To find an Earth-like planet around a Sun-like star by the Doppler method, astronomers must measure a stellar wobble of about 10 cm/s (On the Feasibility of Intense Radial Velocity Surveys for Earth-Twin Discoveries). The star's own surface activity, meaning dark spots, bright magnetic regions and convection, produces apparent velocity signals ten to a hundred times larger (Workshop Summary: Exoplanet Orbits and Dynamics). There is one physical handle on this problem. A planet slides every spectral line rigidly sideways without changing its shape, while activity changes the shape of the averaged line profile, called the cross-correlation function (CCF). We propose to train a neural encoder that is blind to rigid shifts by construction, so it cannot see the planet. It learns, without labels, a compact description of line-profile shape over time. A small time-adaptive head then maps that description to the activity part of the measured velocity. Whatever velocity the shape cannot explain is left for an explicit planetary orbit model.
The training recipe comes from neural recording, a field that faces the same failure mode. In low signal-to-noise EEG, models asked to reconstruct raw waveforms spend their capacity on noise. Models asked to predict an abstract summary of masked, contiguous stretches of signal learn the underlying dynamics instead (EEG2Rep). Stellar line profiles at realistic photon noise are the same kind of problem: the useful signal is a tiny, slowly evolving distortion buried in pixel noise. The new element is the combination of a translation-blind input, which structurally protects the planet signal, with EEG2Rep-style masked latent prediction and a mapping from shape to velocity that is allowed to drift over time.
Core Claim
The claim is that a self-supervised, shift-blind representation of CCF time series, paired with a mapping from shape to velocity that is re-estimated over local time windows, removes more activity velocity than the current best 2D physical model. It should do so while scaling linearly with the number of observations and without absorbing injected planets. Three bodies of prior work set the bar and the constraints.
The first is the 2D line-profile Gaussian process of Yu et al., known as CCF-GP. A Gaussian process is a flexible statistical model of correlated noise. CCF-GP decomposes profile changes into fixed velocity-dependent basis shapes times latent time-series coefficients. It reduced HARPS-N solar velocity residuals from 1.64 to 0.44 m/s. It also separated a planet from activity at an identical 25-day period, a case where the 1D multi-indicator approach of Rajpaul et al. failed. That result motivates us, because it shows the profile holds information that scalar summaries lose. CCF-GP also defines our target. Its coefficients are restricted to a latent process and its time derivative, its kernel is stationary over the whole dataset, and its cost grows cubically with the number of epochs.
The second is the change-point method of Simola et al.. It showed that the correlation between velocity and activity indicators changes magnitude and even sign as active regions evolve. Refitting that correlation separately in each detected segment lowered the median detection threshold on α Cen B from 7.55 to 2.0 m/s. This constrains our design: any fixed mapping from shape to velocity is wrong for real stars. Simola's method, however, works only on scalar indicators. We keep its time-local refitting and apply it to learned 2D shape features.
The third is the neural low-SNR literature, which supplies both the method and a negative result. EEG2Rep supplies the method: an online encoder that sees masked input, a slowly averaged target encoder, a predictor that fills in the target's summary of the masked blocks, contiguous-block masking, and a variance-covariance penalty against collapse. Its negative result is that raw-reconstruction masked autoencoders, which fill in masked samples and score them by squared error, were beaten under Gaussian noise and baseline-shift perturbations. We deliberately do not use the generative-prior route of Li et al. or Palazzo et al.. Those approaches let a strong prior fill in plausible structure, and a hallucinated or suppressed Keplerian signal is fatal for a planet claim. Smedemark-Margulies et al. found that discriminative networks work best inside an explicit Bayesian update. We follow that pattern: the network predicts only the activity velocity, and planets are inferred in an ordinary Keplerian likelihood.
Research Program
The input is the CCF sequence of a star, a velocity-by-epoch matrix. Each epoch's CCF is first placed in a translation-invariant representation. We fit a rigid shift against a slowly updated template and pass the re-centred residual profile to the encoder, together with the Fourier amplitude spectrum of the profile. Fourier amplitudes are unchanged by translation, so they act as a second shift-blind channel. The rigid shift itself goes only to the downstream velocity model. To first order, the encoder therefore has no access to planet-induced motion, and we test this empirically rather than assuming it.
The encoder is a temporal sequence model over epochs, applied to per-epoch profile embeddings. We will try a transformer and a linear-time state-space model in the style of FST-Mamba, since the latter keeps the cost linear for decade-long, dense solar series. Pretraining follows EEG2Rep. We mask contiguous blocks of epochs spanning a fraction of a stellar rotation, and also contiguous velocity bands within profiles, then predict the target encoder's embeddings for the masked parts with a variance-covariance penalty. Two sources of unlabeled data are used: millions of synthetic CCF sequences from the SOAP 2.0 spot and facula simulator, and the real HARPS-N solar series. The simulator provides scale, and the real data closes the gap between simulated and observed spectra that SpectraFM flagged for simulator-trained spectral models.
The head maps embeddings to activity velocity. It is linear, with coefficients re-estimated in sliding windows or within segments found by change-point detection on the embeddings. It is fit jointly with Keplerian terms and a white-noise jitter term, so a planet's rigid velocity is never regressed away by the shape model. On the Sun, where the velocity after removing solar system bodies is activity plus instrument, the head can also be fit directly. For other stars it is fit jointly with the orbits.
Baselines
The naive reference point is no activity correction: a Keplerian fit plus white jitter to the raw velocities. This sets the 1.64 m/s solar residual scale reported by Yu et al. The classical scalar reference is the multi-output Gaussian process of Rajpaul et al. on velocity, line-asymmetry and width indicators. Alongside it we run Simola's change-point regression on the same indicators, whose reported α Cen B median threshold of 2.0 m/s we attempt to reproduce before comparing.
The strongest existing method is CCF-GP. Its reported numbers are an RMS velocity residual of 0.44 m/s on two years of HARPS-N solar data, recovery of a 0.3 m/s, 33-day injected planet during high activity, and robust performance down to SNR of roughly 100 (all from Yu et al.). We will rerun it on exactly our CCFs, epochs and train/test time split. Beating it means a lower out-of-sample residual RMS on held-out solar time, a smaller minimum recoverable semi-amplitude at fixed false-alarm rate in injection tests, and no worse bias in recovered amplitude. All three must hold, with confidence intervals from multiple injection realisations.
Four ablations isolate the key ingredients. The first is principal component analysis of the same shift-removed profiles, with the same windowed head. This tests whether a network is needed at all. The second replaces latent prediction with raw-reconstruction masking at matched size and compute. The third replaces contiguous masks with random masks. The fourth feeds un-centred CCFs to the encoder, which lets us measure how much planet signal leaks in when shift-blindness is removed. All learned variants get the same hyperparameter search budget, selected only on a validation period that precedes the test period.
Why not just…?
Why not just run PCA on the shift-removed profiles and regress? This is the most serious threat. Linear projections of shape residuals already capture much of the activity signal, and the file contains no reported result showing a nonlinear encoder beats them. Our PCA-plus-windowed-head ablation settles the question directly. If it matches the full model, the honest conclusion is that time-local refitting and shift-blindness do the work and the network does not. That would still be a useful and publishable result. We expect the encoder to help mainly when photon noise is added, because noise spreads across principal components. The EEG2Rep noise-robustness result is the analogy behind that expectation, but it is not evidence for spectra.
Why not just use CCF-GP, which already works? On two solar years it reaches 0.44 m/s. It is still above the roughly 0.1 m/s scale needed for an Earth twin, it assumes stationary active-region behaviour that Simola et al. showed is false, and its cubic cost limits it on long, dense series. The head-to-head on held-out solar time and on thinned, noised copies decides this. If CCF-GP wins at native SNR but loses at SNR 50, that also defines the regime where each method belongs.
Why not just train a supervised network on solar CCFs to predict velocity? On the Sun this is feasible, and we include it at matched architecture. It has two weaknesses. It cannot be trained on other stars, whose velocities contain unknown planets. And an unconstrained model fed raw profiles can learn rigid shifts, which would absorb planets. The un-centred-input ablation, run under injection, measures exactly that absorption.
Why not a standard masked autoencoder? EEG2Rep reports that raw-reconstruction masking underperforms latent prediction under noise perturbations in EEG. Whether this transfers to spectra is an open question, and our matched ablation answers it.
Proposed Experiments
The main test uses the HARPS-N solar data with time-blocked splits. Pretraining and head fitting happen on earlier data and evaluation on a later, unseen period, with no overlap within one rotation of the boundary. We report residual RMS, then inject rigid-shift Keplerians into the test CCFs. The injections span 0.1 to 1 m/s and periods of 5 to 300 days, deliberately including periods near the solar rotation period. We measure detection completeness at fixed false-alarm probability, bias in recovered amplitude, and posterior coverage. The ablations are repeated after degrading the spectra with photon noise to SNR 100 and 50.
The SOAP 2.0 case with equal 25-day planetary and rotation periods tests whether we match CCF-GP's success where 1D methods failed. The simulator's known spot configurations also let us check whether the embeddings track spot filling factor. Transfer to other stars uses archival HARPS CCFs. For HD 13808, the method should recover the two Neptune-mass planets of Ahrer et al.. For α Cen B, it should not produce the planet candidate that Rajpaul et al. showed was a sampling and activity artifact.
The decisive checks are two. First, injected planets must be recovered without amplitude suppression at every period, including those next to the rotation period. Second, on held-out solar time and at reduced SNR, the full model must beat both CCF-GP and the PCA-windowed ablation. If the first fails, shift-blindness is not doing its job. If the second fails against PCA, the network is unnecessary, and we will report that.
References
- On the Feasibility of Intense Radial Velocity Surveys for Earth-Twin Discoveries
- Workshop Summary: Exoplanet Orbits and Dynamics
- EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs
- A Gaussian process model for stellar activity in 2D line profile time-series
- A Gaussian process framework for modelling stellar activity signals in radial velocity data
- Accounting for stellar activity signals in radial-velocity data by using change point detection techniques
- Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion
- Generative Adversarial Networks Conditioned by Brain Signals
- Recursive Estimation of User Intent From Noninvasive Electroencephalography Using Discriminative Models
- Hierarchical Spatio-Temporal State-Space Modeling for fMRI Analysis
- SpectraFM: Tuning into Stellar Foundation Models
- The HARPS search for southern extra-solar planets – XLV. Two Neptune mass planets orbiting HD 13808: a study of stellar activity modelling’s impact on planet detection
- Ghost in the time series: no planet for Alpha Cen B