Jeanne Davoult, Valentin T. Bickel, Caroline Haslebacher, Yann Alibert, Daniel Angerhausen, Carles Cantero, Jo Ann Egger, Romain Eltschinger, Yannick Eyholzer, Emily O. Garvin, Salome Gruchola, Adrien Leleu, Sara Marques, Yinan Zhao
10 min
Updated to integrate observational and modeling perspectives from exoplanet atmosphere characterization. This NCCR PlanetS review showcases ML innovations in sequence modeling, pattern recognition, and generative emulation across (exo)planetary science, automating analysis of massive datasets from JWST, rovers, and transit surveys. Complements traditional computational modeling by accelerating spectral retrieval and 3D GCMs.
ML processes heterogeneous data from missions like JWST, HST, TESS, and Kepler, enabling rapid exoplanet detection and atmospheric retrievals from transmission/emission spectra and phase-curves. Automates handling of diverse populations (hot Jupiters to super-Earths), addressing cloud obscuration and C/O ratio uncertainties.
Generative ML emulates self-consistent models (hydrodynamics, radiative transfer, cloud formation), reducing computational costs for 3D simulations like those on K2-18b. Pattern recognition validates brown dwarf analogs and formation diagnostics, bridging observations with physics-based predictions.
The exploration of planetary bodies in our Solar system and beyond relies on the processing and interpretation of large, spatio-temporally inconsistent, and heterogeneous datasets. Recent advances in machine learning (ML) provide unprecedented opportunities to address many fundamental challenges posed by these heterogeneous and hyper-dimensional datasets. This review chapter highlights innovative ML methodologies that were developed and used by NCCR PlanetS members to address three overarching challenges in (exo)planetary science. The first challenge is sequence modelling, which encompasses the intricate analysis of one-dimensional data such as time series of radial velocities and light curves, among other examples. Secondly, there is pattern recognition that involves studying correlations, leveraging convolutional neural networks for feature extraction, mapping and cross correlation among other examples., anomaly detection through variational autoencoders, and unsupervised clustering of mass spectrometric data. Lastly, there are generative models and emulation-based Bayesian analysis, which encompass the development of predictive models for planetary interior structure, employing Deep Neural Networks to understand planet formation mechanisms. These innovative ML methodologies herald a paradigm shift in the processing of data and numerical models that represent inherent challenges in planetary and exoplanetary science, paving the way for revolutionary discoveries and ideas in this field.
Alex: Wait—so it predicts the noise from the raw spectrum shapes themselves? That sounds like it keeps more detail than the usual way.
Sam: Yes. The standard method averages many lines in the spectrum to pull out one wobble number, using something called a cross-correlation function. But that averaging smooths away the subtle changes from star activity, like how blending fruits loses the taste of each one. By working at the full spectrum level first, the CNN captures those changes precisely—widths of lines, their slopes—and links them to fake wobbles.
Alex: Okay, so subtract the predicted star noise, and planets pop out clearer. They tested this on actual stars?
Sam: They applied it to data from Alpha Centauri B, Tau Ceti, and our Sun. On solar spectra with plenty of data, it reached a wobble sensitivity of 0.2 meters per second—about twice as precise as prior limits—enough to spot a planet twice Earth's mass in a habitable orbit. The paper suggests this works because spectrum details preserve activity clues lost in averaging.
Alex: Huh. That's a meaningful step for finding nearby Earth-like worlds.
Alex: So sequence modeling helps clean up radial velocity data. What about direct imaging—taking actual pictures of exoplanets around other stars?
Sam: Direct imaging faces its own hurdles. Stars are millions of times brighter than their planets, so even with tricks like blocking the star's light, leftover glow creates speckles—tiny bright spots from telescope imperfections and air turbulence. These speckles look just like faint planets, fooling detectors. One strategy captures a stack of images while the telescope stays fixed but Earth spins, rotating the planet's position against those static speckles. They call this angular differential imaging, or ADI.
Alex: Like watching a merry-go-round from a stationary spot—the background blurs but fixed rides stay put?
Sam: Precisely. Subtracting rotated images cancels speckles but keeps the moving planet signal. Still, standard cleanup uses principal component analysis—it builds a model of speckles from the images themselves and subtracts it. But that can smear the planet a bit and misses tricky patterns. PlanetS researchers created NA-SODINN, a system trained on labeled examples of real speckle noise versus planets. It scans ADI-processed stacks and flags true planets better than other tools.
Alex: Trained to spot the difference in those cleaned-up image sequences. How much better does it perform?
Sam: In challenges like the Exoplanet Imaging Data Challenge, it consistently beat top algorithms across different telescopes. They re-ran data from 150 stars observed with a European telescope called SPHERE, uncovering more planet candidates. The paper notes this fills gaps between radial velocity finds and imaging, especially for wider orbits.
Alex: Those wide-orbit giants sound promising. But the review also covers mapping features on planets closer to home, right—like craters or cracks on Mars or moons?
Sam: Yes. On Mars, researchers used systems like convolutional neural networks to scan decades of orbital images. These spotted hundreds of new impact craters—pits from meteor hits—linking them to earthquake data from NASA's InSight lander. That raised the known impact rate about twice as high as before, since manual checks missed many.
Alex: So automating that catches things humans overlook in huge image stacks. What about trickier features, like thin lines across icy surfaces?
Sam: Thin lines, or lineaments, on Europa challenge detectors because they're narrow—often just a few pixels wide—and form dense networks over big areas. One tool breaks images into regions, outlines each line, and sorts them by type, like ridges or cracks. It started with a supervised setup that learns from labeled examples, then improved with more data. The same approach worked on Venus, Ganymede, and Enceladus without retraining, showing it transfers across worlds.
Alex: Outlining narrow lines in crowded images without smearing them—and it works on multiple worlds.
Sam: Exactly. With upcoming missions bringing 500 times more Europa images, this prepares for quick global mapping of how those lines formed.
Alex: Huh. Grouping raw data by hidden similarities. How does that play out in other data, say chemical analysis?
Sam: In mass spectrometry—where lasers zap samples to reveal chemical makeup by mass patterns—unsupervised tools simplify high-dimensional readouts. First, they cut noise and scale the data. Then, they squeeze it into fewer dimensions, keeping key differences, like summarizing a book's themes without losing the plot. Clustering then groups similar patterns, spotting minerals or fossils automatically. This runs onboard spacecraft, prioritizing what to send home.
Alex: So for future landers, it flags interesting rocks in real time without Earth waiting.
Sam: Precisely. The review stresses keeping humans in the loop for checks, but automation handles the volume from missions like Europa Clipper.
Alex: Automation like that makes sense for the data flood. Does the review touch on using these tools to speed up the heavy simulations for planet formation or interiors?
Sam: Yes. Complex physics models for things like planet interiors or full system formation take days to run because they solve detailed equations step by step. Researchers train neural networks on millions of those slow runs to mimic the outputs quickly—like a shortcut recipe that skips the full cooking time but gets the same taste. They call this surrogate modeling or emulation. One example, plaNETic, predicts a planet's radius from its likely inner layers, trained on data from a code called BICEPS. Traditional ways to guess interiors use sampling methods that chain thousands of model runs, taking hours or days. With plaNETic, it finishes in under 30 minutes—used in over 30 studies on real exoplanets.
Alex: Huh. Cutting computation by orders of magnitude lets them explore more possibilities.
Alex: But there must be limits—like needing good training data?
Sam: That's a key point. Data is often scarce, noisy, or instrument-limited, so models risk overfitting—memorizing training examples but stumbling on new ones. They're somewhat black-box, harder to interpret than traditional stats, and struggle with uncertainties. Synthetic data helps but may not match reality perfectly. The paper calls for more interpretable models and real data from upcoming missions like PLATO or ARIEL.
Alex: Right, and keeping humans overseeing to catch biases.
Sam: Precisely. Still, it enables real-time decisions, better JWST targets, and steps toward Earth-like detections in habitable zones. This review shows machine learning as a meaningful tool for (exo)planetary science—balancing speed with rigor.
Alex: That's a clear synthesis, Sam. Thanks for breaking it down.
Sam: My pleasure. It points to careful progress ahead. Thanks for listening to ResearchPod.