astronomy/ai produced the result/The Astrophysical Journal 2023 · v2
Image classifiers flag a gas kink in a young star's disc, pointing to a planet
Six off-the-shelf image classifiers were run on archival radio images of the disc around the star HD 142666. Their outputs, and the patterns they responded to, pointed the astronomers to a disturbance in the gas about 75 au from the centre.
spectrum · one line per step, placed by what the step does · bright lines used AI
Kinematic Evidence of an Embedded Protoplanet in HD 142666 Identified by Machine Learning
The Astrophysical Journal, 2023
doi:10.3847/1538-4357/acc737 · record aix-00248 v2 · checked 2026-10-09
- AI was for
- Classification
- Model family
- Convolutional neural network
- Checked by
- None stated
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Young stars are often wrapped in a flat, spinning disc of gas and dust, the raw material from which planets form. The planets themselves are usually hidden inside that material, so astronomers look for the marks they leave. One such mark is in the motion of the gas. Gas orbiting on its own follows a smooth, predictable pattern set by the star's gravity. A planet tugs on the gas nearby and bends that pattern locally. Radio telescopes record this by splitting the light of a carbon monoxide gas line into many slices, each showing gas moving at a slightly different speed. A planet can show up as a small kink in one of those slices, and kinks are easy to miss.
The researchers took existing images of the disc around the star HD 142666, drawn from a public archive of such observations, and looked for a kink of that kind. Rather than scanning the slices by eye first, they put the stack of speed slices through classifiers they had trained earlier, and then followed up on what those classifiers pointed at.
Where AI came in
The classifiers were convolutional neural networks, a type of program that learns to sort images by example. Two standard designs were used, EfficientNetV2 and RegNet, each in three versions taking a different number of speed slices, making six models in all. They had been trained on simulated discs with and without planets, and were run on the real observation without further adjustment. Each returned a score for the two options, planet or no planet; the average for the planet class was over 0.84. The researchers also looked at which parts of the images the networks responded to internally, which indicated the slice and the distance from the centre where the kink lay.
The networks stood in for the initial eye-and-judgement sweep through the speed slices. The paper states that the conventional analysis of that feature was carried out only because the models had pointed to it. The location they indicated then set where a planet was placed in follow-up fluid-dynamics simulations, which were turned into mock images for comparison with the real one; these were not AI methods. A planet of five Jupiter masses at 75 au gave the closest match.
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
Six previously trained convolutional classifiers (EfficientNetV2 and RegNet variants) were applied to archival ALMA 12CO channel maps of the protoplanetary disk HD 142666 from the DSHARP catalogue. The average softmax value was over 0.84 for the planet class, and the models' activation structures highlighted a localized kink in the gas kinematics roughly 75 au from the disk centre, a feature the paper reports as previously unreported. SPH simulations with PHANTOM, varying embedded planet mass between 1 and 5 Jupiter masses and post-processed with MCFOST, reproduced the observed kinematic structure for a 5 Jupiter-mass planet at 75 au. The authors conclude that HD 142666 hosts a planet.
How AI was used
Archival DSHARP 12CO channel maps of HD 142666 were cropped to the disk, a subset of velocity channels selected, and the channels reshaped to 600x600 pixels and normalised to the range 0 to 1. The resulting position-position-velocity cube was passed to six convolutional classifiers from the authors' earlier work: EfficientNetV2 and RegNet implementations built in PyTorch, each in three versions taking 47, 61 or 75 input velocity channels, originally trained on synthetic MCFOST observations of PHANTOM SPH disks with and without planets. The models were run without further fitting to these data. Each produces a two-component softmax output over planet and no-planet classes, together with maps of internal activation structure; both were inspected by the authors, the softmax values to assess whether a planet is present and the activations to judge which velocity channel most likely contains the non-Keplerian feature. The radial location indicated this way set where a planet was placed in follow-up PHANTOM SPH simulations, which were converted into synthetic channel maps with MCFOST and convolved to the observed resolution for comparison against the observation.
The shape of the work
Structural · the record, drawn
no AI
Obtain archival ALMA observations of HD 142666
Obtaining raw data, whether by measurement, download or retrieval.
The HD 142666 data was taken from the DSHARP catalogue.where the paper describes this · verbatim
no AI
Crop, reshape and normalise channel maps
Cleaning, filtering, normalising or labelling data already obtained.
The channels were reshaped to 600×600 pixels and normalized such that all pixel values were between 0 and 1.where the paper describes this · verbatim
AI
Run six previously trained classifiers on the observed cube
Running a trained model over new data to predict, classify or score. The AI stood in for expert judgement.
we apply our previously trained models to ALMA data of the HD 142666 systemwhere the paper describes this · verbatim
no AI
Inspect softmax values and activations to localise the kink
Extracting understanding from model behaviour.
We inspect the softmax values and activation structures to gain insight into whether a planet might be in the systemwhere the paper describes this · verbatim
no AI
Run SPH simulations with embedded planets
Numerical or physics simulation, including where a learned surrogate replaces it.
We run a suite of SPH simulations using PHANTOM, varying the mass of the embedded planet between 1 and 5 MJ.where the paper describes this · verbatim
no AI
Generate synthetic channel maps by radiative transfer
Numerical or physics simulation, including where a learned surrogate replaces it.
For each simulation, we create channel maps using MCFOST in the same way that the original training data was made.where the paper describes this · verbatim
no AI
Compare simulated and observed kinematic structure
Testing outputs against ground truth.
There is strong agreement between this feature and the non-Keplerian channel identified by our modelswhere 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 non-Keplerian feature that the paper reports was located by applying the classifiers and inspecting their activations; the paper states the traditional analysis was only performed because of the information given by the models. Verification was by non-AI hydrodynamics, so a reviewer may prefer 'analysis'.
We use two different architectures: EfficientNetV2 and RegNet.where the paper describes this · verbatim
The HD 142666 data was taken from the DSHARP catalogue.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.
- ValidationNo validation of the AI is described.
- Version of EN47 (EfficientNetV2, 47 input velocity channels)Which version of the model was used is not stated.
- Version of EN61 (EfficientNetV2, 61 input velocity channels)Which version of the model was used is not stated.
- Version of EN75 (EfficientNetV2, 75 input velocity channels)Which version of the model was used is not stated.
- Version of RN47 (RegNet, 47 input velocity channels)Which version of the model was used is not stated.
- Version of RN61 (RegNet, 61 input velocity channels)Which version of the model was used is not stated.
- Version of RN75 (RegNet, 75 input velocity channels)Which version of the model was used is not stated.
About this article
Record aix-00248, version 2, checked by a person on 2026-10-09. 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