Alexis: When and Where is that Flare?
| Nugget | |
|---|---|
| Number: | 99999 |
| 1st Author: | Jorge R. PADIAL DOBLE |
| 2nd Author: | Kelly HOLLEY-BOCKELMANN |
| Published: | October 26, 2026 |
| Next Nugget: | |
| Previous Nugget: | The Hale Sector Boundary links flares to global magnetic fields |
A light curve without a location
Every X5.0 solar flare headline you have read about is named after a single number: the soft X-ray flux of the entire unresolved solar disk measured by the GOES/XRS soft X-ray sensor. This instrument tells us how strongly the Sun is emitting Earth-bound X-rays at any given moment but, as a single-pixel detector, it can never tell us where the X-rays are being produced. Historically, locations must be added afterward by NOAA hand-combing imagers but more than half of all reported C-class flares and above end up with no location at all. This matters far beyond bookkeeping: flare-prediction machine learning models are trained on magnetic field data from HMI active region patches (HARPs/SHARPs) [1] and labelled as flaring or non-flaring magnetic fields with catalogs like SWPCs and SolarSofts. For comparison, what would happen if, for example, you want a model to identify pictures of dogs, but half of the images of dogs the model is trained on are actually cats?
Recreating the light curve, region by region
We (Ref. [1])have built ALEXIS (Automatically Labeled EUV and X-Ray Incident SolarFlares) to ask a simple question: can most of the solar X-ray emission be explained by smaller/constrained regions on the Sun using real data? Specifically, ALEXIS takes the raw GOES XRS data (figure 1A), identifies discrete regions of interest in all the SDO/AIA EUV wavelengths (figure 1B), and reconstructs the X-ray light curve as a weighted linear combination of those regional EUV curves (figure 1C). We find the minimum set of regions that explains the signal, and each surviving region becomes a catalog entry with peak time, coordinates, correctly scaled X-ray class, and the associated HARP number, thus relating the flare to a large amount of SDO and other supporting data (Ref. [2]). We also repair the long-neglected GOES Solar X-ray Imager records, correcting pointing errors and making them usable by standard Python tools (Astropy, SunPy).
The ALEXIS procedures are:
1. Raw input. Left: 131 131Â EUV solar image with the solar limb. Right: normalized GOES XRS flux (dark red, top-right)
2. Region selection and integration. ALEXIS finds candidate flaring regions (clusters 0 and 1) and integrates the pixel flux within each box to form its own light curve (arrows map each curve to its region).
3. Linear-combination fit and flare definition. The optimal weighted linear combination of the two region curves (gray) that reproduces the XRS curve as closely as possible. Think aX+By=Z where X and Y are the blue and orange curves, and Z the gray fit; the goal is to find a and b. The ALEXIS flare events are then identified marking cluster 0 at 18:06 UTC (blue star) and cluster 1 at 18:17 UTC (orange star).
4. Metadata overlay and catalog comparison. ALEXIS places external catalog information -- SolarSoft/SWPC peak times, flare durations (pink: NOAA, yellow: SolarSoft) and flare locations (arrows), with the gray HARP boundary boxes shown for context. The external catalog records a single event peaking at 18:15 UTC (pink/yellow stars). Both mis-identify the timing and location of the flare.
What the current catalogs are missing
On a randomized test bed of 1000 one-hour segments (2010 May to 2020 March), ALEXIS reproduces the XRS signal with an rms error of 0.066&plusminus;0.036, using only a handful of regions per event. When more than one region is needed, the extra components turn out to be physically meaningful: flares firing in different active regions both simultaneously (synchronous) or minutes apart (sympathetic), or bright emission lingering before and after the main event.
For example, Figure 1D shows the light curve for a SWPC/SolarSoft reported M2.8 flare, SOL2011-09-24T18:15 UTC. The ALEXIS pipeline identified that the X-rays are actually produced by two distinct signals: one M5.8 flare at 18:06 UTC near the eastern limb and another M5.9 flare later on the western limb. Of the flares ALEXIS examined, 62% of SWPC entries and 15% of SolarSoft entries need revisiting: incomplete peak times and duration, incomplete locations or active-region associations, or flares that were actually two or more events. Meanwhile, ALEXIS reports roughly 15% more flares than either catalog while missing only 6.7% of a 1057-flare subsample and returning 5% false positives across 1211 detections.
An upgrade path for flare science
These are proof-of-method results from a subsample; the full census, now in preparation, will analyze more than 11 million SDO/AIA images (600+ TB uncompressed) covering over 8000 C-Class and above events from 2010-2020. A complete, self-consistent, interpretable, spatially resolved flare catalog allows for precise and sharper training labels for prediction models, cleaner statistics for flare-energy studies, and raises questions about how flaring regions communicate across the disk.