3DSTokesFlow: simulation-based inference for 3D Stokes profiles using flow matching

This paper introduces 3DSTokesFlow, a novel simulation-based inference framework using conditional flow matching to perform fast, scalable, and spatially correlated Bayesian inversion of 3D solar atmospheric parameters from Stokes profiles, enabling the generation of reliable posterior distributions and the computation of 3D physical quantities like electric currents and magnetic loop emergence.

The Gist

Imagine trying to figure out what's happening inside a storm cloud just by looking at the shadows it casts on the ground. That's essentially what solar physicists do when they look at the Sun. They can't stick a thermometer inside the Sun's atmosphere, so they have to guess the temperature, magnetic fields, and wind speeds by analyzing the light (specifically, the "Stokes profiles") that reaches us.

For decades, this has been like trying to solve a giant, blurry puzzle where many different pictures could fit the same set of clues. Traditional methods were slow, often gave just a single "best guess" without telling you how sure they were, and treated every tiny spot on the Sun as if it were alone, ignoring how its neighbors might be connected.

Enter 3DStokesFlow, a new tool developed by the authors that acts like a super-smart, fast-forwarding detective. Here is how it works, using simple analogies:

1. The Problem: The "Blindfolded Sculptor"

Imagine a sculptor trying to recreate a statue based only on a blurry photograph.

• The Old Way: The sculptor would look at one tiny part of the photo, guess the shape, then move to the next tiny part and guess again. They would ignore the fact that the arm connects to the shoulder. This was slow, and the final statue often had weird, disconnected parts.
• The New Way (3DStokesFlow): Instead of guessing one piece at a time, this new method looks at the entire photo at once. It understands that if the arm is here, the shoulder must be there. It uses the "spatial correlation"—the fact that neighboring pixels on the Sun are related—to make a much more accurate and consistent 3D model of the solar atmosphere.

2. The Secret Sauce: "Flow Matching"

The paper uses a technique called Flow Matching. Think of this like a river flowing from a calm lake into a chaotic, stormy ocean.

• The Lake (Simple Noise): The computer starts with a bag of pure, random static noise (like white noise on an old TV).
• The Ocean (The Real Sun): The goal is to transform that noise into a perfect, realistic map of the Sun's magnetic fields and temperatures.
• The River (The Flow): The AI learns the exact "current" or path needed to turn that random noise into the correct solar map. It doesn't just guess the final picture; it learns the journey from chaos to order. Because it learned this journey on millions of simulated solar storms, it can instantly reverse-engineer the path for real observations.

3. Training on "Fake" Sun Data

You can't train a pilot on a real plane crash, so they use simulators. Similarly, the authors trained 3DStokesFlow on 3D computer simulations of the Sun (specifically "quiet Sun" regions).

• They fed the AI millions of pairs: one side was the "fake" light coming from the simulation, and the other side was the "true" 3D map of what was inside that simulation.
• The AI learned the relationship between the light and the hidden physics. Once trained, it can look at real light from the Sun and instantly reconstruct the hidden 3D map.

4. What It Actually Found (The Results)

The paper claims this method does three specific, impressive things that previous tools struggled with:

• It sees in 3D (Geometric Height): Instead of just saying "this is deep" or "this is shallow" in an abstract way, it maps the Sun's atmosphere in actual geometric height (like kilometers above the surface). This is like turning a flat map into a real, 3D terrain model.
• It finds the "Electric Currents": Because it knows the 3D shape of the magnetic fields, it can calculate where electric currents are flowing. The paper shows that these currents are highly localized, appearing mostly at the boundaries where magnetic fields of different strengths meet. It's like finding the exact spots where electrical wires are fraying and sparking.
• It tracks "Emerging Loops": The authors applied this to real data from the Hinode satellite and watched a small magnetic loop rise through the Sun's surface. They could see the loop's "feet" (where it touches the surface) and its "head" (the top of the loop) moving and changing shape over time, confirming it was a single, connected structure rising up.

5. What It Does Not Do (Limitations)

The paper is very clear about what this tool is not yet:

• It's not for the whole Sun yet: It was trained only on "quiet" areas of the Sun, not the violent, stormy active regions (sunspots) where the physics is much more complex.
• It's not for other telescopes yet: It was trained specifically for the Hinode satellite's resolution. Using it on newer, sharper telescopes (like DKIST) would require retraining, which is computationally difficult.
• It doesn't measure sideways wind: It can tell you how fast the gas is moving up and down, but not how fast it's blowing sideways (though the authors hope to fix this later).

Summary

3DStokesFlow is a new, fast, and highly accurate way to turn 2D pictures of sunlight into 3D maps of the solar atmosphere. By using advanced AI to learn from simulations, it solves the "puzzle" of the Sun's magnetic fields and electric currents much better than before, revealing hidden structures like rising magnetic loops and localized electric currents that were previously impossible to see clearly.

Technical Summary

Technical Summary of 3DStokesFlow: Simulation-Based Inference for 3D Stokes Profiles Using Flow Matching

Problem Statement
The interpretation of observed Stokes profiles to infer solar atmospheric conditions is an ill-posed problem characterized by observational noise and mathematical degeneracies (e.g., the $180^\circ$ ambiguity in magnetic field azimuth). Traditional pixel-by-pixel (1D) inversion codes provide point estimates with unreliable uncertainty quantification, often failing to capture the complex posterior distributions required for rigorous probabilistic inference. Furthermore, existing machine-learning-based Bayesian frameworks are largely restricted to 1D spatial configurations, neglecting crucial spatial correlations between neighboring pixels. Inverting a full 2D field-of-view (FoV) with height stratification results in a posterior distribution of extremely high dimensionality (e.g., half a million dimensions for a $100 \times 100$ pixel map), making standard Bayesian inference computationally intractable.

Methodology
The authors introduce 3DStokesFlow, a novel framework based on conditional flow matching to perform fast, scalable Bayesian inference across an entire 2D FoV.

• Generative Modeling Strategy: The method treats the inversion as a generative problem, learning a continuous transformation (flow) that maps a simple base distribution (standard Gaussian) to the complex target posterior distribution of atmospheric parameters conditioned on observed Stokes profiles.
• Flow Matching: Instead of learning the flow directly, the model learns a time-dependent vector field $v_\theta(x)$ using a conditional loss function. By defining an optimal transport path between a noise sample $X_0$ and a target data sample $X_1$ (where $X_t = (1-t)X_0 + tX_1$), the model minimizes the difference between the learned vector field and the constant target vector field $(X_1 - X_0)$. This bypasses the need to know the true, intractable target vector field.
• Architecture:
  • Conditioning: A U-Net extracts multi-scale spatial features from the observed Stokes profiles (Fe I 630 nm line pair) to create a latent representation.
  • Velocity Model: A second U-Net takes the concatenation of the current state ($X_t$), the conditioning latent representation, and the time step $t$ to predict the vector field.
• Training Data: The model is trained on realistic 3D magnetohydrodynamic (MHD) simulations of the quiet Sun (from the SPIN4D project) generated with the MURaM code. The training set includes various magnetic configurations (zero field to mixed polarities) and covers a range of optical depths. The synthetic Stokes profiles are convolved with Hinode/SOT point spread functions and include Gaussian noise.
• Inference: Once trained, the model samples from the base distribution and integrates the learned vector field conditioned on the observed Stokes profiles to generate samples from the posterior distribution of atmospheric parameters.

Key Contributions and Results

• 3D Stratification and Uncertainty: Validation on independent synthetic datasets demonstrates that the model accurately captures the true 3D stratification of thermodynamic (temperature, pressure, velocity) and magnetic parameters. Crucially, it provides reliable posterior distributions (median and percentiles) rather than single point estimates.
• Geometrical Height Scale: Unlike traditional inversions that operate in optical depth, the model infers a geometrical height scale. This allows for the reconstruction of physical parameters as a function of geometric height, accounting for Wilson depressions and density variations.
• Disambiguation and Currents: By leveraging the inferred geometrical height and spatial correlations, the model enables a robust disambiguation of the transverse magnetic field components without requiring complex field extrapolation methods. This allows for the direct calculation of electric current density maps ($J$) and Lorentz forces.
• Dissipation Maps: The framework facilitates the computation of Ohmic and ambipolar dissipation maps in the photosphere directly from observations.
• Application to Real Data:
  • Quiet Sun: Applied to Hinode/SP quiet Sun observations, the model reveals highly localized electric currents at magnetic boundaries and produces dissipation maps consistent with theoretical expectations (Ohmic dissipation localized at magnetic borders; ambipolar dissipation in regions with measurable transverse fields).
  • Loop Emergence: The model successfully traces the 3D geometry of small-scale emerging magnetic loops, reconstructing their evolution from the photosphere up to $\sim300$ km, confirming the loop-like structure in a consistent geometrical height scale.

Significance and Claims
The paper claims that 3DStokesFlow represents a significant advancement by enabling fast and scalable Bayesian inference for full 2D FoVs while exploiting spatial correlations. The authors emphasize that this approach moves beyond the limitations of 1D pixel-by-pixel inversions and previous machine-learning attempts restricted to 1D atmospheres.

The primary significance lies in the ability to:

1. Provide reliable posterior distributions for atmospheric parameters, offering a rigorous quantification of uncertainties.
2. Derive geometrical height information directly from the inversion, which is typically a difficult constraint to satisfy.
3. Compute electric current densities and dissipation maps purely from observations without relying on additional assumptions like magnetohydrostatic equilibrium or field extrapolation, which were previously necessary for such calculations.

The authors note that the current implementation is limited to quiet Sun photospheric regions and disk-center observations of the Fe I 630 nm pair, as these are the most realistic and available MHD simulations. They acknowledge that extending the method to active regions, higher atmospheric layers, and higher spatial resolutions (e.g., DKIST) requires further development, potentially involving latent space modeling or transformer-based encoders. However, the current work establishes a proof-of-concept for simulation-based inference that yields physically consistent 3D maps of the solar atmosphere.