Exoplanet Hunter
Classification of exoplanet candidates
From NASA observations to exoplanet candidates, understand data and compare models. From XGBoost to ExoMiner, then to a Transformer from scratch at the heart of my thesis subject.

- 01
Observe the stars
Measure their brightness over time.
- 02
Spot a signal
A drop in light may indicate a passage.
- 03
Sort the candidates
Distinguish between possible planets and false positives.
Spotting planets around other stars
An exoplanet is a planet located outside our Solar System. The challenge of the NASA Space Apps Challenge 2025 is to leverage open observations and AI to help identify them. Faced with thousands of signals, it is necessary to direct the analysis towards the most credible candidates.
Our response: help vet candidates
Exoplanet Hunter allows you to import observations, run a classification and inspect the results. To understand how the model does this sorting, you first need to understand what telescopes measure: starlight.
- 01
The curve
When and how light varies.
- 02
Measurements
Period, duration and depth of the passage.
- 03
Known status
An example to learn how to classify.
From measured light to catalogs
Kepler, K2 and TESS measure the brightness of stars over time: it is a light curve. A planet passing in front of its star can produce a repeated dip, called a transit. But another star or an instrumental defect can create a similar signal.
Two ways of describing the same observation
The curve maintains the shape of the signal. The catalog summarizes its properties: period between two passages, duration, depth and characteristics of the star. Each line describes an object or a candidate, sometimes with a status already assigned. These tables constitute the starting point of the first model.
- 01
Identify
Recognize the format and the target to predict.
- 02
Prepare
Select measurements and process zero values.
- 03
Learn
Link these measurements to known statuses.
Make measurements usable
Before learning, you need to know which columns describe the signal and which values are missing. The missions use different names and statuses. The application recognizes the format, locates the target column and selects the numerical measurements available for this catalog.
Process absences without losing all observations
The code removes entirely empty columns, the object’s name and the status to predict. It fills missing values with their column mean, falling back to zero. This retains incomplete rows but may smooth out their distinctive features: that choice also needs evaluation.
- 01
Catalog measurements
Describe each candidate using numbers.
- 02
Successive trees
Gradually correct classification errors.
- 03
A first reference
Measure what tables alone allow.
Start with a tabular baseline
The data are now measurement lines. XGBoost is suitable for this representation: it learns non-linear relationships between variables, without requiring a complete curve as input. This first model establishes a reference to then judge the interest of a more complex analysis.
Combine small rules to improve sorting
Each tree splits examples using thresholds on the measurements. Subsequent trees correct the previous ensemble’s errors. From known statuses, the model learns to assign a score to each class. We still need to check whether it succeeds on examples it has not seen during training.
- 01
Learn
Fit the trees on the training set.
- 02
Test
Compare predictions to set aside statuses.
- 03
Examine errors
Identify information that might be missing.
Understand errors before making them more complicated
A model can memorize examples well without knowing how to generalize. The pipeline reserves by default 20% of lines for testing, maintaining the proportion of classes. The confusion matrix then shows which statuses are confused, beyond just the rate of correct answers.
Measure what is retained and what is missed
Precision indicates how reliable a selection is; recall measures the proportion of sought cases that were found. The prototype computes these metrics but fills missing values before splitting the data. A stricter evaluation must estimate these means from training data alone and keep stars separate across the datasets.
- 01
The digital summary
Compact, but limited to the selected measures.
- 02
The shape of the signal
Details and repetitions to examine.
- 03
The rest of the study
Prepare the curves before modeling them.
What the tables don't show
XGBoost only sees the selected metrics. Two events can have a similar duration and depth, but different forms. Examining the curve allows you to look for these differences: is the dip regular, is it repeated, is there another drop between two passages?
Give models more complete information
The following therefore explores the form of the signal, in addition to its summaries. This motivates the use of specialized models like ExoMiner and the Transformer track. Before explaining them, we must prepare this new entry: a series of noisy and sometimes interrupted measurements.
- 01
Filter metrics
Discard invalid values and check anomalies.
- 02
Remove trend
Reduce slow variations in the light background.
- 03
Normalise the flux
Compare shapes on a common scale.
Bring out the transit without erasing it
Brightness may drift slowly with the activity of the star or instrument. Outliers can also mimic an event. Cleaning seeks to reduce these disturbances so that the model learns the shape of the transit, rather than measurement defects.
Filter, remove trend, normalize
The preparation studied removes invalid measurements, estimates a smooth trend, then puts flux on a comparable scale. Filtering must preserve plausible dips: removing every sharp drop could erase a planet. The corrected curve can then be aligned with repeated transits.
- 01
A candidate period
Estimate when the phenomenon repeats.
- 02
A global view
Observe the cycle and other variations.
- 03
A local view
Carefully examine the shape of the passage.
Compare passages on two scales
Once the signal is prepared, we use a candidate period to superimpose the passages: this is phase folding. For example, if an event recurs every ten days, the measurements spaced ten days apart are aligned. A repeated pattern becomes easier to examine.
Keep both context and the detail of the dip
The overall view shows the entire cycle and possible secondary declines; the local view zooms around the transit. A BLS search can estimate the period by testing for repeated troughs. An incorrect period confuses the alignment: preparing the views is therefore already part of the experiment.
- 01
Multiple diagnoses
Examine transit and false positive indices.
- 02
Specialized branches
Learn a representation of each diagnosis.
- 03
A combined score
Gather the clues to classify the candidate.
Combine clues with a specialised model
These views explain the interest of ExoMiner: an isolated dip is not enough to conclude. This specialized network combines several diagnoses inspired by scientific examination, such as the shape of the transit, secondary declines and the differences between even and odd passages.
Integrate an existing reference into the application
The project launches the NASA pipeline from TESS identifiers and observation sectors, then returns its scores. ExoMiner thus provides specialized analysis in addition to XGBoost. Its principle of crossing clues leads to the thesis question: can we learn their relationships with a Transformer?
- 01
Represent
Signal views and star context.
- 02
Connect by attention
Learn what clues illuminate others.
- 03
Aggregate with CLS
Form a representation of the candidate.
Architecture proposed in the research report; its performance remains to be measured.
Learn the relationships between clues
Following on from ExoMiner, the initial idea of my thesis is to build a Transformer from scratch. The goal is to relate the local shape of the transit to the full cycle and properties of the star, rather than interpreting each clue separately.
Attention connects information
The signal is divided into digital elements, tokens. STAR and PHYS add stellar and physical context. The Q, K and V projections compare these elements, weight their relationships and transmit useful information. The CLS token gathers the result to classify the candidate.
- 01
Transform
A learned representation of relationships.
- 02
TSFRESH
Explicit measurements of the time series.
- 03
Late merger
Attach both summaries before filing.
Add a statistical signal reading
Attention learns its own representations. Can explicit statistics complete it? TSFRESH calculates, for example, variability and autocorrelations: the latter indicate how closely a signal resembles a lagged version of itself.
Bring two representations together before the decision
The report proposes selecting these descriptors, then projecting them through a small network, the MLP. Its output is joined to the Transformer’s CLS summary: late fusion. Adding this branch guarantees no improvement; the model must be compared with and without it.
- 01
Fix the protocol
Same candidates and separation by star.
- 02
Remove a component
Compare with and without each source of information.
- 03
Measuring intake
Relate gains to errors and cost.
Check what each addition brings
The journey led from tabular measurements to ExoMiner diagnostics, then to a research architecture. To find out if this complexity is useful, the report provides comparisons on common data and the removal of one branch at a time: an ablation study.
Gains to demonstrate on new stars
Separating the stars prevents close observations from being found in both training and testing. We then compare precision, recall and calculation cost. Removing TSFRESH or the physical context allows us to isolate their contribution. These experiments are a proposed protocol, not results already obtained.
- 01
Prioritize
Retain candidates for review.
- 02
Check the clues
Return to curves and ambiguous cases.
- 03
Continue analysis
Compare the prediction with observations.
Return to the candidate behind the score
Even a well-evaluated model produces errors. To meet the initial challenge, its scores must help choose the observations to examine. The application allows you to consult the selected candidates and export a selection for further analysis.
Automatic sorting for scientific examination
The approaches presented classify events already proposed upstream; they may miss those that this first search did not identify. A high score therefore does not, by itself, confirm a planet. We must return to the signal, diagnostics and additional observations.
The journey in Exoplanet Hunter
Discover Exoplanet Hunter in this video presentation of the project.
The app in action
An overview of the project interface, to connect the previous explanations to its use.