Bayesian optimization of stellarator alpha-particle confinement using data-informed parameter spaces and dimensionality reduction

This paper proposes two data-informed parameterization methods—quantile transformation of Fourier modes and PCA-based dimensionality reduction—to enable efficient Bayesian optimization of stellarator shapes for superior alpha-particle confinement, overcoming the limitations of traditional Fourier-based approaches regarding parameter bounds and expressiveness.

The Gist

Imagine you are trying to design the perfect shape for a futuristic, donut-shaped fusion reactor (called a stellarator). The goal is to keep super-hot, fast-moving particles (like alpha particles) trapped inside the magnetic "bottle" so they don't escape and hit the walls.

The problem is that figuring out the right shape is incredibly difficult. It's like trying to sculpt a masterpiece while blindfolded, using a tool that has thousands of knobs to turn. If you turn the knobs randomly, the shape might twist into a knot (self-intersect) and break, or it might be so boring that it doesn't trap anything.

Here is how the authors of this paper solved that problem, using simple analogies:

1. The Problem: A Messy Control Panel

Traditionally, scientists describe the shape of the reactor using a long list of numbers called Fourier amplitudes (think of these as the settings on a giant sound equalizer).

• The Issue: Some knobs on this equalizer control tiny details, while others control huge shapes. If you try to turn them all at once, the "tiny" knobs get lost, and the "huge" knobs might twist the reactor into a knot.
• The Constraint: You can't just turn the knobs anywhere. If you turn them too far, the reactor shape breaks. If you don't turn them far enough, you can't find the perfect shape. It's a frustrating "Goldilocks" problem where the range of safe settings is tiny and hard to find.

2. The Solution: A New Map and a New Compass

The authors created two new ways to navigate this design space, using a "map" based on shapes that already exist and work well.

Method A: The "Quantile Transformation" (The Fair Ruler)
Imagine you have a bag of marbles of all different sizes. If you want to pick one at random, you usually pick based on size, which is unfair if you have 1,000 tiny marbles and only 1 giant one.

• What they did: They took all the existing, successful reactor shapes and created a "fair ruler." They mapped every possible shape to a number between 0 and 1.
• The Result: Now, instead of turning a knob that might break the machine, you just pick a number between 0 and 1. Every number is equally likely to produce a valid, non-broken shape. It's like turning a dial from "0" to "100" where every setting is guaranteed to be safe.

Method B: The "PCA" (The Compression Tool)
Imagine you have a library of 1,000 different reactor blueprints. You notice that most of them are just slight variations of a few core ideas.

• What they did: They used a mathematical technique called Principal Component Analysis (PCA) to find the "main ingredients" of these shapes. Instead of needing 1,000 knobs, they found that you only need about 20 "master knobs" to create almost any valid shape.
• The Result: This shrinks the control panel from a massive wall of switches down to a small, manageable remote control. It makes the search for the perfect shape much faster and easier.

3. The Test: The "Smart Search"

To prove these new maps work, the authors used a Bayesian Optimization algorithm.

• The Analogy: Imagine you are looking for the best spot to set up a campfire in a huge forest, but you can only light one fire at a time. A "dumb" search would try random spots. A "smart" search (Bayesian) remembers where the fires burned well and where they failed, then uses that memory to guess the next best spot.
• The Twist: They added a "smart timer." If a shape looks like it's going to let particles escape quickly, they stop the test early to save time. If a shape looks promising, they let the test run longer to get a precise score.

4. The Results: Breaking the Rules

The authors found five new reactor shapes that trap particles incredibly well.

• The Surprise: For decades, scientists thought you had to make the reactor perfectly symmetric (like a perfect circle or a specific spiral) to trap particles.
• The Discovery: These new shapes are not perfectly symmetric. They are weird, twisted, and irregular. Yet, they trap the particles better than many of the "perfect" shapes.
• The Takeaway: It turns out you don't need a perfect, symmetrical shape to build a great fusion reactor. You just need a shape that is "smart" enough to keep the particles inside, even if it looks messy.

Summary

The paper introduces a new way to design fusion reactors by:

1. Normalizing the controls so you can't accidentally break the design.
2. Simplifying the controls so you don't have to search through millions of options.
3. Proving that you can build excellent reactors that look nothing like the "perfect" shapes scientists used to think were required.

It's like realizing that to build a great house, you don't need to follow a rigid, symmetrical blueprint; you just need to use a smart set of tools to find a shape that keeps the rain out, even if the roof looks a little wobbly.

Technical Summary

Technical Summary: Bayesian Optimization of Stellarator Alpha-Particle Confinement

Problem Statement
Modern stellarator design relies on optimizing the plasma boundary shape, typically parameterized by Fourier amplitudes. This traditional parameter space presents significant challenges for global optimization algorithms, particularly Bayesian optimization (BO). First, the scales of Fourier modes vary drastically (high-order modes have much smaller amplitudes than low-order modes), violating the assumption that input dimensions should have similar scales for efficient optimization. Second, many global algorithms require bound constraints (box constraints), but setting these bounds for Fourier amplitudes is non-trivial: wide constraints often lead to self-intersecting, unphysical boundaries where Magnetohydrodynamic (MHD) equilibrium calculations fail, while tight constraints limit the expressiveness of the geometry, potentially excluding optimal shapes. Furthermore, high-dimensional Fourier spaces make it computationally prohibitive to include high-fidelity physics calculations, such as direct guiding-center tracing for energetic particle confinement, within the optimization loop.

Methodology
The authors propose two new "data-informed" parameter spaces designed to address scaling, bound constraints, and dimensionality issues. Both methods utilize a dataset of existing stellarator boundaries (drawn from experimental configurations like W7-X and LHD, and databases like QUASR and ConStellaration) to define a prior probability distribution on boundary shapes.

1. Quantile-Transformed Fourier Space:

   • A quantile transformation is applied independently to each Fourier amplitude in the dataset.
   • This maps the empirical cumulative distribution function (CDF) of each parameter to a uniform distribution on the unit interval $[0, 1]$.
   • The transformed variables become the new degrees of freedom. This approach naturally bounds parameters to $[0, 1]$, ensures comparable scales across dimensions, and is robust to outliers. The transformation is piecewise-linear, preserving compatibility with gradient-based methods if needed.

2. PCA-Based Parameter Space:

   • Principal Component Analysis (PCA) is first applied to the boundary data (represented as $(R, Z)$ coordinates on a grid).
   • A quantile transformation is then applied to the amplitudes of the principal components.
   • This method offers dimensionality reduction, capturing the variance of complex boundary shapes with a small number of parameters while maintaining natural bounds.

Both methods allow for weighted sampling to prevent over-representation of specific classes of configurations (e.g., quasisymmetric designs) in the training data. The authors also detail procedures to fix the major radius, minor radius, and aspect ratio independently of the parameter space definition.

Optimization Application
The new parameter spaces are demonstrated via Bayesian optimization to minimize alpha-particle losses.

• Objective Function: A novel objective function is defined based on the time required for alpha-particle losses to reach a threshold ($M=2\%$). If losses exceed the threshold early, the objective is penalized based on the time to reach it; otherwise, it is based on the total loss at a fixed time ($t_{max}=0.1$ s). This acts as a multi-fidelity method, terminating tracing early for poor configurations to save computational resources.
• Algorithm: Asynchronous parallel Bayesian optimization is employed using the Ax and TURBO libraries. The workflow utilizes multiple GPUs (via the CATAPULT code) to trace guiding-center trajectories concurrently, with a manager process updating the Gaussian process surrogate model as results return.
• Physics Model: MHD equilibria are computed using vmec++. Guiding-center trajectories for 3.5 MeV alpha particles are traced without collisions in Boozer coordinates. The device scale is fixed to a maximum magnetic field of 12 T, with pressure profiles typical of high-field stellarator reactors.

Key Results
The authors present five optimized stellarator configurations (Configurations 1–5) generated using these methods:

• Confinement Performance: All five configurations exhibit excellent alpha-particle confinement. Configurations 1–3 (unconstrained) achieve energy losses $< 0.06\%$. Configurations 4–5, which include a constraint for Mercier stability, achieve losses $\le 1\%$.
• Comparison to Baselines: When compared to existing configurations (including ConStellaration designs optimized for quasi-isodynamicity and W7-X variants) under identical scaling and initialization, the new configurations show significantly lower losses (often an order of magnitude better).
• Field Topology: The optimized fields are notably far from ideal quasisymmetric (QS), omnigenous, or quasi-isodynamic (QI) patterns. For instance, the magnetic field contours near $B_{max}$ in most configurations do not close poloidally, toroidally, or helically, violating omnigenity. Configuration 3 exhibits a hybrid structure resembling piecewise-omnigenous fields.
• Neoclassical Transport: Despite the lack of strict symmetry, the optimized configurations exhibit low effective ripple ($\epsilon_{eff} < 1\%$) in the core, comparable to or better than W7-X, suggesting that optimizing for fast-ion confinement inherently improves neoclassical thermal transport.
• Parameter Space Efficacy: The data-informed spaces significantly improve the "usable fraction" ($f_{us}$) of the design space (the fraction of parameter space where MHD equilibrium converges) compared to baseline Fourier spaces. The PCA method further demonstrates that a rich variety of strongly shaped boundaries can be represented with a reduced number of dimensions (e.g., 20–25 principal components).

Significance and Claims
The paper claims that the proposed data-informed parameter spaces effectively resolve the trade-off between expressiveness and the usability of the design space for global optimization. By mapping data distributions to a unit hypercube via quantile transformations, the methods provide natural bounds and scaling that facilitate the use of derivative-free algorithms like Bayesian optimization.

The primary scientific contribution is the demonstration that excellent alpha-particle confinement can be achieved in stellarator configurations that are far from quasisymmetric or quasi-isodynamic. This challenges the prevailing view that strict adherence to QS or QI conditions is necessary for good confinement, suggesting that these conditions may be overly constraining. The work provides a framework for directly optimizing high-fidelity, non-smooth physics objectives (like Monte Carlo guiding-center tracing) by combining dimensionality reduction, robust parameterization, and asynchronous parallel computing.