Machine Learning Transforms Exo-Planetary Science Workflows

A multi-author review led by Jeanne Davoult and collaborators from NCCR PlanetS consolidates how Machine Learning is reshaping planetary and exoplanetary science. The chapter synthesizes practical methods and case studies that address data volume, heterogeneity, and the inverse problems common to planetary research, framing ML as a transformative toolkit for both observational and simulation pipelines.

The review organizes the field around three core methodological pillars and provides concrete examples and algorithmic choices:

• •Sequence modelling for one-dimensional signals, including radial velocity time series and photometric light curves, with architectures that combine temporal convolution, recurrent elements, and attention mechanisms to recover periodicities and transit signals.
• •Pattern recognition and unsupervised discovery, illustrating convolutional neural networks for feature extraction, mapping and cross-correlation workflows, VAE-style anomaly detection, and clustering pipelines applied to mass spectrometry and spectral datasets.
• •Generative modelling and emulation-based Bayesian analysis, showing how DNNs and surrogate models accelerate forward model evaluation, enable differentiable emulators for planetary interior structure inference, and integrate with Bayesian posteriors to explore formation scenarios.

The review is practical and method-centric rather than purely theoretical. It highlights how ML reduces computational cost for expensive numerical models, increases sensitivity to weak signals in noisy telescopic data, and enables modular pipelines that combine supervised, unsupervised, and generative approaches. This matches broader trends where domain scientists use ML as an accelerator for simulation-heavy disciplines rather than as a black-box replacement. The emphasis on emulators and Bayesian coupling is important: it connects ML advances with the statistical rigor required for parameter inference in planetary science.

Expect growing adoption of ML-based emulators in mission pipelines and more cross-disciplinary toolkits that standardize preprocessing for heterogeneous, multi-instrument datasets. Key open questions remain around uncertainty calibration, model interpretability for physical inference, and reproducible benchmarks that compare ML surrogates to classical methods.