🔭 astrophysics

Predicting the thermodynamics in the chromosphere from the translation of SDO data into the IRIS $^{2}$ inversion results using a visual transformer model

The paper introduces SDO2IRIS $^2$ , a visual transformer model that translates SDO/AIA and SDO/HMI observations into chromospheric thermodynamic parameters (temperature, electron density, and velocities) by learning from IRIS Mg II h&k inversions, achieving strong predictive correlations for temperature and electron density with rapid execution times suitable for standalone or complementary solar physics analysis.

Original authors: Alberto Sainz Dalda, Vishal Upendran, Juno Kim, Kyuhyoun Cho, Paul S. Killam, Viggo Hansteen, Bart De Pontieu

Published 2026-04-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Alberto Sainz Dalda, Vishal Upendran, Juno Kim, Kyuhyoun Cho, Paul S. Killam, Viggo Hansteen, Bart De Pontieu

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 Picture: A Solar "Translator"

Imagine the Sun is a massive, complex machine. Scientists have two main cameras watching it:

SDO (Solar Dynamics Observatory): This is like a wide-angle security camera. It sees the whole Sun all the time, in high definition, but it mostly sees the "surface" and the outer atmosphere. It's great for seeing the big picture, but it can't see deep inside the Sun's lower atmosphere (the chromosphere) very well.
IRIS (Interface Region Imaging Spectrograph): This is like a high-powered microscope. It zooms in on tiny, specific spots to see the nitty-gritty details of the gas and heat deep inside the Sun's atmosphere. However, it can only look at a tiny slice of the Sun at a time, and it's slow.

The Problem: We have a lot of data from the "wide-angle" camera (SDO), but we don't know the deep physics (temperature, speed of gas, density) for most of it because the "microscope" (IRIS) isn't looking there.

The Solution: The authors built a super-smart AI translator called SDO2IRIS2. It learns how to look at the "wide-angle" photos and guess what the "microscope" would have seen if it were looking at that same spot.

How the AI Works: The "Visual Transformer"

Think of the AI model as a master chef who has tasted thousands of dishes (IRIS data) and seen the ingredients used to make them (SDO photos).

The Training: The AI was fed 1,500 examples where it saw both the SDO photo and the corresponding IRIS "microscope" data. It learned the recipe: "When the SDO photo looks like this (bright here, dark there, with this magnetic field), the temperature inside is usually this hot, and the gas is moving this fast."
The Magic Trick (Visual Transformer): Instead of just looking at one pixel at a time, this AI uses a Visual Transformer. Imagine looking at a puzzle. A normal AI might look at one piece and guess the picture. A Visual Transformer looks at the whole puzzle at once, understanding how the pieces relate to each other globally and locally. It understands the "context" of the solar storm, not just the pixels.
The Output: Once trained, you can feed it any SDO photo (even from 2010, before IRIS was even working), and it instantly predicts the temperature, gas speed, and density of the chromosphere for that entire image.

What Did They Predict?

The AI predicts four things about the Sun's atmosphere:

Temperature ( $T$ ): How hot is it?
Electron Density ( $n_e$ ): How crowded is the gas?
Turbulent Speed ( $v_{turb}$ ): How chaotic is the gas movement?
Line-of-Sight Velocity ( $v_{los}$ ): Is the gas moving toward us or away from us?

The Results: A Mixed Bag of Success

The paper tested the AI on data it had never seen before. Here is how it did:

Temperature & Density (The Stars): The AI is excellent at guessing temperature and density. It got the right answer about 80% of the time with very high accuracy. It's like a weather forecaster who can predict the temperature almost perfectly just by looking at a satellite cloud map.
Turbulence (The Good): It did a pretty good job guessing how chaotic the gas is (about 70% accuracy).
Movement Speed (The Struggle): It struggled to guess exactly how fast the gas is moving toward or away from us. The correlation was weak.
- Analogy: Imagine trying to guess how fast a car is driving toward you just by looking at a photo of the road. You can guess the road is there, but the exact speed is hard to tell without seeing the car move. The AI sees the "road" (the magnetic fields and light) but can't perfectly calculate the "speed" (Doppler shift) from a still image.

Why Does This Matter?

1. Unlocking Decades of Data:
SDO has been taking photos of the whole Sun every 12 seconds since 2010. That's a lot of data. Before this tool, we couldn't know the deep physics of most of those photos. Now, we can apply SDO2IRIS2 to every single SDO image ever taken and get a full 3D map of the Sun's atmosphere.

2. Speed:
Calculating these physics values the old way (using complex physics equations) takes hours or days. The AI does it in minutes (or even seconds on a powerful computer).

3. Filling the Gaps:
Sometimes IRIS isn't looking at a specific solar storm. With this tool, we can use SDO data to create a "proxy" (a stand-in) for what the physics would be, giving us context for events we couldn't study in detail otherwise.

The Catch (Limitations)

The Edge of the Sun: The AI was trained on data from the center of the Sun. If you look at the very edge (the limb), the AI gets confused, just like a map that works great in the city center but fails at the border.
The "Ghost" Lines: Because the training data had little marks (fiducial lines) to help align the telescope, the AI sometimes predicts these marks even when they aren't there. It's like a chef who always puts a garnish on a dish because the training photos had garnish, even if the customer didn't order it.
Resolution: The AI is limited by the resolution of the SDO camera. It can't see tiny details that the IRIS microscope sees, so some fine structures get blurred out.

The Bottom Line

SDO2IRIS2 is a game-changer. It turns the SDO satellite from a simple "photo taker" into a "physics simulator." It allows scientists to instantly understand the heat, density, and chaos of the Sun's atmosphere across the entire solar disk, using a smart AI that learned the secrets from the high-resolution IRIS telescope.

Note: The paper is dedicated to the memory of Alan Title, a pioneer who led the missions that made this research possible.

1. Problem Statement

The solar chromosphere is a complex layer of the solar atmosphere characterized by rapid spatial and temporal variations in thermodynamic properties and magnetic fields, operating under non-local thermodynamic equilibrium (non-LTE) conditions. Deriving physical parameters (temperature, velocity, density) from observations is challenging:

Traditional Inversion: Methods like the IRIS2 inversion tool, which analyzes Mg II h&k spectral lines observed by the Interface Region Imaging Spectrograph (IRIS), provide accurate thermodynamic maps but are computationally expensive and time-consuming.
Data Availability: The Solar Dynamics Observatory (SDO) has provided full-disk, high-cadence imaging (AIA) and magnetograms (HMI) since 2010, but these lack the direct spectral resolution to derive chromospheric thermodynamics without IRIS observations.
The Gap: There is a need for a fast, efficient method to estimate chromospheric thermodynamics using the widely available SDO data, leveraging the high-fidelity results from IRIS2 as a ground truth.

2. Methodology

The authors developed SDO2IRIS2, a deep learning tool based on a Visual Transformer (ViT) architecture designed to translate SDO observations into IRIS2-derived thermodynamic parameters.

Data Preparation

Training Data: 1,585 IRIS Mg II h&k datasets were inverted using IRIS2 to generate target maps for four variables: Temperature ( $T$ ), Line-of-Sight Velocity ( $v_{los}$ ), Turbulent Velocity ( $v_{turb}$ ), and Electron Density ( $n_e$ ) at three optical depths ( $\log(\tau) = -5.2, -4.2, -3.2$ ).
Input Data: SDO/AIA intensity images (channels 171, 193, 304, 1600, 1700 Å) and SDO/HMI photospheric magnetograms were co-aligned and resampled to match the IRIS slit geometry ("slit-like rasters").
Preprocessing: Data were standardized, zero-padded to ensure divisibility by patch sizes, and split sequentially (80% training, 10% validation, 10% testing) to prevent temporal leakage (consecutive rasters in different sets).

Model Architecture

Base Model: A Visual Transformer (ViT) acting as an Image-to-Image translator.
Input: A tensor of size $C_{in} \times H \times W$ , where $C_{in}$ includes AIA channels, HMI magnetogram, and sinusoidal position encodings (totaling 30 input channels).
Mechanism: The input image is split into non-overlapping patches, linearly projected into tokens, and processed through 8 layers of Transformer encoders (8 heads, 512 feed-forward size) to capture both local and global spatial contexts. A decoder then maps these embeddings to the 12 output channels (4 variables $\times$ 3 optical depths).
Loss Function: A composite loss function was used:
$L_{MS} = L_1(y, \hat{y}) + \alpha \frac{1}{K} \sum_{k} L_1(\vec{\nabla}A_k[y], \vec{\nabla}A_k[\hat{y}])$
This combines standard pixel-wise $L_1$ loss with a spatial gradient loss (computed over binned scales $k=1, 2, 4$ ) to ensure spatial coherence between prediction and target.
Baseline: A Deep Neural Network (DNN) without spatial context (pixel-to-pixel) was used as a baseline for comparison.

3. Key Contributions

SDO2IRIS2 Tool: The first application of a Visual Transformer to translate full-disk SDO/AIA and HMI data into chromospheric thermodynamic parameters typically derived from IRIS spectral inversions.
Feature Attribution: Utilization of Shapley values (KernelSHAP) to quantify the contribution of each input channel (AIA bands and HMI magnetogram) to the model's performance, revealing that lower atmosphere channels (1600, 1700 Å) and the transition region channel (304 Å) are the primary drivers of accuracy.
Generalization: The model is designed to handle variable spatial dimensions via a "rolling-patch" approach, allowing it to predict thermodynamics for arbitrary fields of view, not just the specific IRIS slit locations used in training.
Efficiency: The model reduces computation time from hours (for full IRIS2 inversions) to minutes (≤10 mins on CPU, ≤5 mins on GPU).

4. Results

The model was evaluated on a test set of 185 datasets using Pearson correlation coefficients ( $r$ ) between SDO2IRIS2 predictions and IRIS2 ground truth.

Performance Metrics (after outlier removal):
- Temperature ( $T$ ) & Electron Density ( $n_e$ ): Showed strong correlations ( $r \approx 0.80$ ) for approximately 80% of the data in the high and mid-chromosphere ( $\log(\tau) = -5.2, -4.2$ ).
- Turbulent Velocity ( $v_{turb}$ ): Showed moderate-to-strong correlations ( $r \approx 0.63$ ) for approximately 70% of the data.
- Line-of-Sight Velocity ( $v_{los}$ ): Showed weak correlations across all experiments. The model tends to underestimate the magnitude of velocities compared to IRIS2.
Feature Importance:
- The most critical inputs were AIA 1600 Å, 1700 Å, and 304 Å.
- Adding the HMI magnetogram provided a slight improvement and helped represent magnetic structures better but did not drastically change correlation coefficients.
- Including all AIA channels (including coronal 193/171 Å) offered marginal gains for $T$ and $n_e$ but slightly degraded performance for $v_{turb}$ in the temperature minimum region.
Visual Analysis:
- The model successfully reproduces large-scale structures (sunspots, plage, light bridges).
- Limitations: The model struggles with fine-scale internetwork features due to the lower spatial resolution of SDO/AIA compared to IRIS. It also reproduces "fiducial marks" (calibration lines) present in the training data, which can appear as artifacts in predictions.
- SAA Handling: The model successfully predicts thermodynamics in regions where IRIS data is corrupted by the South Atlantic Anomaly (SAA), as SDO/AIA is unaffected, demonstrating robustness.

5. Significance and Future Work

New Usability for SDO: SDO2IRIS2 unlocks the ability to derive chromospheric thermodynamics for the entire SDO archive (2010–present), even for periods when IRIS was not observing. This creates a continuous, global map of chromospheric conditions.
Synergy: It serves as a powerful complementary tool for simultaneous observations, providing context for other instruments or filling gaps when spectral data is unavailable.
Future Evolution: The authors propose extending the tool to AIA2IRIS2+, incorporating inversions from additional spectral lines (C II, photospheric lines) to cover a broader atmospheric range from the photosphere to the high chromosphere.

In conclusion, SDO2IRIS2 represents a significant advancement in solar physics, utilizing state-of-the-art Vision Transformers to bridge the gap between high-cadence imaging and detailed thermodynamic modeling, offering a fast, scalable solution for analyzing the solar chromosphere.