AI is helping astronomers discover new planets

BLOG: Heidelberg Laureate Forum

Laureates of mathematics and computer science meet the next generation
Heidelberg Laureate Forum

This article is part of an ongoing series of articles on how Artificial Intelligence is currently being used to make the world a better place. For previous stories, check out how AI is looking for delicious meat alternatives and how algorithms can partner up with human intuition to make fundamental discoveries.

Scouring through data from the Kepler telescope, a deep neural network called ExoMiner has discovered 301 new exoplanets, hinting that algorithms could discover plenty more planets in the near future.

Artistic depiction of exoplanets. Image credits: NASA / ESA.

The first exoplanet (a planet outside our solar system) was discovered in 1992. Just a few more were discovered until the year 2000. By 2010, still, fewer than 500 exoplanets were known. Now, in 2022, we know 4,905 — and out of them, 2,662 were discovered by the Kepler Space telescope. In fact, Kepler may have already discovered many more exoplanets that haven’t been confirmed yet.

Kepler and Transiting Exoplanet Survey Satellite (another exoplanet-searching mission) have discovered over 100,000 potential signals that could come from exoplanets. But in order to be certain, the data needs to be processed and analyzed. This is where ExoMiner comes in.

ExoMiner is a new deep neural network (a type of Artificial Intelligence) that also uses NASA’s Pleiades supercomputer, one of the world’s most powerful supercomputers, with specialized pipelines developed by the space agency. ExoMiner is essentially a robust classifier — an algorithm that categorizes data into “classes”. What’s most impressive about it is its ability to distinguish actual exoplanets from false positives. To do this, it draws not only from various tests and properties that human astronomers use to confirm exoplanets, but also from its past mistakes.

It’s also got one particular advantage compared to other neural networks: researchers can understand how it works.

“Unlike other exoplanet-detecting machine learning programs, ExoMiner isn’t a black box – there is no mystery as to why it decides something is a planet or not,” said Jon Jenkins, exoplanet scientist at NASA’s Ames Research Center in California’s Silicon Valley, in a statement. “We can easily explain which features in the data lead ExoMiner to reject or confirm a planet.” Oftentimes, even the researchers using AI aren’t sure why it does what it does.

“When ExoMiner says something is a planet, you can be sure it’s a planet,” added Hamed Valizadegan, ExoMiner project lead and machine learning manager with the Universities Space Research Association at Ames. “ExoMiner is highly accurate and in some ways more reliable than both existing machine classifiers and the human experts it’s meant to emulate because of the biases that come with human labeling.”

In order to understand how ExoMiner analyzes planet candidates, we first need to look at how exoplanets are discovered. Several methods are used, but by far the most common is the transit method. Basically, we can’t really “see” planets because they don’t emit light. But we can see stars, and we can measure the light that comes from stars — and here lies the key of the transit method.

Let’s say you’re measuring the light from a star, and there’s also an exoplanet orbiting around that star. When the exoplanet passes between the star and your observation point, it obstructs a small part of the light coming from the star. This dip in brightness can be used to detect exoplanets.

The transit method: A depiction of how the dip in brightness can be used to detect planets. Image credits: NASA.

But not any dip is indicative of an exoplanet. ExoMiner analyzes this transit data, attempts to filter and process the data, and then classifies the signal as an exoplanet or not-exoplanet.

Example of transit data analyzed by ExoMiner. Image from Valizadegan et al (2022).

So far, the neural network was applied to Kepler data, and in a new study published in arXiv, a team of researchers presented the discovery of 301 new exoplanets. None of them appear to lie in the habitable zone or to be Earth-like, but they could help us better understand our corner of the universe.

“These 301 discoveries help us better understand planets and solar systems beyond our own, and what makes ours so unique,” said Jenkins. 

In addition, the approach could lead to much more exoplanet discoveries. Furthermore, researchers say, as ExoMiner is fine-tuned even more, it could also have a go at data from TESS or the European Space Agency’s upcoming PLAnetary Transits and Oscillations of stars (or PLATO) mission.

“Now that we’ve trained ExoMiner using Kepler data, with a little fine-tuning, we can transfer that learning to other missions, including TESS, which we’re currently working on,” said Valizadegan. “There’s room to grow.”

It’s not the first time an AI has been used to detect exoplanets. In 2017, NASA announced a collaboration to use AI to find exoplanets; they even hosted an “Ask Me Anything” session about the first exoplanets discovered thusly. In 2019, a team led by an undergraduate student used AI to discover two new exoplanets. Now, it’s become apparent that the method is scalable and reliable — in the test set that NASA used, ExoMiner discovered 93.6% of all exoplanets, while the best existing classifier had a success rate of 76.3%.

Astronomers are increasingly looking at Big Data and Artificial Intelligence to help them make sense of the universe. A problem that modern astronomy is faced with is that sometimes, there’s just too much data to analyze. It’s a good problem to have, but it’s a problem nonetheless. Specialized algorithms looking for patterns in this data can be of great help to astronomers, and it’s not just for detecting exoplanets.

Astronomers have been using AI to find gravitational waves, gravitational lenses, or to calibrate instruments. A separate team has even used AI to detect the fingerprints of different molecules on exoplanets, and we’re probably just starting to scratch the surface of how AI can be used in astronomy.

Discovering exoplanets is a particularly important and exciting application of this type of technology. The number of exoplanets known to humanity is growing at an accelerated pace, and now we have a new tool that can be used to detect even more planets, accelerating our progress in exploring the universe. Who knows what lies out there? Maybe somewhere, there’s another AI just discovering Earth.

Avatar photo

Posted by

Andrei is a science communicator and a PhD candidate in geophysics. He is the co-founder of ZME Science, where he published over 2,000 articles. Andrei tries to blend two things he loves (science and good stories) to make the world a better place -- one article at a time.


  1. Significant work has gone into perfecting the ExoMiner expert system, which uses a variety of deep neural networks for subtasks pertinent to the overall task of determining whether astronomical data potentially related to an exoplanet transit meets the many criteria used by experts in the field.
    So ExoMiner is not a single neural network that works end-to-end (data in, result out, inner workings unknown), but consists of several diligently trained neural networks and for the final decision uses a neural network that uses the results of the DNNs of subnets used. I therefore call ExoMiner an expert system based on deep learning subsystems. Because of this multifactorial approach, ExoMiner can even tell how its decision comes about, since ExoMiner keeps an eye on the DNN subsystems involved.
    In discussing ExoMiner in the arxiv article ExoMiner:
    A Highly Accurate and Explainable Deep Learning Classifier that Validates 301 New Exoplanets
    we read the following sentences, which clearly demonstrate the expert nature of the ExoMiner approach:

    The key to ExoMiner’s performance is its architecture, which mimics the process by which domain experts discover transit signals by examining multiple examine types of diagnostic tests in the form of scalar values ​​and time series data. This architecture also allowed us to design a preliminary explainability framework for branch closures that provides interpretability in terms of which diagnostic tests ExoMiner uses to verify a signal.

    Conclusion: ExoMiner is an example of an expert system based on multiple neural networks. It can explain its decisions because it knows the influence of the subsystems on the overall result.
    Attention: The ExoMiner approach means considerable work for the AI ​​and for domain experts. This approach is therefore not cheap and not suitable for areas with less available domain knowledge.

    • Addendum: ExoMiner is most comparable to a Mixture of Experts system. Because the opinion/assessment of the individual AI experts is ultimately processed by a toplevel expert to form an overall judgment – is it an exoplanet or is it not – one could speak of a hierarchical mixture of experts system. In Wikipedia you can read about it:

      Mixture of experts (MoE) refers to a machine learning technique where multiple expert networks (learners) are used to divide a problem space into homogeneous regions.[1] It differs from ensemble techniques in that typically only a few or 1 expert model will be run, rather than combining results from all models.

      An example from computer vision is combining one neural network model for human detection with another for pose estimation.

      Hierarchical mixture
      If the output is conditioned on multiple levels of (probabilistic) gating functions, the mixture is called a hierarchical mixture of experts.[2]

      A gating network decides which expert to use for each input region. Learning thus consists of learning the parameters of:

      individual learners and
      gating network.

      Classification: ExoMiner is an expert system which bases its conclusion on a Mixture of Experts, where each expert is a DNN.

Leave a Reply

E-Mail-Benachrichtigung bei weiteren Kommentaren.
-- Auch möglich: Abo ohne Kommentar. +