🔭 astrophysics

Accelerating 3D Non-LTE Synthesis with Graph Neural Networks

This paper introduces a Graph Neural Network framework that accelerates 3D non-LTE radiative transfer calculations by orders of magnitude while maintaining high accuracy, thereby enabling efficient and routine spectropolarimetric inversions of the solar chromosphere with full horizontal radiative coupling.

Original authors: A. Vicente Arévalo, A. Asensio Ramos, C. J. Díaz Baso

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: A. Vicente Arévalo, A. Asensio Ramos, C. J. Díaz Baso

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Isolated Column" Mistake

Imagine you are trying to understand the weather in a specific city. In the past, solar physicists modeled the Sun's atmosphere (the chromosphere) by treating every single vertical column of air as if it were an isolated island. They would look at one column, calculate the temperature and light, and move to the next, ignoring the fact that wind and clouds blow sideways between them.

This is called the 1.5D approximation. It's like trying to predict the weather in New York by only looking at the air directly above a single street corner, completely ignoring the wind blowing in from New Jersey or the humidity coming from the ocean.

In the Sun's atmosphere, this is a big problem. Light (photons) travels huge distances horizontally before being absorbed. The state of the gas in one spot depends heavily on what is happening in the spots next to it. To get the physics right, scientists need to solve the problem in full 3D, accounting for all that sideways interaction.

The Bottleneck: The "Slow Calculator"

The only way to get the perfect 3D answer right now is to use super-complex computer codes (like Multi3D). These codes work like a very slow, stubborn calculator. They guess the answer, check the math, realize they were wrong, guess again, check again, and repeat this process millions of times until the numbers finally settle down.

This process is so slow that it can take millions of CPU hours to simulate just a tiny snapshot of the Sun. Because it's so slow, scientists can't use it to "invert" observations (work backward from what we see to figure out what the Sun is actually doing). They are forced to stick with the inaccurate "isolated column" method.

The Solution: The "Graph Neural Network" (GNN)

The authors of this paper built a new tool using Graph Neural Networks (GNNs). Think of this not as a calculator, but as a highly trained intuition engine.

Instead of solving the math step-by-step, they taught a computer to "see" the physics.

The Map: They turned the 3D solar atmosphere into a giant graph. Imagine a 3D spiderweb where every knot (node) is a piece of the Sun's atmosphere holding data like temperature, magnetic fields, and density. The strings (edges) connecting the knots represent the physical distance between them.
The Training: They fed this network data from a perfect, slow simulation (the "ground truth"). The network learned the rules of how light travels sideways and how atoms behave in different conditions. It learned that if a knot has a certain temperature and is close to a knot with a strong magnetic field, the atoms in the first knot will behave in a specific way.
The Result: Once trained, the network doesn't "calculate" the answer iteratively. It just predicts it instantly based on the patterns it learned.

How It Works: The "Star" Shape

To make this fast, they didn't connect every single knot to every other knot (which would be too heavy for a computer). Instead, they used a "star-like" connectivity.

Imagine you are the center of a star. You only need to listen to the people standing in a circle around you to understand the neighborhood. You don't need to listen to people on the other side of the planet. The network focuses on the "central column" it is trying to predict and gathers information from the surrounding "context" nodes. This keeps the math light and fast.

The Results: Fast and Accurate

The team tested this new "intuition engine" on a simulation of the Sun's atmosphere.

Speed: It is roughly one million times faster than the traditional slow calculators. What used to take weeks or months now takes minutes.
Accuracy: In the lower and middle parts of the atmosphere, the predictions are almost perfect (over 99% correlation).
The Upper Atmosphere: In the very top layers, where the physics is messier and light travels even further, the errors get slightly bigger, but they remain unbiased (they don't systematically lie in one direction).
Spectral Synthesis: When they used these predicted numbers to generate images of the Sun's light (specifically the Calcium 8542 Å line), the resulting images looked almost identical to the slow, perfect 3D simulations.

Why This Matters

This paper doesn't claim to solve the Sun's mysteries today, but it removes the biggest roadblock. By creating a "fast forward model," they have paved the way for future software that can finally do 3D inversions.

In the future, scientists will be able to take real observations from telescopes and use this fast network to instantly figure out the 3D structure of the Sun's magnetic fields and temperature, without having to wait years for the computer to crunch the numbers. It turns a "super-complex, slow puzzle" into a "quick, accurate prediction."

Technical Summary: Accelerating 3D Non-LTE Synthesis with Graph Neural Networks

Problem Statement
Accurate modeling of solar chromospheric spectra requires solving the radiative transfer (RT) problem under non-local thermodynamic equilibrium (non-LTE) conditions. In these tenuous layers, atomic level populations are strongly coupled to the radiation field across vast spatial volumes due to non-local scattering. Traditional 3D non-LTE codes (e.g., Multi3D, PHOENIX/3D, PORTA) solve this via computationally expensive iterative schemes, such as Multi-level Accelerated Lambda Iteration (MALI), often requiring millions of CPU hours per snapshot. This computational bottleneck prevents the use of full 3D physics in spectropolarimetric inversions. Current inversion codes (e.g., NICOLE, STiC) are forced to rely on 1.5D approximations, treating each pixel as an independent column and neglecting horizontal radiative transfer. This assumption fails in dynamic chromospheric regions with strong horizontal gradients (e.g., fibrils, shock fronts), introducing biases in inferred temperature and magnetic field stratification and failing to reproduce key features like Hα fibrils.

Methodology
The authors propose a Graph Neural Network (GNN) framework to serve as a fast surrogate model for solving the 3D statistical equilibrium (SE) equations. The approach extends previous 1.5D work to the full 3D domain for the Ca II atom.

Graph Construction: The solar atmosphere is discretized as a directed graph rather than a regular Cartesian grid.
- Nodes: Encode local physical properties at specific spatial points, including temperature ( $T$ ), velocity ( $v$ ), magnetic field ( $B$ ), electron density ( $n_e$ ), gas pressure ( $p_{gas}$ ), mass density ( $\rho$ ), and geometric height ( $z$ ).
- Edges: Encode geometric Euclidean distances between spatial points. Unlike the 1.5D approach which used optical depth distance, the 3D model uses geometric distance, allowing the network to learn opacity structures from the combination of distance and local densities.
- Topology: To manage memory constraints, a "star-like" connectivity scheme is employed. The graph focuses on inferring populations for a central column, connecting it to surrounding context nodes within a defined radius. This avoids the $O(n^2)$ memory cost of a fully connected graph while preserving direct radiative coupling between the target and its 3D environment.
Architecture: The model utilizes an Encode-Process-Decode architecture:
- Encoder: Projects input features and edge attributes into a 128-dimensional latent space using Multi-Layer Perceptrons (MLPs).
- Process: Information propagates across the domain through 16 message-passing steps. Edge and node states are updated sequentially, mimicking the physical propagation of the radiation field. This depth is significantly reduced compared to 1.5D models due to the richer 3D connectivity.
- Decoder: Projects the final latent states back to physical space to predict normalized populations for the five bound levels of Ca II and the Ca III ground state.
Training: The network is trained on a snapshot from a Bifrost radiation-MHD simulation (24 $\times$ 24 $\times$ 14.4 Mm $^3$ ). Ground truth populations are computed using Multi3D. The training strategy involves constructing graphs from subvolumes with a dynamic striding technique (randomly selecting strides $S \in \{1, \dots, 5\}$ ) to ensure robustness to different spatial scales. The loss function is a standard Mean Squared Error (MSE) on transformed (logarithmic) population values.

Key Results

Population Prediction: The trained GNN accurately predicts Ca II level populations. Correlations exceed 0.99 in the photosphere and mid-chromosphere. Errors increase in the upper chromosphere where non-local effects dominate, but the errors remain unbiased and follow Gaussian distributions.
Speedup: Inference is approximately $10^6$ times faster than traditional iterative SE solvers. While a full 3D NLTE solution may take millions of CPU hours, GNN inference takes minutes.
Spectral Synthesis: Synthesis of the Ca II 8542 Å line using the predicted populations yields intensity profiles with mean residuals of $\lesssim 2\%$ relative to the full 3D solution. The synthesized maps and mean profiles are nearly indistinguishable from the ground truth (Multi3D) and significantly better than 1.5D or LTE approximations.
Error Distribution: The majority of inference errors are concentrated in the upper chromosphere (where populations are low and non-local effects are strong) and lower layers for Ca III. However, the relative errors in the normalized space remain Gaussian across all atmospheric regimes.

Significance and Claims
The paper claims that this 3D GNN framework successfully bypasses the computational bottleneck of iterative SE solvers while preserving essential non-LTE physics, including horizontal radiative transfer. The primary significance lies in enabling routine 3D non-LTE inversions. By acting as a fast, differentiable forward model, the GNN allows inversion codes to retrieve 3D atmospheric properties without relying on the restrictive 1.5D approximation.

The authors note that while the model is robust, it is currently trained on specific simulation conditions (Bifrost "enhanced network") and may require training on more diverse datasets (e.g., flaring regions, quiet Sun extremes) for broad application. They also identify future work directions, including extending the architecture to predict off-diagonal density matrix elements for full-Stokes polarized synthesis (including Zeeman splitting and Hanle effect) and handling partial frequency redistribution. The method is presented as a foundational step toward integrating full 3D non-LTE physics into standard solar inversion workflows.