AI Finds New Earths: How Deep Learning is Cleaning Up Starlight to Detect Rocky Planets

Redacción HC
29/04/2024

The search for Earth-like exoplanets — small, rocky worlds orbiting stars like our Sun — has long been one of the great quests of modern astronomy. But while our telescopes are powerful and our instruments are precise, the stars themselves can deceive us. Their flickers, spots, and flares create noise that masks the delicate signals of orbiting planets, especially those no bigger than Earth.

Now, a new study accepted in Astronomy & Astrophysics and led by an international team including researchers from Switzerland, the U.S., Italy, and Peru, demonstrates that deep learning may be the key to tuning out the noise. By applying convolutional neural networks (CNNs) directly to high-resolution stellar spectra, the team has achieved unprecedented sensitivity in detecting Earth-like planets through radial velocity — a technique that tracks the subtle wobble of stars.

The Problem: When Stars Speak Too Loudly

The radial velocity (RV) method detects planets by measuring the motion of stars induced by gravitational tugs. But stars are not static. Their surfaces boil with convection cells, rotate with magnetic spots, and flare unpredictably. These processes distort the spectral lines used to infer RV, often mimicking or drowning out planetary signals.

For Earth-sized planets, the induced stellar motion is less than 1 meter per second — about the speed of a slow walk. Differentiating that from stellar noise has been a major barrier.

"Can a deep learning model trained on spectral profiles distinguish between real planetary signals and stellar activity?" That's the central question of this groundbreaking study.

The Deep Learning Solution

The researchers trained a CNN on thousands of spectral line profiles from three well-known stars: Alpha Centauri B, Tau Ceti, and the Sun (via HARPS-N solar telescope). These spectra included both real data and synthetic planetary signals — carefully injected to simulate Earth-like planets with velocity amplitudes as low as 0.2–0.5 m/s and orbital periods from 10 to 300 days.

This model learned to associate subtle changes in the shape, asymmetry, and width of the spectral lines with activity noise, and to filter it out.

Performance Highlights

1. Alpha Centauri B and Tau Ceti:
   • CNN reduced stellar noise enough to detect signals down to 0.5 m/s, enabling discovery of planets around 4 Earth masses in habitable zones.
2. The Sun:
   • Even more impressive, the model achieved 0.2 m/s precision, corresponding to potential planets as small as 2.2 Earth masses.
3. Outperforms traditional methods:
   • Standard RV pipelines based on line indicators or Gaussian process models couldn't match the precision achieved by this CNN approach.
   • The AI model extracted information from line shapes that classical tools overlook.

"Deep learning doesn't need explicit physical modeling — it learns directly from the data patterns," noted co-author Xavier Dumusque.

Why It Matters: From Hardware to Software Innovation

This research represents a major step forward for software-driven astronomical breakthroughs. Instead of building new telescopes, scientists are improving what we can see with the data we already have.

Impacts Across Astronomy

• Exoplanet detection: Paves the way for detecting rocky worlds in the habitable zones of nearby stars — a key goal of upcoming missions like PLATO and telescopes like the Extremely Large Telescope (ELT).
• Observational strategy: The method works on existing high-resolution instruments like HARPS, HARPS-N, and ESPRESSO, and could be adapted for future ones.
• Scientific policy: Justifies more investment in AI applications within astronomy, which can accelerate discoveries without the need for costly new infrastructure.

Challenges Ahead and Next Steps

Despite the promising results, the CNN still needs broader testing:

• It has so far been trained and validated mostly on quiet, well-behaved stars. More turbulent stellar environments will test its generalization capacity.
• The model's portability across different instruments, exposure times, and spectral calibrations remains to be confirmed.

The team recommends:

1. Embedding the CNN in current RV pipelines,
2. Expanding training sets with more active stars, and
3. Combining CNN-based detection with robust statistical modeling for higher confidence in planet claims.

Latin American Contributions

The study includes Isidro Gómez‑Vargas, affiliated with the Instituto de Astronomía of the UNAM (Peru/Mexico). His participation underscores Latin America's growing presence in front-line astronomical research.

With numerous observatories across Chile, Brazil, and Mexico already collecting RV data, this model opens exciting possibilities for regional scientists to enhance their planet-hunting capabilities using AI.

Conclusion: AI Is Clearing the Way to New Earths

This study is not just another incremental improvement in exoplanet detection. It's a paradigm shift — showing that AI can isolate the faint fingerprints of rocky planets from the noisy chatter of stars.

The implications are far-reaching. As we refine these models and train them on more diverse data, we move closer to confidently spotting Earth's twin — a world like ours, orbiting a distant star, hidden until now beneath layers of cosmic noise.

As the study authors argue, "This approach enables new discoveries without requiring hardware upgrades — a leap in detection capabilities through data science."

Stay tuned: the next Earth might already be in the data. We just needed a smarter way to listen.

Topics of interest

Technology