU. In the following section, we explain the drift-kinetic model. 3 Neoclassical model of plasma The dynamics of a magnetized hot plasma differ significantly from that of an unmagnetized fluid. Unlike isotropic hard-sphere collisions that govern the behavior of an uncharged fluid, a plasma behaves differently in directions perpendicular and parallel to the magnetic field lines because of Coulomb collisions. In magnetized plasmas, particles traverse helical trajectories around magnetic field lines, exhibiting motion characterized by gyration around these lines and drifts across them. The classical transport model assumes a simplistic view of particle collisions and does not adequately incorporate the effects of these drifts. To correctly account for the effect of drifts, trapped and passing particles, and the magnetic geometry, we use the neoclassical transport theory. There are three fundamental length and time scales relevant to magnetized plasmas. The time scales correspond to the particle transit frequency v_{\mathrm{th,s}}/L , where v_{\mathrm{th,s}}=\sqrt{2T_{s}/m_{s}} is the thermal speed, the Coulomb collision frequency \nu_{ss^{{}^{\prime}}}\propto T^{3/2} , and the gyration (cyclotron) frequency \Omega_{s}=Z_{s}e\lvert B\rvert/(m_{s}c) where s,s^{{}^{\prime}} are the species of interest. For each time scale, the corresponding length scales are the gradient scale length of the magnetic field L , the mean free path \lambda_{\mathrm{mfp}} , and the gyroradius \rho_{s}=v_{\mathrm{th},s}/\Omega_{s} , respectively. In a magnetized plasma, \nu_{ss^{\prime}}\sim\frac{v_{\mathrm{th,s}}}{L}\ll\Omega_{s}\qquad\lambda_{\mathrm{mfp}}\sim L\gg\rho_{s} (4) Using a random walk estimate, we can calculate the classical heat transport coefficient in the perpendicular direction as D_{\perp}\sim\nu_{ss^{{}^{\prime}}}\rho_{s}^{2}\sim T^{5/2} ( helander2005collisional ) whereas, using neoclassical theory, we have \Delta x\sim\rho_{s}\lvert B\rvert/\lvert B_{\mathrm{p}}\rvert , \lvert B\rvert and \lvert B_{p}\rvert are the total and poloidal magnetic field strength, respectively which gives D_{\perp}\sim\nu_{ss^{{}^{\prime}}}\rho_{s}^{2}\lvert B\rvert^{2}/\lvert B_{p}\rvert^{2}\sim T^{5/2}\lvert B\rvert^{2}/\lvert B_{p}\rvert^{2} . The ratio \lvert B\rvert/\lvert B_{\mathrm{p}}\rvert strongly depends on the magnetic field geometry and significantly affects the regime of neoclassical transport. A magnetized plasma can either be weakly or strongly collisional. This is defined by the collisionality \nu_{*}\equiv L/\lambda_{\mathrm{mfp}}=\nu/(v_{\mathrm{th,s}}/L) . In a strongly collisional plasma, particles undergo frequent collisions without covering a significant distance along a magnetic field line,