A Convolutional Neural Network-Derived Catalog of Solar Flares from Soft X-Ray Observations

This paper presents a new, highly sensitive catalog of solar flares derived from GOES soft X-ray data using a convolutional neural network, which identifies over 111,000 candidates and reveals a steeper power-law distribution for small events while correcting for previous under-counting biases in flare statistics.

The Gist

The Big Picture: Finding the "Whispers" in a Roar

Imagine the Sun is a giant, noisy stadium. For decades, scientists have been trying to count the fans cheering (solar flares). But they've been using a very old, slightly broken microphone (the standard GOES catalog) that only picks up the loudest screams. If a fan whispers or if two people cheer at the exact same time, the microphone misses them or thinks it's just background noise.

This paper introduces a new, super-smart "AI listener" (a Convolutional Neural Network, or CNN) that can hear the whispers and separate overlapping cheers. By using this new listener, the scientists found more than seven times as many solar flares as the old catalog did.

The Problem: The "Obfuscation" Effect

The old way of counting flares had a major flaw called obscuration.

• The Analogy: Imagine you are trying to count individual raindrops falling on a roof. But right now, it's a heavy thunderstorm. The roar of the heavy rain drowns out the sound of the smaller drops. You only count the big splashes.
• The Science: When the Sun is very active (solar maximum), the background "noise" of X-rays is so loud that the standard algorithms can't see the smaller flares. They also struggle when two flares happen close together, often merging them into one big event or missing the second one entirely. This led scientists to believe that big flares were rare and that small flares didn't happen often.

The Solution: The AI Detective

The authors built a Convolutional Neural Network (CNN). Think of this AI not as a simple calculator, but as a detective who has studied thousands of hours of solar "crime scenes" (data).

1. Training the Detective: Instead of teaching the AI to look for a whole storm (start-to-end), they taught it to recognize the specific moment a storm starts (the "rise episode"). It's easier to spot the exact moment a wave begins to rise than to guess exactly when it crashes back down.
2. The "Human" Check: To teach the AI, the scientists manually looked at 145 random days of data and marked the flares themselves. This created a "gold standard" reference book for the AI to learn from.
3. The Result: The AI scanned the entire Sun from 2018 to 2025. It found 111,580 flare candidates. The old catalog only found 14,612.

How They Knew It Was Real (The Bayesian Filter)

Since the AI found so many more events, the scientists had to make sure it wasn't just hallucinating noise. They used a Bayesian Inference system.

• The Analogy: Imagine the AI spots a shadow and says, "That's a person!" To be sure, it asks three questions:
  1. How big is the shadow? (Peak Flux)
  2. How fast did it appear? (Rise Time)
  3. Does anyone else see it? (Coincidence with other catalogs)
• If the shadow is tiny and fast, the AI says, "Probably just a trick of the light." If it's big and matches what other telescopes saw, the AI says, "Definitely a person." They kept the events where the AI was at least 20% confident, which still gave them a massive new list of flares.

What They Learned: The Sun is a Random Party

With this new, complete list of flares, the scientists re-examined two big questions:

1. Do small flares follow a specific pattern?
Yes. They found that the sizes of flares follow a "Power Law."

• The Analogy: Think of a forest fire. There are many tiny sparks, fewer small fires, fewer medium fires, and very few massive infernos. The new catalog showed that this pattern holds true even for the tiniest sparks, extending the pattern down to sizes the old catalog couldn't see.

2. Do big flares "pause" the Sun?
For years, scientists debated: If a huge flare happens, does the Sun need a "recovery time" before the next one? (Like a person needing a nap after running a marathon).

• The Old View: The data suggested that after a big flare, there was a long wait for the next one.
• The New View: The AI showed that this "long wait" was an illusion caused by the old microphone missing the small flares that happened during the recovery. When you count all the flares, the Sun doesn't seem to pause. It behaves like a random process (like popcorn popping). A big pop doesn't stop the next kernel from popping; it just happens to be a bigger one.

The Conclusion

This paper is a game-changer because it fixes the "blind spots" in our view of the Sun.

• Before: We thought the Sun was quiet most of the time, with rare, massive explosions.
• Now: We know the Sun is constantly "flickering" with thousands of tiny events we couldn't see before.

The new catalog is like upgrading from a grainy black-and-white TV to a 4K Ultra HD stream. It reveals that the Sun's energy release is a continuous, random, and much more active process than we ever imagined. This helps scientists better predict space weather, which can affect our satellites, power grids, and GPS on Earth.

Technical Summary

1. Problem Statement

Solar flare statistics are fundamental to understanding energy storage and release mechanisms in the solar atmosphere. However, existing studies often yield conflicting results regarding flare size distributions and waiting times. The primary cause of this discrepancy is obscuration: the inability of traditional flare detection algorithms (FDAs) to detect weak or temporally overlapping flares during periods of high solar activity.

• Limitations of Current Catalogs: The standard NOAA GOES catalog relies on 1-minute averaged Soft X-Ray (SXR) data and arbitrary rules (e.g., requiring four consecutive minutes above a threshold). This approach suppresses the detection of small events and masks flares that occur in rapid succession, leading to systematic under-counting and biased statistical properties.
• The Gap: There is a need for a high-resolution, automated detection method that can resolve overlapping events and identify flares at the "small-flux" end of the distribution without relying on arbitrary start/end definitions that fail during high-activity periods.

2. Methodology

The authors developed a novel framework using Convolutional Neural Networks (CNNs) to identify solar flares directly from high-resolution (1-second cadence) GOES SXR time series data.

• Data Source: GOES-R series satellite data (GOES-16, 17, 18, 19) from January 1, 2018, to August 22, 2025. The data utilizes 1-second resolution SXR flux in the 1–8 Å channel.
• Reference Catalog Construction:
  • To train the CNN, a new reference catalog was manually created by visually inspecting 145 randomly selected days spanning solar minimum to maximum.
  • Key Strategy: Instead of defining full start-to-end flare profiles (which are ambiguous during overlaps), the CNN was trained specifically to identify flare rise episodes. This relaxes the "slow-driving" assumption (that one flare must end before the next begins), allowing the detection of overlapping events.
  • The reference set contains 7,700 manually identified rise episodes.
• CNN Architecture:
  • Input: 600-sample (10-minute) overlapping windows of normalized SXR flux.
  • Structure: A hybrid architecture combining:
    • Multi-scale Convolutional Branches: Four parallel branches with kernel sizes of 3, 5, 7, and 11 to capture local temporal features.
    • BiLSTM (Bidirectional Long Short-Term Memory): Captures gradual flux changes and short-range dependencies.
    • Transformer Encoder: Uses self-attention mechanisms to capture long-range temporal relations and global context.
  • Output: Binary classification (Class 1: Rise Episode, Class 0: Background).
• Bayesian Confidence Estimation:
  • Since CNNs produce raw detections, a Bayesian framework was applied to quantify the probability ($P$) that a detection is a True Positive (TP) versus a False Positive (FP).
  • Features: The model uses peak flux amplitude ($A$), rise time ($\tau$), and a coincidence flag ($C$) indicating overlap with existing catalogs (GOES, HEK, SolO).
  • Logic: If a candidate coincides with a known catalog event ($C=1$), it is assigned $P=1$. Otherwise, $P$ is calculated based on the joint distribution of amplitude and rise time derived from the reference catalog.

3. Key Contributions

1. New High-Resolution Catalog: The generation of a catalog containing 111,580 flare candidates for the 2018–2025 period, which is 7.6 times larger than the corresponding GOES catalog (14,612 events).
2. Overcoming Obscuration: By focusing on rise episodes and using 1-second data, the method successfully detects overlapping flares and weak events that are invisible to 1-minute averaged algorithms.
3. Bayesian Confidence Framework: A rigorous method to assign probabilistic confidence to each detection, allowing researchers to filter the catalog based on desired reliability (e.g., $P > 0.9$ retains 62,896 events).
4. Resolution of Statistical Debates: The study provides a dataset large enough to re-evaluate long-standing debates regarding the correlation between flare sizes and waiting times.

4. Key Results

• Event Counts: The CNN catalog detects significantly more small events. For raw peak fluxes, A-, B-, and C-class events increased from 95/3,516/9,306 (GOES) to 406/25,071/83,205 (CNN).
• Power-Law Distributions:
  • Raw Flux: The CNN catalog shows a steeper power-law index ($\alpha = -2.59 \pm 0.02$) compared to GOES ($\alpha = -2.25 \pm 0.04$), indicating higher sensitivity to small events.
  • Background-Subtracted Flux: After removing background, the indices converge ($\alpha_{CNN} = -1.97 \pm 0.02$ vs. $\alpha_{GOES} = -2.05 \pm 0.04$). Crucially, the CNN catalog extends the power-law distribution by one order of magnitude at the low-flux end ($10^{-7}$ to $10^{-3}$ Wm$^{-2}$).
• Waiting Time Statistics:
  • Both catalogs broadly follow a piecewise Poisson process (memoryless behavior), consistent with Self-Organized Criticality (SOC) models.
  • The CNN catalog reveals a mean flare rate of 1.35 flares/hour, compared to 0.21 flares/hour for GOES.
• Obscuration and Asymmetry:
  • Previous studies reported a correlation where larger flares are followed by longer waiting times (time asymmetry).
  • The CNN analysis shows that this asymmetry disappears when weak/overlapping events are included. The apparent correlation in the GOES catalog was an artifact of obscuration (under-counting small events immediately following large ones).
  • The corrected data supports a time-symmetric, memoryless energy release process.

5. Significance

• Paradigm Shift in Flare Detection: This work demonstrates that deep learning can outperform traditional rule-based algorithms in complex, high-noise environments like the solar corona, particularly during solar maximum.
• Resolution of Physical Debates: By correcting for observational bias (obscuration), the study clarifies that solar flares likely behave as a spatially uncorrelated, memoryless process (SOC) rather than a stress-relaxation cycle with strong size-waiting time correlations.
• Foundation for Future Research: The new catalog provides a complete, consistent dataset for studying flare statistics, mapping events to Active Regions (ARs), and improving space weather forecasting models. The authors note that the catalog is publicly available, facilitating further statistical analysis and machine learning applications in heliophysics.