💻 computer science

Modeling Doppler Shifts in Radial-Velocity Data with Deep Learning toward Earth-mass Exoplanet Detection

This paper presents a robust deep-learning framework, implemented in the `doppleriann` Python package, that combines physics-motivated spectral representations with advanced training strategies to reliably detect Earth-mass exoplanets in real stellar radial-velocity data by overcoming the challenges of stellar activity.

Original authors: Isidro Gómez-Vargas, Xavier Dumusque, Yinan Zhao, Khaled Al Moulla, Michael Cretignier

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Isidro Gómez-Vargas, Xavier Dumusque, Yinan Zhao, Khaled Al Moulla, Michael Cretignier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding a Needle in a Haystack (That's Also Shaking)

Imagine you are trying to hear a tiny, specific whisper (an Earth-sized planet) in a room where a giant is constantly shouting, breathing heavily, and shifting their weight (the star's activity). This is the challenge astronomers face when looking for Earth-like planets using the Radial Velocity method.

This method works by listening to the "wobble" of a star. As a planet orbits, it tugs on the star, causing the star's light to shift slightly in color (the Doppler shift). For an Earth-sized planet, this shift is incredibly tiny—about the speed of a slow walk (0.1 meters per second). The problem is that the star itself is noisy. It has sunspots, magnetic storms, and surface boiling that create "fake" wobbles much louder than the planet's tug.

The New Tool: A "Smart Filter" Made of Deep Learning

The authors of this paper built a new kind of computer brain (a Deep Learning model) designed to filter out the star's noise and hear the planet's whisper.

Instead of feeding the computer the raw, messy starlight, they gave it a special, compressed version of the data called a "Spectral Shell."

The Analogy: Imagine looking at a complex painting. A standard computer might just look at the total amount of red paint (Flux). But this new method looks at how the paint is applied and the temperature of the brushstrokes.
The Innovation: The team realized that different parts of a star's surface behave differently depending on how "deep" you look into it. They created two versions of their data:
1. Flux-based: Looking at the brightness of the light.
2. Temperature-based: Looking at the "formation temperature" of the light (essentially, which layer of the star's atmosphere the light is coming from).

They found that the Temperature-based version was like putting on special glasses that made the star's noise disappear, leaving the planet's signal much clearer.

How They Trained the Computer

You can't just teach a computer to find planets by showing it real data, because we don't know for sure if the tiny signals are real planets or just noise. So, the scientists played a game of "hide and seek" with their own data:

The Setup: They took 10 years of real sunlight data captured by the HARPS-N telescope.
The Trick: They secretly "injected" fake planetary signals into the data. They told the computer, "We added a planet here with a specific wobble."
The Test: They trained the computer to find these fake signals.
- Strategy A (The Blind Test): They trained the computer on 80% of the data and tested it on 20% it had never seen before. This checks if the computer is actually learning the rules or just memorizing the answers.
- Strategy B (The Round-Robin): They rotated the data, training on different chunks and testing on the rest, to ensure the computer was robust.

What They Found

The results were impressive, especially for the Temperature-based model:

The Sensitivity: The best model could reliably detect planetary signals as small as 20 cm/s (about 0.7 km/h). To put that in perspective, that's like detecting a planet that is tugging the star with the force of a gentle breeze.
The Range: It worked for planets orbiting anywhere from 10 days to 550 days away from their star.
The "Ghost" Signal: In many cases, traditional methods (looking at the raw light) saw only the star's noise. But the new model, using the temperature data, successfully isolated the planet's signal, showing a clear peak in the data where the planet was hiding.

The "Magic Glasses" Analogy

Think of the star's noise as a foggy window.

Old methods tried to wipe the window with a cloth (mathematical corrections), but the fog kept coming back.
This new method didn't just wipe the window; it changed the color of the glass. By looking at the "temperature" of the light instead of just the brightness, the fog (stellar activity) became transparent, and the planet (the signal) stood out clearly against the background.

The Takeaway

The authors didn't just build a model; they released a free software package called doppleriann so others can use it.

Their main conclusion is that by combining physics (understanding how starlight forms at different temperatures) with artificial intelligence, we can finally start hearing the whispers of Earth-sized planets that were previously drowned out by the star's shouting. This is a crucial step toward finding a second Earth.

Important Note: The paper focuses strictly on analyzing data from our Sun (as a testbed) and detecting simulated signals. It does not claim to have found a new Earth-like planet yet, but rather proves the method works well enough to try finding one in the future.

Technical Summary: Modeling Doppler Shifts in Radial-Velocity Data with Deep Learning toward Earth-mass Exoplanet Detection

Problem Statement
Detecting Earth-mass exoplanets via the radial-velocity (RV) method requires measuring stellar Doppler shifts with a precision of approximately 0.1 m/s. However, stellar activity (oscillations, granulation, and magnetic activity) induces RV variations that often exceed planetary signals, dominating the noise budget even with state-of-the-art spectrographs like ESPRESSO. While machine learning (ML) techniques, particularly deep learning (DL), have shown promise on simulated data, many existing methods struggle to generalize to real, unseen stellar spectra or require computationally intensive retraining for specific targets. Furthermore, many approaches rely on Cross-Correlation Functions (CCFs) which may not fully disentangle activity from planetary signals.

Methodology
The authors propose a deep-learning framework designed to model and predict Doppler shifts (DS) directly from high-resolution stellar spectra, aiming to isolate planetary signals from stellar variability. The methodology consists of three core components:

Physics-Motivated Spectral Representations: Instead of using raw spectra or standard CCFs, the authors utilize "spectral shells," a reduced-dimensionality projection of spectra. They introduce two specific representations:
- Flux-based: Projects normalized flux against the flux gradient with respect to velocity.
- Temperature-based: Replaces normalized flux with the line-formation temperature ( $T_{1/2}$ ), derived from radiative transfer modeling. This approach leverages the physical understanding that stellar activity signatures depend on the formation depth (temperature) of spectral lines.
  Both representations are constructed on a $9 \times 9$ grid and weighted to minimize the contribution of noisy or low-information regions.
Deep Learning Architecture:
- Model: Supervised Convolutional Neural Networks (CNNs) with a 1D convolutional architecture (chosen for efficiency over 2D given the structured nature of the shell data).
- Outputs: The network predicts two values: the total RV (to validate against traditional CCF methods) and the injected Doppler shift (the planetary signal).
- Training Strategy: The models are trained on HARPS-N solar spectra (2036 daily spectra from 2015–2024) with synthetic Keplerian planetary signals injected. The injection grid covers semi-amplitudes from 0.1 to 5 m/s and periods from 10 to 100 days (extended to 550 days for evaluation).
- Optimization: Hyperparameters are optimized using a Genetic Algorithm (NSGA-II). Predictive uncertainty is quantified using Monte Carlo Dropout (MC-DO).
- Validation: Two strategies are employed to test generalization:
  - Hold-out (HO): A fixed set of 400 spectra is held out entirely from training.
  - Cross-Validation (CV): An $N$ -fold strategy (5-fold and 10-fold) where the model is trained on subsets and tested on unseen folds, concatenating predictions to reconstruct the full time series.
Software Release: The authors release doppleriann, a Python package implementing the spectral shell construction, signal injection, and the deep-learning framework.

Key Results
The study evaluates the models based on detection probability, and the recovery of amplitude, phase, and orbital period.

Superiority of Temperature-Based Representations: Models using line-formation temperature shells consistently outperform flux-based models across all metrics. They demonstrate higher detection rates, lower amplitude differences, and better phase/period recovery. This confirms that encoding photospheric depth information helps disentangle activity from planetary signals.
Detection Limits:
- Hold-out Strategy: The temperature-based model reliably recovers signals with semi-amplitudes $\ge 0.35$ m/s.
- Cross-Validation Strategy: The temperature-based model achieves higher sensitivity, reliably recovering signals with semi-amplitudes as low as 0.20 m/s (20 cm/s) for periods between 10 and 550 days.
- Strict Criteria: When requiring the recovered signal to be the dominant peak in the periodogram (rather than just a peak above the False Alarm Probability threshold), the detection limit for the CV-Temp model rises slightly to 0.25 m/s.
Performance Metrics:
- Amplitude: Relative amplitude differences generally remain below 30% for successful recoveries, with temperature models showing greater stability, especially at low amplitudes and long periods.
- Phase & Period: Phase differences are typically $< 0.35$ cycles, and period differences are $< 3\%$ across the parameter space.
Generalization: The models successfully generalize to unseen temporal data (in the HO test) and unseen spectral subsets (in CV), demonstrating robustness without retraining for specific stars.

Significance and Claims
The paper claims that combining physically motivated spectral representations (specifically line-formation temperature) with deep learning provides a viable pathway toward detecting Earth-mass planets in real RV data. The significance of the work lies in:

Physical Grounding: By incorporating line-formation temperature, the model leverages known physics of stellar activity (depth dependence) rather than relying solely on statistical correlations, improving the separation of activity noise from planetary signals.
Generalization: Unlike many previous DL approaches that require retraining per star, this framework demonstrates the ability to generalize to unseen spectra and temporal baselines using lightweight architectures.
Sensitivity: The approach achieves detection limits (20–25 cm/s) that approach the requirements for detecting Earth-mass planets in the habitable zone of solar-type stars, surpassing the performance of traditional CCF-based analysis in the presence of stellar activity.
Rigorous Validation: The use of both hold-out and cross-validation strategies, coupled with uncertainty quantification, provides a statistically rigorous assessment of the model's reliability and robustness against overfitting.

The authors conclude that while the current results are based on solar data, the framework offers a promising, physically grounded, and statistically rigorous method for enhancing the sensitivity of radial-velocity surveys to low-amplitude planetary signals.