🔭 astrophysics

Improving Solar Flare Soft X-ray Classification With FOXES: A Framework For Operational X-ray Emission Synthesis

This paper introduces FOXES, a Vision Transformer-based framework that synthesizes spatially-resolved solar soft X-ray irradiance from Extreme Ultraviolet observations to overcome the location and viewpoint limitations of current GOES satellite data, thereby enabling more accurate flare classification and multiviewpoint space weather monitoring.

Original authors: Griffin T. Goodwin, Alison J. March, Jayant Biradar, Christoph Schirninger, Robert Jarolim, Angelos Vourlidas, Viacheslav M. Sadykov, Lorien Pratt

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Griffin T. Goodwin, Alison J. March, Jayant Biradar, Christoph Schirninger, Robert Jarolim, Angelos Vourlidas, Viacheslav M. Sadykov, Lorien Pratt

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 "Blind" Satellite

Imagine the Sun is a giant, glowing stage where solar storms (flares) happen. To keep Earth safe, we need to know exactly how strong these storms are.

Currently, we rely on a satellite called GOES that sits in Earth's orbit. Think of GOES as a very sensitive microphone placed far away from the stage. When a solar flare happens, the microphone picks up the "volume" (the X-ray energy) of the scream. It can tell us how loud the scream is (Class C, M, or X), but because it's just a single microphone, it cannot tell us where on the stage the actor is screaming.

This causes two big problems:

Confusion: If two actors scream at the same time, the microphone just hears one big noise. We don't know which actor did it or if they are helping each other.
Blind Spots: If an actor is on the far side of the stage (the back of the Sun), the microphone might not hear them clearly, or we might not know they are screaming at all.

The Solution: The "Super-Eyes" (FOXES)

The authors of this paper built a new tool called FOXES (Framework for Operational X-ray Emission Synthesis).

Instead of just listening with a microphone, FOXES uses high-definition cameras (EUV images from the SDO satellite) that can see the entire stage in detail. It uses a type of Artificial Intelligence called a Vision Transformer (think of it as a super-smart detective that looks at patterns in a photo).

Here is the magic trick:
FOXES learns to look at the detailed pictures of the Sun and say, "Based on what I see in these photos, I can predict exactly how loud the 'microphone' (GOES) would hear it, and exactly where on the stage the sound is coming from."

It essentially creates a "Virtual Microphone" that can be placed anywhere, even on the back of the Sun.

How It Works (The Analogy)

Imagine you are trying to guess the total weight of a pile of bricks just by looking at a photo of them.

The Old Way (Baseline): You just count the total number of pixels in the photo and multiply by a number. It's a rough guess. You might be right for small piles, but you'd be way off for huge, dense piles.
The FOXES Way: The AI looks at the photo, identifies every single brick, sees how they are stacked, notices which ones are glowing hot, and calculates the weight of each individual brick. Then, it adds them all up.
- Result: It gives you the total weight (the GOES measurement) AND a map showing exactly which bricks contributed the most weight.

What Did They Find?

It's Surprisingly Accurate: When they tested FOXES, it was incredibly good at guessing the "loudness" of the flares. It was much more accurate than the old "rough guess" method.
It Solves the "Two Actors" Problem: In one test, two flares happened at the same time. The old catalogs got confused and only reported one big event. FOXES looked at the photo, saw two distinct glowing spots, and correctly said, "Hey, there are actually two separate flares happening here!"
The "Back of the Sun" Limitation: The AI is great, but it has one weakness. If a flare happens right on the very edge of the Sun (the limb), the AI sometimes underestimates the strength. It's like trying to hear someone whispering from behind a wall; the sound gets muffled. The authors plan to fix this in future versions.

Why Does This Matter?

This isn't just about better math; it's about safety and exploration.

Space Weather: If a massive storm hits the back of the Sun, we usually don't know until it's too late. FOXES could act as a "virtual GOES" for other satellites (like STEREO or Solar Orbiter) that are looking at the Sun from different angles. This means we could get a 360-degree view of solar storms.
Mars Missions: If humans go to Mars, they need to know about solar storms. Mars doesn't have a GOES satellite. But if we have a satellite with a camera (EUV) orbiting Mars, FOXES could translate those photos into "flare strength" data, warning astronauts of danger even when they are on the other side of the solar system.

The Bottom Line

The authors built an AI that turns pictures of the Sun into sound measurements of solar storms. It's like giving us a pair of glasses that lets us see the invisible energy of the Sun, helping us predict space weather more accurately and keep our technology (and future astronauts) safe.

1. Problem Statement

Solar flares are critical drivers of space weather, capable of disrupting satellite operations, communications, and power grids. The standard method for classifying flare strength relies on the Geostationary Operational Environmental Satellite (GOES) series, which measures spatially integrated Soft X-ray (SXR) irradiance in the 1–8 Å wavelength range.

However, this approach has significant limitations:

Lack of Spatial Resolution: GOES provides a single flux value for the entire solar disk. It cannot distinguish between multiple simultaneous flares or accurately locate the source, leading to misattributed flare locations and ambiguous classifications during high-activity periods.
Single Viewpoint: GOES observes only from Earth's orbit (Sun-Earth line). Missions observing from other angles (e.g., STEREO, Solar Orbiter) lack SXR imagers, making it difficult to characterize flares occurring on the far side of the Sun or at the limb.
Data Quality: Existing flare catalogs often suffer from temporal gaps and inaccuracies in location/strength due to the inability to disentangle concurrent events.

The core research question is: Can spatially resolved Extreme Ultraviolet (EUV) observations be used to infer and synthesize the integrated SXR flux measured by GOES?

2. Methodology: The FOXES Framework

The authors introduce FOXES (Framework for Operational X-ray Emission Synthesis), a machine learning model designed to translate EUV imagery into SXR flux predictions.

Data Collection and Preprocessing

Input Data: Seven EUV channels from the Solar Dynamics Observatory (SDO)/Atmospheric Imaging Assembly (AIA): 94, 131, 171, 193, 211, 304, and 335 Å. These channels are sensitive to hot flaring plasma.
Target Data: 1-minute averaged 1–8 Å SXR flux from GOES (Release Version 2-2-0).
Dataset: Approximately 3,200 hours of observations (1.4 TB) spanning January 2012 to September 2025.
- Preprocessing: AIA images were cropped to 1.1 solar radii, corrected for instrument degradation, normalized, and resampled to $512 \times 512$ pixels. GOES data was scaled to remove operational factors.
- Splitting: Data was temporally split to prevent leakage: Training (Jan 2012–Dec 2022, Jul 1–20 2023), Validation (Jan–Jun 2023, Jul 22–30 2023), and Testing (Aug 2023–Sep 2025).

Model Architecture: Vision Transformer (ViT)

FOXES utilizes a Vision Transformer architecture, chosen for its ability to capture global context and provide interpretable attention maps.

Input Processing: The input is a stack of 7 channels ( $7 \times 512 \times 512$ ). The image is segmented into $8 \times 8$ pixel patches, resulting in 4,096 patches per image.
Embedding: Each patch is flattened into a 448-dimensional vector ( $7 \text{ channels} \times 8 \times 8$ ) and projected into a 256-dimensional embedding space.
Transformer Encoder: The model employs 8 stacked encoder blocks containing:
- Layer normalization.
- Multi-head self-attention: Specifically, a masked-attention mechanism is used where a patch can only attend to patches outside a $9 \times 9$ local grid. This forces the model to learn global context (background state) while preserving local brightness via residual connections.
- A 1024-dimensional MLP.
Output Head: A regression head predicts the SXR flux contribution for each patch.
- Output 1: The sum of all patch fluxes (Global SXR prediction).
- Output 2: A spatially resolved flux map (identifying source regions).
Training: Trained for 100 epochs on two NVIDIA A100 GPUs using Huber Loss ( $\delta=0.3$ ) in log-space, with class-weighted sampling to address the imbalance between quiet Sun and X-class flares.

Baseline Comparison

A simple Linear Regression model was developed as a baseline. It used the disk-integrated signal of the 7 EUV channels (a $7 \times 1$ vector) to predict GOES flux, lacking any spatial information.

3. Key Contributions

End-to-End Translation: FOXES is the first deep learning model to map full-disk, multi-channel EUV imagery directly to integrated SXR flux and spatial distributions.
Virtual Spatially-Resolved GOES: By distributing the predicted flux across the solar disk, FOXES effectively creates a "virtual" GOES X-ray Sensor (XRS) on board SDO, capable of localizing flare sources.
Multiviewpoint Capability: The framework offers a pathway to synthesize SXR data for missions outside the Sun-Earth line (e.g., STEREO, Solar Orbiter) that lack SXR imagers but possess EUV instruments.
Disentanglement of Sympathetic Events: The model can separate contributions from multiple active regions occurring simultaneously, a task where current catalogs often fail.

4. Results

The model was evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Bias Error (MBE), and Pearson correlation ( $r$ ) in log-space.

Performance vs. Baseline:
- Baseline: MAE of 0.307 dex (predictions off by a factor of ~2.0). It struggled with low and high fluxes, often biased toward C-class predictions.
- FOXES: MAE of 0.051 dex (predictions off by a factor of ~1.12).
- Improvement: FOXES reduced the MAE by 83% and RMSE by 79% compared to the baseline. The Pearson correlation improved from 0.737 to 0.990.
Flare Class Performance: FOXES performed exceptionally well across C and M classes. It showed slight underprediction for X-class flares (likely due to AIA filter saturation and insensitivity to plasmas >30 MK) and limb flares (due to optical thickness and cropping).
Spatial Localization:
- The model successfully identified source regions for individual flares.
- In a case study of simultaneous flares, FOXES correctly identified two distinct events and their relative intensities, whereas standard catalogs often attributed the flux to a single region or missed the secondary event.
Ablation Study (Channel Importance):
- Channels peaking at high temperatures (131, 193, 211 Å) were most critical.
- Removing the 94, 131, and 335 Å channels simultaneously caused a significant performance drop (MAE ~0.49 dex), suggesting these are vital for high-energy flare detection.

5. Significance and Future Implications

Enhanced Space Weather Forecasting: By providing accurate, spatially resolved flare data from off-Sun-Earth-line vantage points, FOXES enables forecasting for the entire heliosphere, crucial for protecting assets and future human missions (e.g., to Mars).
Improved Flare Catalogs: FOXES can generate a more comprehensive and accurate catalog of flare events, resolving ambiguities in current databases regarding location and strength during complex events.
Operational Utility: The model acts as a "virtual instrument," fusing the capabilities of SDO/AIA and GOES/XRS. It allows researchers to "calibrate" GOES curves using spatial context.
Future Work: The authors plan to retrain the model on data from STEREO and Solar Orbiter (which have fewer EUV channels) to create a robust "virtual GOES" for the entire solar sphere. Further investigation is needed to handle limb flares and optically thick emissions.

In conclusion, FOXES represents a significant leap forward in solar physics, moving from purely statistical flare classification to a spatially aware, AI-driven synthesis of solar activity that bridges the gap between EUV imaging and X-ray flux measurement.