Locating Hidden Exoplanets in ALMA Data Using Machine Learning

Draft version November 18, 2022
Typeset using L
A
TEX twocolumn style in AASTeX631
Locating Hidden Exoplanets in ALMA Data Using Machine Learning
J. P. Terry ,1, 2
C. Hall ,1, 2
S. Abreau ,3, 4
and S. Gleyzer 5
1Department of Physics and Astronomy, The University of Georgia, Athens, GA 30602, USA.
2Center for Simulational Physics, The University of Georgia, Athens, GA 30602, USA.
3Department of Medicine, Division of Cardiology, University of California, San Francisco, CA, 94143, USA.
4Cardiovascular Research Institute, San Francisco, CA, 94158, USA.
5Department of Physics and Astronomy, The University of Alabama, Tuscaloosa, AL 35487, USA
(Received Augusts 5, 2022; Revised November 16, 2022; Accepted November 18, 2022)
ABSTRACT
Exoplanets in protoplanetary disks cause localized deviations from Keplerian velocity in channel
maps of molecular line emission. Current methods of characterizing these deviations are time consum-
ing, and there is no unified standard approach. We demonstrate that machine learning can quickly
and accurately detect the presence of planets. We train our model on synthetic images generated from
simulations and apply it to real observations to identify forming planets in real systems. Machine
learning methods, based on computer vision, are not only capable of correctly identifying the presence
of one or more planets, but they can also correctly constrain the location of those planets.
Keywords: Hydrodynamics — Radiative transfer — Accretion disks — Methods: numerical — Catalogs
— Planets and satellites: formation
1. INTRODUCTION
In recent years, the kinematic analysis of protoplane-
tary disks has proven important in demonstrating that
ring and gap-like features are indeed due to the presence
of unseen planets (Pinte et al. 2018; Teague et al. 2018).
The spiral wake from planets causes localized deviations
from Keplerian motion known as “kinks”. Other mecha-
nisms, such as gravitational instability (Hall et al. 2020),
vertical shear instability (Barraza-Alfaro et al. 2021),
and magnetorotational instability (Simon et al. 2015),
may leave similar kinematic imprints.
Recent kinematic observations of the HD 97048 sys-
tem revealed a localized velocity deviation signaling the
presence of a forming exoplanet (Pinte et al. 2019) too
faint to be observed directly at ∼mm wavelengths. To
estimate the location of the forming exoplanet, several
computationally expensive hydrodynamics simulations
were run, synthetic observations were generated and
compared “by-eye” to real observations to obtain the
best fit.
Machine learning offers an alternative route. The use
of machine learning in astronomy has exploded in the
last decade (Jo & Kim 2019; Möller & de Boissière 2020;
Alexander et al. 2020). The study of protoplanetary
disks largely relies on images, so machine learning tech-
niques for computer vision are perfectly suited for ap-
plication to these kinematics observations. Computer
vision is a field of AI that focuses on image and video
analysis. From simple transcription of hand-written dig-
its (Cireşan et al. 2011) to locating and classifying ob-
jects pixel-by-pixel (Minaee et al. 2020), computer vi-
sion has a wide range of uses. The introduction of ad-
vanced architectures, such as residual (He et al. 2015)
and inception (Szegedy et al. 2014) blocks, along with
qualitatively new algorithms, such as the vision trans-
former (Vaswani et al. 2017; Dosovitskiy et al. 2020), has
led to computer vision models exceeding human experts
at times (Zhou et al. 2021).
Given the applicability and strong performance of
computer vision techniques, we have applied them to
the kinematic analysis of protoplanetary disks.
Previous work has successfully applied deep learning
to the problem of inferring properties of planets within
protoplanetary disks (Auddy et al. 2021, 2022) using
models trained on simulated data, but these efforts do
not use kinematic information and instead rely on gaps
as the signatures. We demonstrate that a trained net-
work can detect the presence of planets in large datasets
much faster than a human, recover all previously iden-
tified planets, and possibly identify ones that have been
arXiv:2211.09541v1
[astro-ph.EP]
17
Nov
2022

2 Terry et al.
missed in previous analyses. Inferring other properties,
such as the number of planets and their masses, may
be within the reach of machine learning techniques as
well, which we leave to future work. This paper presents
a proof of concept application of a relatively simple
machine algorithm applied to protoplanetary accretion
disks.
The paper is arranged as follows: Section 2 describes
the training data and the details of the machine learning
models and their training. Section 3 presents the results,
interpretations, suggestions for future work, and limita-
tions. Section 4 gives our conclusions.
2. METHODS
Training a model to predict the presence of a planet
requires a large amount of data spanning a large param-
eter space. We run 1,000 simulations in this work. The
parameter space is sampled using a Latin Hypercube
(LHC) (McKay et al. 1979), which accounts for previ-
ously sampled points to evenly distribute values across
the entire parameter space. We sample values from
observed ranges inferred from disk surveys (Andrews
et al. 2009; Andrews et al. 2018; Huang et al. 2018) and
widely accepted simulational parameters (Pinte et al.
2018, 2019, 2020).
Table 1 shows the variables sampled and their min-
imum and maximum value. Approximately one third
of the simulations have no planets in to address any
class imbalance (i.e., not having an equal number of “at
least one planet” class and “no planet” class simula-
tions) while ensuring diversity in the number of planets.
After sampling the stellar mass for use in the simulation,
stellar temperatures and radii, both of which are used in
the radiative transfer calculations, are determined using
isochrones calculated by Siess et al. (2000).
2.1. Hydrodynamical simulations
We run 1,000 3D smoothed particle hydrodynamics
(SPH) simulations using the PHANTOM (Price et al. 2018)
code. System properties, such as stellar mass, disk prop-
erties, observational resolution etc. were LHC-sampled
from Table 1. A significant fraction (25%) of these sim-
ulations were withheld for testing. Each simulation uses
106
SPH particles and evolves for 100 orbits. A given
planet’s accretion radius is set to 1/4 of its Hill radius.
The surface density profile is set to Σ ∝ R−p
, and the
sound speed is set to cs ∝ R−q
.
2.2. Velocity channel maps
The simulation results are used to create velocity
channel maps in only 13
CO to keep the size of the pa-
rameter space manageable, although 12
CO and C18
O are
also reasonable choices. The channel maps are created
using the MCFOST radiative transfer code (Pinte et al.
2006, 2009). For each simulation, multiple points be-
tween 10 and 100 orbits are chosen randomly. Since one
third of the simulations have no planet, six snapshots
are chosen from each of those simulations while only two
snapshots are chosen for simulations with planets. This
eliminates the class imbalance. From each snapshot, a
velocity channel cube is made using either the 13
CO J
2→1 (220.4 GHz) or 3→2 (330.6 GHz) transition, which
is randomly sampled. These transitions are used be-
cause they were also used to identify the planets in HD
163296 and HD 97048 (Pinte et al. 2018, 2019). The
final result is a velocity cube, where each velocity chan-
nel is an image with 600x600 pixels at a resolution of 1
au/pixel. This velocity cube is used as the model input.
Channel maps are calculated using 108
photon packets.
The dust-to-gas ratio is 1:100, and dust is composed of
carbon and silicates (Draine & Lee 1984). Stellar radii
and temperatures are calculated with isochrones given
by Siess et al. (2000) using the LHC-sampled masses and
ages. A spherical beam is used.
2.3. Simulated Observations
The velocity channel maps are convolved spatially and
spectrally, and noise is added by LHC sampling from
Table 1. RMS noise is added independently for each
pixel, which more faithfully replicates observational ef-
fects such as hot pixels. This method results in an av-
erage RMS of ∼ 10% for each image.
Except for the specific values, this process largely fol-
lows that done in Terry et al. (2022). Figures 1 and 2
show the results of convolving selected velocity channels
for a disk with and without a planet, respectively. Fig-
ure 1 shows that while the perturbation is strongest in
the velocity channel that covers the planet (rightmost
column), the perturbation is not necessarily completely
localized and can be visible in distant regions of the disk.
This is an expected behavior. The deviation is strongest
at the location of the planet, but the spiral wake leaves
further imprints as it disperses throughout the disc.
By itself, a single velocity channel is poorly suited
for our method. Only a few channels display kinks, and
channels with kinks can neighbor the exact channel con-
taining the planet. We therefore stack all C velocity
channels into a single image of dimension (H, W, C),
where H and W are the height and width of the image
(600x600 pixels for these purposes).
The number of velocity channels in an observation
varies, but mostly fall in the range ∼ 40 − 80 channels
that cover the disk (Andrews et al. 2018; Öberg et al.
2021). We therefore use C=47, 61, and 75 channels for

Locating Exoplanets Using Machine Learning 3
Parameter Variable Minimum Value Maximum Value Use
Number of planets N 0 4 PHANTOM, MCFOST
Disk Inner Radius (au) Rout 150 250 PHANTOM, MCFOST
Disk Outer Radius Rin 0.1Rout 0.1Rout PHANTOM, MCFOST
Planetary Mass (M⊕) Mplanet 5 120 PHANTOM, MCFOST
Stellar Mass (M ) M? 0.5 2.0 PHANTOM, MCFOST
Planet Semimajor Axis (au) a Rin + 0.2Rout Rin + 0.8Rout PHANTOM
Mass Ratio Mdisc / M? 5×10−4
10−2
PHANTOM
Disk Viscosity αSS 5×10−4
2.5×10−3
PHANTOM
Surface Density Profile (Σ = Σ0R−p
) p 0.1 1.0 PHANTOM
Sound Speed Profile (cs = cs,0R−q
) q 0.1 0.75 PHANTOM
Stellar Age (Myr) t 0.5 5.0 MCFOST
Inclination (degrees) i 10 80 MCFOST
Azimuth (degrees) φ 0 360 MCFOST
Distance (pc) d 100 200 MCFOST
Spectral Resolution (km/s) ∆v 0.01 0.1 Synthetic Observations
Spatial Resolution (arcseconds) 00
0.025 0.25 Synthetic Observations
RMS Noise (%) δ 0.05 20 Synthetic Observations
Table 1. Simulation parameter ranges for the LHC sampling and the stage at which they were applied.
Figure 1. Example raw (top row) and convolved, noisy (bottom row) channel maps in a disk with a planet present. The planet
(circled in white) is visible as a kink in the right column. The opposite velocity channel is shown in the left column, and the
systemic channel is shown in the middle column. The beam size is indicated in the bottom middle image. This disk is one of the
smallest and farthest simulated and is observed with some of the worst spatial resolutions, which is why the beam is so large.
each observation. The number of channels must be odd
to cover the disk symmetrically and still retain the sys-
temic (δv = 0) channel. The input velocity of the sim-
ulated channels ranges from |δv| = |vsystemic − v| ≤ 2.5
km/s to fully cover the disk.

4 Terry et al.
Figure 2. Example raw (top row) and convolved, noisy (bottom row) channel maps in a disk without a planet present. The
beam size is indicated in the bottom middle image.
2.4. Machine Learning Models
The input for all models is a (600x600xC) image, and
their output is a two-component vector with softmax ac-
tivation (Goodfellow et al. 2016). The softmax function
maps the output vector such that the sum of the compo-
nents equals one. Each component can be interpreted as
the predicted probability that the input belongs to the
corresponding class (i.e., “contains at least one planet”
class or “does not contain a planet” class). An Adam
optimizer (Kingma & Ba 2014) is used to add momen-
tum, which remembers previous gradients, and gradient
scaling to a simple gradient descent algorithm. This
helps escape local minima and improves the speed and
range of convergence. The cross-entropy loss (Hinton
et al. 1995) penalizes confident predictions that are in-
correct. Training stops when there is no improvement
in the validation loss after eight complete epochs, i.e.
eight sweeps through the training data.
Multiple architectures are used to compare perfor-
mance on two tasks: predicting if the system hosts at
least one planet and determining its azimuthal location.
We use two different models based on PyTorch (Paszke
et al. 2019) implementations: EfficientNetV2 (Tan & Le
2021) and RegNet (Xu et al. 2021). Neither model uses
the default hyperparameters or pre-trained weights.
EfficientNetV2 is an updated version of Efficient-
Net (Tan & Le 2019) that maximizes performance and
minimizes the number of trainable parameters and train-
ing time. RegNet, on the other hand, seeks to maximize
performance by adding variants of convolutional recur-
rent neural networks to ResNet (He et al. 2015). This
can extract greater spatial information at the cost of
significantly more trainable parameters.
We perform Bayesian hyperparameter tuning using
WANDB (Biewald 2020) to find separate sets of hyper-
parameters that minimize the validation loss for C=47,
61, and 75. The default parameters for these sweeps
are based on default versions “EfficientNetV2 S” and
“RegNetY 16GF.”
Each architecture is evaluated for accuracy and
area under the receiver operating characteristic curve
(AUC) (Hanley & McNeil 1982). Accuracy in this con-
text is defined as whether a model correctly predicted
that a system with planet(s) has at least one planet, and
similarly for systems with no planets. We choose AUC
because it gives a robust measure of the performance of
the model across different decision thresholds, i.e. the
softmax threshold required to claim the presence of a
planet. Substantial confidence is required to claim a
planet detection, but setting the decision threshold too
high can lead to false negatives. False positives may be

more acceptable because further analysis can overrule
them, whereas a false negative may lead to no further
analysis and result in the planet not being detected. The
optimum threshold for this boundary is open for discus-
sion by the community. We show multiple thresholds
(50%, 75%, 90% and 95%) in Figures 3 and 4 to demon-
strate how the threshold for claiming the presence of at
least one planet affects the results .
While some protoplanetary disks exhibit gaps in con-
tinuum that are highly suggestive of a planet, this
does not help one find the azimuthal location. Planet-
induced non-Keplerian kinks tend to be strongest near
the planet, so our model must determine if and where
the kink is in our data in order to give information on
the azimuthal location.
One simple approach to this is to look at the activa-
tions inside the model itself. The activation strengths
highlight regions which the model finds particularly in-
formative for classification. EfficientNetV2 uses SiLU
activations (Elfwing et al. 2017), and RegNet uses
ReLU (Glorot et al. 2011).We use mean-subtracted ac-
tivations to achieve this. For a given pixel with an ac-
tivation value, x, the mean-subtracted activation, x0
, is
defined as x0
= |x − hxi|, where hxi is the mean acti-
vation over the entire image. Activations occur most
strongly at the location of the strongest deviation from
Keplerianity, i.e. the kink.
3. RESULTS AND DISCUSSION
3.1. Machine learning Models
We train six models in total: an EfficientNetV2 and
a RegNet for 47, 61, and 75 velocity channels. Using a
typical decision threshold of 0.5, each model, other than
the 47-channel RegNet, has an accuracy of at least 93%.
When this decision threshold is increased to 0.95, four
models have an accuracy of at least 92%. Five models
have an AUC of at least 0.98. This confirms that our
models have learned to correctly detect the presence of
planets using kinematic signatures.
Table 2 shows the number of parameters as well as
relevant performance measures for each final model. All
EfficientNetV2 models converge to a similar number of
trainable parameters, whereas the number of parame-
ters for the RegNets increases sharply with the number
of channels. RegNet is a larger and more complicated
architecture, so this is expected. The performance of the
RegNet models varies widely compared to the Efficient-
NetV2 models. The accuracy of EfficientNetV2 models
varies by at most 8%, whereas RegNet varies by as much
as 31%. By all performance metrics, while the average
EfficientNetV2 outperforms the average RegNet, the 61-
channel RegNet outperforms all other models.
Figure 3 shows the confusion matrix, which describes
the distribution of correct and incorrect predictions, for
all models applied to the withheld test set. Panel (a)
shows the results using a 50% decision threshold, i.e. a
planet is predicted if the confidence is over 50%. Pan-
els (b), (c), and (d) show the same for cutoffs of 75%,
90%, and 95%. Figure 4 gives several relevant metrics
for model performance and all ROC curves. Claiming a
planet is present should come with high confidence, so
a 50% decision threshold isn’t necessarily what should
be used in practice. It is encouraging that, for all other
models than the 47-channel RegNet, there are no qual-
itative changes in the results when a higher decision
threshold is used; the main effect is simply increasing
the number of false negatives and decreasing the num-
ber of false positives.
If there are multiple planets in a disk, then multiple
kinks may be present. It is therefore prudent to check
that the quality of the predictions does not depend on
the number of planets in the disk. We check the pos-
sibility of this dependence by calculating the accuracy
for each model for disks with a given number of planets.
Figure 5 gives the results of these calculations. The p-
values from Pearson correlation tests shown in Figure 6
demonstrate that only the 47-channel RegNet using a
decision threshold of 50% has a statistically significant
relationship between accuracy and the number of plan-
ets (p < 0.05). While the 47-channel RegNet using a
decision threshold of 95% does not have a statistically
significant relationship, it is clear that the performance
is worse when there are planets, i.e. there are many false
negatives. This is in line with this model’s relatively low
AUC and Figure 3. The performance of all other models
does not differ significantly with the number of planets.
3.2. Simulation Activations
Many disks do not show obvious signs of planets, ei-
ther in continuum or line emission. Inspecting the acti-
vations of the disks helps overcome this. Figure 7 shows
the result from a disk that is difficult to classify by eye.
The models predict the presence of an exoplanet with
high confidence, activating on a subset of channels that
are deemed important by the model. This gives informa-
tion on the azimuthal location, and the radial location
may be inferred by either the location with the strongest
kink or coincident gaps in continuum images.
Panels (c) and (d) in Figure 7 show the raw and con-
volved velocity channels, respectively. If one were sim-
ply inspecting the velocity channels by eye, it would be
trivially easy to miss the kink even in the unconvolved
data. The convolved, noisy data shows essentially no
sign of the planets. Despite lacking any clear signatures

6 Terry et al.
Model Parameters Accuracy (50%) Accuracy (95%) AUC
(millions) (%) (%)
EfficientNetV2: 47 Channels 20.2 97 ± 0.5 96 ± 0.5 0.99 ± 0.002
RegNet: 47 Channels 51.0 78 ± 1.1 65 ± 1.3 0.86 ± 0.010
RegNet: 61 Channels 62.8 98 ± 0.4 96 ± 0.6 > 0.99 ± 0.001
RegNet: 75 Channels 114 95 ± 0.6 92 ± 0.7 0.98 ± 0.003
Table 2. Final model descriptions and metrics. 95% confidence intervals are calculated by bootstrapping each metric 1,000
times using random selections of 80% of the test data. The percentages next to the accuracy labels denote the decision threshold.
that can be observed visually, the models correctly clas-
sify the disks and locate the planets. This demonstrates
that our models are able to accurately find planets that
the by-eye method may fail to find.
Figure 8 shows a typical activation for a disk with no
planet. The entire disk is activated equally, indicating
that there is no localized region that exhibits behaviors
indicative of a planet.
3.3. Application to Real Observations
Next, we demonstrate the effectiveness of our model
by applying it to real telescope observations. Kinematic
detections of planets have been demonstrated for many
disks (Pinte et al. 2020), but here we focus on the HD
163296 and HD 97048 disks as a proof of concept (Pinte
et al. 2018, 2019). Using the same data as Pinte et al.
(2018) and Pinte et al. (2019), we demonstrate that our
models replicate the prediction and estimated location
of the forming exoplanet. Figure 9 (a) shows the HD
97048 in continuum, and panel (b) shows the velocity
channel with the planet’s signature circled.
Table 3 shows the softmax values for the observations.
For HD 97048, all RegNet models correctly predict a
planet with > 95% confidence. Two EfficientNetV2
models correctly predict a planet with ≥ 85% confi-
dence. Figure 9 shows that the velocity channels that
contain the planet are strongly activated. The activa-
tion in Figure 9 (d) is strong on the planet itself rather
than simply the entire velocity channel.
The results for HD 163296 are not as confident, but
the two best-performing models predict the presence of
at least one planet with over 99% confidence.
Both classes of model perform better when there are
more input channels, a trend not seen for HD 97048.
A possible explanation is the fact that there are about
twice as many velocity channels that cover HD 163296
when compared to HD 97048. While HD 97048 has
roughly 75 channels covering the disk, HD 163296 has
roughly 150. All models applied to HD 163296 must
therefore combine significantly more velocity channels,
which may decrease its ability to extract information
from the images. The activation structure tends to be
less informative, perhaps for the similar reasons, so it is
omitted. Despite this limitation, the high confidence of
some of the models should encourage a human observer
to seriously scrutinize the kinematic structure.
3.4. Limitations and Future Work
Our work comes with several limitations. All simu-
lations were run with SPH, so we did not demonstrate
that our methods would also be applicable to data from
grid-based codes. However, there is typically a qualita-
tive agreement between results from SPH and grid-based
methods (de Val-Borro et al. 2006; Agertz et al. 2007).
A good direction for future work would be to train on
simulation data from a variety of hydrodynamics codes,
and check that the same results can be obtained.
Any differences in performance of the models between
the codes, e.g. if the models consistently performed bet-
ter on SPH data, would be difficult to interpret and the
cause could be hard to identify. It is possible that the
method of domain adaptation (Ben-David et al. 2010),
which has already found success in astronomy (Vilalta
et al. 2019; Alexander et al. 2021; Ćiprijanović et al.
2022), could be used to encourage the models to over-
come any differences between datasets. This is an av-
enue that is ripe for exploration in future work.
In order to keep the parameter space manageable and
demonstrate a proof of concept in this work, we therefore
trained on data produced by only one hydrodynamics
code (PHANTOM).
The absence of true image segmentation (Minaee et al.
2020) or object detection (Zhao et al. 2018) during train-
ing is another limitation that requires addressing in fu-
ture work. While we present results in the form of ac-
tivation structures that suggest the location(s) of the
planet(s), these activations do not always perfectly coin-
cide with the planets, nor capture all planets. User anal-
ysis is required in order to pinpoint the radial (and some-
times azimuthal) location(s) of the planets even with
the information given by the activation structures. The
main utility of the presented models is therefore to give

(a) (b)
(d)
(c)
Figure 3. Confusion matrix from withheld test set for all models. The x axis is the percent of disks in that class that were
predicted accordingly. (a): Using a 50% decision threshold. (b): Using a 75% decision threshold. (c): Using a 90% decision
threshold. (d): Using a 95% decision threshold. The top left and bottom right blocks of each figure show the counts of true
negatives and true positives, respectively. The top right and bottom left blocks of a figure show the false positives and false
negatives, respectively. A perfectly accurate model will have no entries in the top right and bottom left.
Model HD 97048 Softmax HD 163296 Softmax
EffecientNetV2: 47 Channels 85% 37%
RegNet: 47 Channels > 99% 74%
RegNet: 61 Channels 96% > 99%
RegNet: 75 Channels > 99% > 99%
Table 3. Prediction confidences for all models tested on HD 97048 and HD 163296.

8 Terry et al.
Figure 4. Left: Various metrics calculated from the withheld test set for all models. Right: Corresponding ROC curves. Error
bars show 95% confidence intervals that are calculated by bootstrapping each metric 1,000 times using random selections of
80% of the test data. The percentages next to the accuracy labels denote the decision threshold.
Figure 5. Accuracy of all models for a given number of planets. Left: 50% decision threshold. Right: 95% decision threshold

0.5 0.75 0.9 0.95
Decision Threshold
0.0
0.2
0.4
0.6
0.8
1.0
p-value
EfficientNetV2 (47 channels)
RegNet (47 channels)
Figure 6. p-values from testing the correlation between ac-
curacy and the number of planets at different decision thresh-
olds. The black line is p = 0.05, which is the standard cutoff
for statistical significance.
a user an indication as to whether a planet is present and
motivate and direct the analysis of the velocity channels.
Segmentation and object detection would address these
shortcomings by precisely pinpointing the location(s) of
the planet(s) without any necessary further analysis.
The assertion of a vertically isothermal disk is a sim-
plifying assumption that is commonly used in simula-
tions of protoplantary discs (see, e.g., Dipierro et al.
2015; Pinte et al. 2019) . It has been demonstrated that
the dispersion relation and morphology of the planet’s
signature can change significantly if the disk is verti-
cally stratified (Ogilvie & Lubow 2002; Juhász & Rosotti
2018).
However, the presence of any disturbance is more im-
portant than its specific morphology. There does exist
the possibility that vertical stratification dampens the
signatures to an extent that the models’ performances
appreciably decrease, but our results demonstrate that
even extremely subtle signals can be found by the mod-
els. Regardless, a more realistic treatment of the ther-
mal structure may be useful in future works.
4. CONCLUSIONS
Our results show that a machine learning model
trained on synthetic data can determine the presence
and location of forming exoplanets in real telescope ob-
servations of protoplanetary accretion disks. We apply
this model to the HD 97048 and HD 163296 systems.
Multiple models predict the presence of a planet with
greater than 99% confidence in both systems. The pres-
ence and location of the planet in HD 97048 is corrobo-
rated by activating strongly on the same location given
by Pinte et al. (2019).
Five out of six models have greater than 90% accuracy
when using a 50% decision threshold. When using a 95%
decision threshold, four out of six models have an accu-
racy greater than 90%. Five models have an AUC of at
least 0.98. While EfficientNetV2 and RegNet perform
comparably on the synthetic test data, RegNet outper-
forms EfficientNetV2 on the real telescope observations.
This work is the first step in a series of works that use
machine learning to automate the detection and analysis
of objects within protoplanetary disks. Applying full
segmentation and object detection methods will further
increase the ability of our models to locate planets and
infer properties, such as planetary mass. We leave this
work to the future. We envision final models that can
be used easily for multiple purposes and be distributed
freely throughout the community.
As ALMA continues to deliver larger and larger disk
survey datasets and next generation telescopes —such
as JWST, ngVLA and the SKA —come online, the
time-cost of analyzing more and more data will increase.
Our proof-of-concept work demonstrates that even when
high confidence is demanded, machine learning can auto-
mate this task and match and exceed human proficiency.
5. ACKNOWLEDGEMENTS
We would like to thank the referee for in-
sightful comments that improved the paper.
This paper makes use of the following ALMA
data: ADS/JAO.ALMA#2016.1.00826.S and
ADS/JAO.ALMA#2013.1.00601.S. ALMA is a part-
nership of ESO (representing its member states),
NSF (USA) and NINS (Japan), together with NRC
(Canada), MOST and ASIAA (Taiwan), and KASI
(Republic of Korea), in cooperation with the Republic
of Chile. The Joint ALMA Observatory is operated by
ESO, AUI/NRAO and NAOJ. The National Radio As-
tronomy Observatory is a facility of the National Science
Foundation operated under cooperative agreement by
Associated Universities, Inc. SPH results are visualized
using SPLASH (Price 2007). JT was a participant in the
2022 Google Summer of Code program. This study was
supported in part by resources and technical expertise
from the Georgia Advanced Computing Resource Cen-
ter, a partnership between the University of Georgia’s
Office of the Vice President for Research and Office of
the Vice President for Information Technology.

10 Terry et al.
-1
-0.9
-0.8
log
column
density
[g/cm
2
]
50 AU
(a) (b)
(c) (d)
Activations Gas Density
Raw Channel Convolved Channel
Figure 7. (a): Mean-subtracted activations using the 47-channel EfficientNetV2. The innermost planet has M=0.28MJ, the
second planet has M=0.20MJ, the third planet has M=0.36MJ, and the outermost planet has has M=0.02MJ. (b): Disk column
density after an azimuthal rotation of φ = 131◦
and inclined by i = 36◦
. (c): Unconvolved velocity channel with the most
obvious kink. (d) Corresponding convolved velocity channel. Based on the kink, the approximate location of the planet causing
the most significant kink is in the white circle. The planets were correctly predicted with a confidence of > 99%.
REFERENCES
Agertz, O., Moore, B., Stadel, J., et al. 2007, MNRAS, 380,
963, doi: 10.1111/j.1365-2966.2007.12183.x
Alexander, S., Gleyzer, S., McDonough, E., Toomey,
M. W., & Usai, E. 2020, ApJ, 893, 15,
doi: 10.3847/1538-4357/ab7925
Alexander, S., Gleyzer, S., Reddy, P., Tidball, M., &
Toomey, M. W. 2021, arXiv e-prints, arXiv:2112.12121.
https://arxiv.org/abs/2112.12121
Andrews, S. M., Wilner, D. J., Hughes, A. M., Qi, C., &
Dullemond, C. P. 2009, The Astrophysical Journal, 700,
1502, doi: 10.1088/0004-637x/700/2/1502
Andrews, S. M., Huang, J., Pérez, L. M., et al. 2018, ApJL,
869, L41, doi: 10.3847/2041-8213/aaf741
Auddy, S., Dey, R., Lin, M.-K., Carrera, D., & Simon, J. B.
2022, ApJ, 936, 93, doi: 10.3847/1538-4357/ac7a3c

Example with no planet, 811, 646
-1.1
-1
-0.9
log
column
density
[g/cm
2
]
50 AU
(a) (b)
Activations Gas Density
Figure 8. (a): Typical example of mean-subtracted activa-
tions for a disk with no planet. (b): Disk column density
after an azimuthal rotation of φ = 153◦
and inclined by
i = 13◦
.
Auddy, S., Dey, R., Lin, M.-K., & Hall, C. 2021, ApJ, 920,
3, doi: 10.3847/1538-4357/ac1518
Barraza-Alfaro, M., Flock, M., Marino, S., & Pérez, S. 2021,
A&A, 653, A113, doi: 10.1051/0004-6361/202140535
Ben-David, S., Blitzer, J., Crammer, K., et al. 2010,
Machine Learning, 79, 151,
doi: 10.1007/s10994-009-5152-4
Biewald, L. 2020, Experiment Tracking with Weights and
Biases. https://www.wandb.com/
Ćiprijanović, A., Kafkes, D., Snyder, G., et al. 2022,
Machine Learning: Science and Technology, 3, 035007,
doi: 10.1088/2632-2153/ac7f1a
Cireşan, D. C., Meier, U., Masci, J., Gambardella, L. M., &
Schmidhuber, J. 2011, arXiv e-prints, arXiv:1102.0183.
de Val-Borro, M., Edgar, R. G., Artymowicz, P., et al. 2006,
MNRAS, 370, 529, doi: 10.1111/j.1365-2966.2006.10488.x
Dipierro, G., Price, D., Laibe, G., et al. 2015, MNRAS, 453,
L73, doi: 10.1093/mnrasl/slv105
Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. 2020,
arXiv e-prints, arXiv:2010.11929.
Draine, B. T., & Lee, H. M. 1984, ApJ, 285, 89,
doi: 10.1086/162480
Elfwing, S., Uchibe, E., & Doya, K. 2017, arXiv e-prints,
arXiv:1702.03118. https://arxiv.org/abs/1702.03118
Glorot, X., Bordes, A., & Bengio, Y. 2011, in AISTATS
Goodfellow, I., Bengio, Y., & Courville, A. 2016, Deep
Learning (MIT Press)
Hall, C., Dong, R., Teague, R., et al. 2020, ApJ, 904, 148,
doi: 10.3847/1538-4357/abac17
Hanley, J. A., & McNeil, B. J. 1982, Radiology, 143 1, 29
He, K., Zhang, X., Ren, S., & Sun, J. 2015, arXiv e-prints,
Hinton, G. E., Dayan, P., Frey, B. J., & Neal, R. M. 1995,
Science, 268, 1158. http://www.jstor.org/stable/2888376
Huang, J., Andrews, S. M., Dullemond, C. P., et al. 2018,
ApJL, 869, L42, doi: 10.3847/2041-8213/aaf740
Jo, Y., & Kim, J.-h. 2019, MNRAS, 489, 3565,
doi: 10.1093/mnras/stz2304
Juhász, A., & Rosotti, G. P. 2018, MNRAS, 474, L32,
doi: 10.1093/mnrasl/slx182
Kingma, D. P., & Ba, J. 2014, arXiv e-prints,
McKay, M. D., Beckman, R. J., & Conover, W. J. 1979,
Technometrics, 21, 239.
http://www.jstor.org/stable/1268522
Minaee, S., Boykov, Y., Porikli, F., et al. 2020, arXiv
e-prints, arXiv:2001.05566.
Möller, A., & de Boissière, T. 2020, MNRAS, 491, 4277,
doi: 10.1093/mnras/stz3312
Öberg, K. I., Guzmán, V. V., Walsh, C., et al. 2021, ApJS,
257, 1, doi: 10.3847/1538-4365/ac1432
Ogilvie, G. I., & Lubow, S. H. 2002, MNRAS, 330, 950,
doi: 10.1046/j.1365-8711.2002.05148.x
Paszke, A., Gross, S., Massa, F., et al. 2019, arXiv e-prints,
Pinte, C., Harries, T. J., Min, M., et al. 2009, A&A, 498,
967, doi: 10.1051/0004-6361/200811555
Pinte, C., Ménard, F., Duchêne, G., & Bastien, P. 2006,
A&A, 459, 797, doi: 10.1051/0004-6361:20053275
Pinte, C., Price, D. J., Ménard, F., et al. 2018, ApJL, 860,
L13, doi: 10.3847/2041-8213/aac6dc
Pinte, C., van der Plas, G., Ménard, F., et al. 2019, Nature
Astronomy, 3, 1109, doi: 10.1038/s41550-019-0852-6
Pinte, C., Price, D. J., Ménard, F., et al. 2020, ApJL, 890,
L9, doi: 10.3847/2041-8213/ab6dda
Price, D. J. 2007, PASA, 24, 159, doi: 10.1071/AS07022
Price, D. J., Wurster, J., Tricco, T. S., et al. 2018, PASA,
35, e031, doi: 10.1017/pasa.2018.25
Siess, L., Dufour, E., & Forestini, M. 2000, A&A, 358, 593.
https://arxiv.org/abs/astro-ph/0003477
Simon, J. B., Hughes, A. M., Flaherty, K. M., Bai, X.-N.,
& Armitage, P. J. 2015, ApJ, 808, 180,
doi: 10.1088/0004-637X/808/2/180
Szegedy, C., Liu, W., Jia, Y., et al. 2014, arXiv e-prints,
Tan, M., & Le, Q. V. 2019, arXiv e-prints,
—. 2021, arXiv e-prints, arXiv:2104.00298.
Teague, R., Bae, J., Bergin, E. A., Birnstiel, T., &
Foreman-Mackey, D. 2018, ApJL, 860, L12,
doi: 10.3847/2041-8213/aac6d7

12 Terry et al.
885 µm ∆v = 0.96 km/s
Activation 1 Activation 2
(a) (b)
(c) (d)
Figure 9. (a): HD 97048 at 885 µm. (b): ∆v = 0.96 km/s channel. Bottom: Mean-subtracted activations from two different
layers for HD 97048 using the 47-channel RegNet. The estimated planet location is given inside the white circle. The beams
for the observations are given in the bottom right corner.
Terry, J. P., Hall, C., Longarini, C., et al. 2022, MNRAS,
510, 1671, doi: 10.1093/mnras/stab3513
Vaswani, A., Shazeer, N., Parmar, N., et al. 2017, arXiv
Vilalta, R., Dhar Gupta, K., Boumber, D., & Meskhi,
M. M. 2019, PASP, 131, 108008,
doi: 10.1088/1538-3873/aaf1fc
Xu, J., Pan, Y., Pan, X., et al. 2021, arXiv e-prints,
Zhao, Z.-Q., Zheng, P., Xu, S.-t., & Wu, X. 2018, arXiv
Zhou, W., Yang, Y., Yu, C., et al. 2021, Nature
Communications, 12, 1259,
doi: 10.1038/s41467-021-21466-z

Locating Hidden Exoplanets in ALMA Data Using Machine Learning

More Related Content

Similar to Locating Hidden Exoplanets in ALMA Data Using Machine Learning (20)

More from Sérgio Sacani (20)

Recently uploaded (20)

Locating Hidden Exoplanets in ALMA Data Using Machine Learning