astronomy/ai produced the result/Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences 2026 · v2
Neural network trained on simulated data hunts solar wave trains in radio spectra
Researchers trained a neural network on synthetic solar flare signals to spot quasi-periodic fast propagating wave trains, then ran it over a year of Japanese radio observations, reporting 50 independent candidate events.
spectrum · one line per step, placed by what the step does · bright lines used AI
Detector for fast wave trains in the solar radio emission
Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences, 2026
doi:10.1098/rsta.2025.0056 · record aix-00274 v2 · checked 2026-10-23
- AI was for
- Detection, Classification
- Model family
- Convolutional neural network
- Checked by
- Held-out51 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
The Sun's outer atmosphere, the corona, carries waves. Among them are quasi-periodic fast propagating wave trains, or QFPs: bursts of ripples that travel outwards from a flare, arriving as a rapid series of pulses rather than a single push. They matter because the speed and spacing of such ripples carry information about the gas they move through. The trouble is that they are hard to find. A wave train shows up as a faint wobble riding on top of a much larger, slowly changing brightness, and real measurements are full of noise and sharp spikes that can mimic a wobble. Only a few dozen observed examples existed to learn from.
The authors set out to build an automatic detector and point it at radio data. Radio spectra record how much the Sun emits at each frequency over time, so a search through them could turn up wave trains that nobody had looked for by hand. They used observations from the Japanese HiRAS instrument covering 20 megahertz to 2.5 gigahertz during 2011, concentrating on dates when a large-scale wave had already been catalogued crossing the Sun.
Where AI came in
A one-dimensional convolutional neural network did the deciding. Convolutional networks learn to recognise shapes in data by sliding small pattern templates across it; here the data was a brightness-versus-time curve, with the slow background removed and the leftover wobble rescaled. The network answered one yes-or-no question: does this curve contain a wave train? Because real examples were scarce, it learned from artificial curves built by adding together a drifting background, noise, and either a digitised wave train, a deliberately misleading signal such as a spike or a plain sine wave, or nothing at all.
The network stood in for the eye and judgement of a specialist sifting through spectra. On a test set of 51 real curves from published events, plus 1000 noise and spike curves, it reached 38% precision, 61% recall and 95% specificity. Applied to the radio data, after a simpler correlation check narrowed down where to look, it produced 132 candidate detections, consolidated by human inspection into 50 independent candidate events, 13 of them linked to large-scale coronal waves.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
A one-dimensional convolutional neural network was trained on synthetic solar flare time profiles to decide whether a detrended profile contains a quasi-periodic fast propagating (QFP) wave train, and then applied to HiRAS dynamic radio spectra covering 20 MHz to 2.5 GHz during 2011. Of 102 global coronal EUV waves listed in the LCPFs catalogue for 2011, 50 had HiRAS data available, and the search over those dates produced 132 candidate detections, of which 75 were judged phenomenologically similar to a wave train. After accounting for single events split across frequency bands, the authors report 50 independent QFP-candidate events, 13 of them connected with global waves. On a test sample of 51 real time profiles taken from published events plus 1000 noise and spike profiles, the detector gave 38% precision, 61% recall and 95% specificity at threshold T=0.5.
How AI was used
Because only a few dozen observed wave trains exist, the authors generated synthetic flare time profiles of fixed length by additively combining a slowly varying trend, a signal component (a wave train digitised from published simulations and observations, a 'bad' signal such as a spike, sinusoid or beat, or none) and white plus red noise, with augmentation by time shifting and scaling of 38 basic wave train profiles. Profiles were detrended with a forward-backward Butterworth filter and normalised to their own standard deviation, then split 80:15:5 into training, validation and test sets for supervised binary classification with a 1D CNN using BCEWithLogitsLoss and AdamW; the network score is thresholded at a chosen value T. The trained detector was checked on the synthetic test split and on time profiles extracted from SDO/AIA time-distance maps for 16 published events together with added noise and spike profiles. For the radio search, HiRAS dynamic spectra on marker dates from the LCPFs global-wave catalogue were detrended per frequency, split into 320, 640 and 1280 second windows and binned to a fixed input length; mean cross-correlation coefficients between neighbouring frequency channels were used to select which regions of the spectrum to pass to the network, and positive triggers were then inspected visually with their time profiles and wavelet spectra and grouped into independent events.
The shape of the work
Structural · the record, drawn
Build synthetic training set of flare time profiles
Producing candidate objects that did not previously exist.
we generate synthetic solar flare time profiles containing wave trains in sufficient quantitieswhere the paper describes this · verbatim
Detrend and normalise time profiles
Cleaning, filtering, normalising or labelling data already obtained.
We have preferred the Butterworth filter for the trend removal because it works faster than other filters.where the paper describes this · verbatim
Train binary CNN classifier
Fitting model parameters, including fine-tuning an existing model. The AI stood in for expert judgement.
The network was trained as a binary classification with a loss function BCEWithLogitsLoss and AdamW optimizer.where the paper describes this · verbatim
Test detector on synthetic and real profiles
Testing outputs against ground truth.
We collect all the selected time profiles (51 in total) to create the sample of real data to test our neural network.where the paper describes this · verbatim
Select and preprocess HiRAS dynamic spectra
Cleaning, filtering, normalising or labelling data already obtained.
The resulting high-frequency components formed a new dynamic spectrum, which was then split into time windows of three different scaleswhere the paper describes this · verbatim
Restrict search regions by cross-correlation metric
Reducing a candidate set by filtering or ranking, in a single pass.
The mean value over 20 obtained cross-correlation coefficients forms the new dynamic spectrum as metric for the next detection step.where the paper describes this · verbatim
Run detector over selected time profiles
Running a trained model over new data to predict, classify or score. The AI stood in for expert judgement.
The time profiles in these regions were passed as an input into the neural network detector.where the paper describes this · verbatim
Visual control and consolidation of detections
Reducing a candidate set by filtering or ranking, in a single pass.
as well as wavelet spectra of the high-frequency component for additional visual controlwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The reported QFP-candidate events are the output of the trained neural network detector applied to HiRAS dynamic spectra; the paper's result is the detection list itself.
CNN inputs a detrended flare time profile at fixed size and performs binary classificationwhere the paper describes this · verbatim
We obtained 38% precision and 61% recall and 95% specificity for threshold T=0.5where the paper describes this · verbatim
They are publicly available at https://solarobs.nict.go.jp/.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of 1D convolutional neural network QFP detectorWhich version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00274, version 2, checked by a person on 2026-10-23. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Neural network trained on simulated data hunts solar wave trains in radio spectra. Record aix-00274, version 2, checked 2026-10-23. https://aixsci.org/articles/aix-00274/