astronomy/ai produced the result/Astronomy and Astrophysics 2025 · v2
Neural networks replace a slow statistical fit to find solar oscillations in Hα images
Astronomers studied sequences of solar images from the GONG telescope network to find oscillations on the Sun. Two convolutional neural networks were trained to stand in for a Bayesian fitting step that otherwise took a minimum of ten seconds per pixel.
spectrum · one line per step, placed by what the step does · bright lines used AI
Fast Bayesian spectral analysis using convolutional neural networks: Applications to GONG Hα solar data
Astronomy and Astrophysics, 2025
doi:10.1051/0004-6361/202452928 · record aix-00118 v2 · checked 2026-10-08
- AI was for
- Simulation surrogate
- Model family
- Convolutional neural network
- Checked by
- Held-out100000 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
The Sun is restless, and parts of it wobble. Long, cool threads of gas called filaments, suspended above the solar surface, can swing back and forth with regular periods. To spot such rhythms, astronomers take a long run of images and, for each tiny patch of the picture, ask which repeating cycles are present in how its brightness changes. The answer comes as a power spectrum: a plot of how much signal sits at each frequency. The hard part is deciding what counts as real. Every spectrum carries noise that is itself uneven, stronger at some frequencies than others, so a peak only means something once that noisy background has been modelled.
The researchers worked with Hα images from the GONG network, taken at one-minute intervals by six telescopes around the world. Hα is a particular red wavelength of light emitted by hydrogen, and it shows filaments clearly. After assembling each day into a cleaned image stack, they computed a spectrum for every pixel and described the background as a mixture of two kinds of noise. Fitting that background properly, with uncertainty attached, is a Bayesian calculation, and it is slow: the authors give a minimum of ten seconds per pixel, which works out at roughly 230 days of computing for a single observation day on one processor core.
Where AI came in
To avoid that wait, the team trained two one-dimensional convolutional neural networks. A network of this kind reads a sequence of numbers and learns to recognise patterns in its shape, in the same way an image network learns edges and textures. Each network takes one normalised spectrum and returns three numbers: one network gives the parameters of the best-fit background curve, the other the parameters of a curve marking the 95% confidence level, above which a peak is treated as a detection. The training examples came from the slow Bayesian fits themselves, supplemented by two million synthetic spectra built by perturbing those fits.
So the networks stand in for the fitting step, not for the astronomy around it. The rest of the pipeline is unchanged: the spectra are still computed in the usual way, and pixels are still flagged by comparing measured power with the confidence curve. On a test set of 100,000 real spectra, the networks' curves differed from the Bayesian ones by a mean relative error of 8% for the best-fit line and 4% for the confidence line. Run over the whole solar disk, the trained networks produced their parameters in about five minutes. Using them, the analysis recovered filament oscillation events already catalogued for two days in 2014, and found oscillations on a day in 2024 that had not been examined before.
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
The authors analysed Hα image sequences from the GONG telescope network by building daily data cubes and computing a generalized Lomb-Scargle periodogram for every pixel, modelling the background as a combination of red and white noise fitted with Bayesian MCMC. Because that fit took a minimum of 10 seconds per pixel, they trained two convolutional neural networks on 2×10^6 synthetic periodograms, derived from MCMC posteriors, to predict the parameters of the best-fit line and of the 95% confidence line directly from a periodogram. On a test set of 10^5 observational periodograms, the CNN parameters differed from the MCMC ones by a mean relative error of 8% for the best-fit line and 4% for the confidence line, with about 2% of periodograms averaging more than 10% error. Using the CNN curves as the detection threshold, the analysis recovered filament oscillation events catalogued for 1 January 2014 and 13 February 2014, and identified oscillations, flare-related and otherwise, on 22 October 2024, a day not previously studied.
How AI was used
Two 1D convolutional neural networks were trained to emulate a Bayesian MCMC fit of a red-plus-white-noise background model to Lomb-Scargle power spectra. Each network takes a normalised periodogram as input and outputs the three background parameters (A, α and B) for either the best-fit curve or the 95% confidence curve. The architecture is four convolutional blocks, each of two 1D convolutional layers with 64 kernels and ReLU activation followed by max-pooling with filter size 2, with kernel sizes 5 and 8 in the first block and 11 and 10 in the rest, feeding dense layers of 1024, 512 and 256 units with 20% dropout and a three-unit sigmoid output. Training targets came from MCMC fits computed with PyMC on real GONG data, augmented by 2×10^6 synthetic periodograms generated by perturbing the parameterised MCMC posteriors and multiplying model curves by exponentially distributed draws; the synthetic set was split 80-20 for training and validation, inputs scaled to the unit interval and outputs log-transformed and min-max scaled. Training used the Adam optimiser with mean-squared-error loss, a learning-rate decay schedule, early stopping on validation loss, batch size 32 and up to 100 epochs on an NVIDIA RTX A5000 GPU. The trained networks were then run over every pixel of the solar disk, and detections were taken where the measured PSD met or exceeded the predicted 95% confidence curve within frequency bins of 0.055 mHz.
The shape of the work
Structural · the record, drawn
no AI
Assemble daily Hα image sequences from the GONG network
Obtaining raw data, whether by measurement, download or retrieval.
In this work, we analyze the temporal sequences of GONG H α images.where the paper describes this · verbatim
no AI
Correct and clean the data cube
Cleaning, filtering, normalising or labelling data already obtained.
Once the image sequence is constructed, we remove limb darkening from the images.where the paper describes this · verbatim
no AI
Compute per-pixel power spectral density
Encoding data into features, descriptors, embeddings or graphs.
For this reason, we use the generalized Lomb-Scargle periodogram to generate the PSD.where the paper describes this · verbatim
no AI
Build training targets with Bayesian MCMC and synthetic periodograms
Cleaning, filtering, normalising or labelling data already obtained.
We generate a total of 2×106 synthetic periodograms, each associated with a best-fit and confidence interval parameter set.where the paper describes this · verbatim
AI
Train the two CNN models
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
We used the Adam optimizer and a Mean-Squared Error for training the model.where the paper describes this · verbatim
AI
Predict background-fit and confidence-line parameters for every pixel
Running a trained model over new data to predict, classify or score. The AI stood in for statistical model.
capable of estimating the best-fit and confidence line parameters for each pixel in the whole solar disk in approximately 5 minuteswhere the paper describes this · verbatim
no AI
Flag pixels whose power exceeds the confidence threshold
Reducing a candidate set by filtering or ranking, in a single pass.
the ratio of the PSD to the 95% confidence curve is calculated, considering only the cases of positive detectionwhere the paper describes this · verbatim
no AI
Compare CNN output with MCMC fits and with catalogued events
Testing outputs against ground truth.
we have created a testing dataset of 105 periodograms selected randomly from different observation dayswhere 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 oscillation detections reported for all three observation days are derived from the CNN-predicted best-fit and 95% confidence curves, and the paper's stated aim is to demonstrate the CNN replacement for the MCMC fit.
We built two CNN models to obtain the same results as the MCMC approach.where the paper describes this · verbatim
The mean relative error for the best-fit line is approximately 8%, while for the confidence interval, it is 4%.where the paper describes this · verbatim
The code is publicly available in the following repository:where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- Version of CNN best-fit parameter modelWhich version of the model was used is not stated.
- Version of CNN 95% confidence-line parameter modelWhich version of the model was used is not stated.
About this article
Record aix-00118, version 2, checked by a person on 2026-10-08. 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