AESTRA II: Generative Spectral Modeling of the Sun as a Star for Precise Radial Velocities

The paper demonstrates that AESTRA, a generative spectral modeling framework applied to NEID Sun-as-a-star data, significantly outperforms traditional activity-indicator detrending methods by recovering 238 injected Earth-analog planets (including 13 with semi-amplitudes below 0.3 m/s) while maintaining zero spurious detections, thereby overcoming key limitations in extreme-precision radial velocity measurements.

The Gist

Imagine you are trying to listen to a very quiet whisper (a planet orbiting a star) in a room that is constantly shaking, humming, and filled with echoes (the star's own activity and Earth's atmosphere). For decades, astronomers have struggled to hear these "whispers" because the noise is so loud it drowns them out.

This paper introduces a new tool called Æstra (pronounced "Astra") that acts like a super-smart noise-canceling headphone specifically designed for starlight. Here is how it works, using simple analogies:

The Problem: The Noisy Room

When astronomers look at a star like our Sun, they aren't just seeing a steady light. The star is a living, breathing ball of gas.

• Stellar Activity: The star's surface is churning with magnetic storms and giant bubbles (like boiling water), which change the shape of the light waves coming from it.
• Earth's Atmosphere: As that starlight passes through our own sky, water vapor and other gases in Earth's atmosphere absorb tiny bits of the light, creating their own "fingerprint" on the data.
• The Goal: They want to detect the tiny wobble caused by an Earth-like planet. This wobble is so small (about 10 centimeters per second) that it's like trying to hear a mosquito buzzing while a jet engine is running next to you.

The Old Way: Guessing and Subtracting

Traditionally, astronomers tried to fix this by:

1. Measuring the total "noise" using simple tools (like measuring the width of a single line in the light spectrum).
2. Guessing how much of the wobble was caused by the star's noise and subtracting it.

The Analogy: Imagine you are trying to hear a friend talk, but you only measure the volume of the background music and subtract that number from the total sound. It's a rough guess. If the music changes its tone or rhythm in complex ways, this simple subtraction fails, and you still can't hear your friend.

The New Way: Æstra (The Generative Model)

The authors built Æstra, which doesn't just guess; it learns what the noise looks like by breaking the light down into its original ingredients.

Think of the starlight data as a complex smoothie.

• The Old Way: You take a sip and guess, "This tastes like strawberries, so I'll subtract strawberries."
• The Æstra Way: It uses a special machine (a neural network) to separate the smoothie back into its exact ingredients: the strawberry juice (the star's changing shape), the water from the ice (Earth's atmosphere), and the sugar (instrumental glitches).

How it works step-by-step:

1. Deconstruction: Æstra looks at the raw light spectrum and learns to identify three distinct things:
   • The Star's "Mood": How the star's surface is churning and changing the shape of its light lines.
   • Earth's "Fingerprint": The specific, narrow lines caused by water vapor in our atmosphere.
   • The "Glitch": Smooth, broad changes caused by the telescope or instrument.
2. Cleaning: It mathematically removes the "Earth's Fingerprint" and the "Glitch," leaving behind a clean version of the star's light that still contains the star's natural "mood" changes.
3. The "Latent" Map: It creates a compact summary (a "latent vector") of the star's mood. Think of this as a weather report for the star. Instead of looking at the whole sky, you just look at the weather report to know if it's stormy.
4. Finding the Planet: Finally, it looks at the total wobble. It uses the "weather report" to say, "Okay, 90% of this wobble is just the star's mood changing. The remaining 10% that doesn't fit the weather report? That must be the planet."

The Results: Hearing the Whisper

The team tested this on years of data from the NEID telescope, which watches our own Sun (acting as a stand-in for other stars). They played a trick: they secretly injected fake "planet signals" (whispers) into the data and asked, "Can you find them?"

• The Old Method: When they used traditional tools to clean the data, they only found 9 of the 500 fake planets. They missed almost everything, especially the quietest ones.
• The Æstra Method: Using their new noise-canceling system, they found 238 of the 500 fake planets.

The Key Takeaway:
Æstra didn't just find more planets; it found the quietest ones that the old methods completely missed. It successfully detected signals as small as 0.3 meters per second (about the speed of a slow walk), whereas the old method gave up on anything slower than 0.5 meters per second.

Why This Matters

The paper claims that by using the entire spectrum of light (the whole smoothie) rather than just a few summary numbers, Æstra can separate the star's noise from the planet's signal much better. This brings us one step closer to finding Earth-like twins around other stars, which have been hiding in the noise until now.

In short: Æstra is a smart filter that learns to distinguish between the star's natural "voice" and the planet's tiny "whisper," allowing astronomers to hear signals that were previously too quiet to detect.

Technical Summary

Technical Summary: AESTRA II: Generative Spectral Modeling of the Sun as a Star for Precise Radial Velocities

Problem Statement
The detection of Earth-analog exoplanets via extreme-precision radial velocities (EPRVs) is currently limited not by photon noise or instrumental stability, but by spectral variability arising from stellar activity, telluric absorption, and instrumental systematics. Traditional approaches often compress spectral data into one-dimensional time series (e.g., Cross-Correlation Functions or CCFs) before attempting to decorrelate activity signals. However, this compression discards wavelength-dependent information crucial for distinguishing between Doppler shifts and activity-induced line-profile distortions. Furthermore, standard preprocessing methods for telluric absorption often rely on external atmospheric templates or fixed instrumental line-spread functions (LSFs), which may be imperfect at sub-m s$^{-1}$ precision levels.

Methodology
The authors apply Æstra, a generative spectrum modeling framework, to high-resolution, high-cadence Sun-as-a-star observations from the NEID spectrograph. The methodology proceeds in three main stages:

1. Spectral Decomposition and Cleaning:

   • The framework treats each observed spectrum as a composition of three components: stellar line-shape variability, micro-telluric absorption, and continuum variability.
   • Using a shared encoder and component-specific decoder branches, the model empirically decomposes the spectra without relying on external atmospheric or stellar templates.
   • The telluric component is constrained to describe narrow, Earth-frame absorption features, while the continuum component captures smooth, low-frequency variations. The stellar component captures localized line-profile distortions.
   • Learned telluric and continuum components are removed from the spectra, while stellar variability is preserved. The cleaned spectra from 42 high-quality echelle orders (spanning 4300–6230 Å) are merged into a single activity-preserving spectrum.

2. Latent Representation of Stellar Activity:

   • The cleaned spectra are shifted to remove the measured apparent radial velocity (RV), isolating residual line-shape variability.
   • An attentive autoencoder (Æstra) maps these spectra to a low-dimensional latent vector ($S=20$ dimensions). This vector serves as a compact, empirical representation of the stellar activity state.
   • Validation shows the leading latent component is strongly correlated ($r=0.889$) with the independent Ca II H&K chromospheric activity index, despite the H&K region being excluded from the training wavelength range.

3. Joint RV Inference:

   • A small feed-forward neural network (Activity RV Estimator) maps the latent activity vector to an activity-driven RV contribution ($v_{act}$).
   • The total apparent RV is modeled as the sum of the activity contribution, candidate planetary signals (modeled as sinusoids), and a trend term.
   • The model is trained to separate activity-driven variability from coherent Doppler signals. The planetary signals are initialized from a traditional period search but are refined jointly with the activity model.

Key Contributions and Validation Strategy
The paper introduces a unified generative approach that operates directly on the spectrum level rather than the compressed RV level. To validate the method, the authors conducted 500 single-planet injection–recovery tests on the NEID solar time series.

• Injection Design: Synthetic planetary signals with periods of 2.5–400 days and semi-amplitudes ($K$) of 0.1–0.7 m s$^{-1}$ were injected into the apparent RV time series.
• Period-Blind Recovery: The analysis was designed to be period-blind; candidates were selected without prior knowledge of the injected parameters.
• Calibration: Detection criteria were calibrated on the injection suite to yield zero spurious detections (false positives).

Results
Under the strict criterion of zero false positives:

• Æstra Performance: The method successfully recovered 238 out of 500 injected planets. Notably, it recovered 13 planets with semi-amplitudes below $K = 0.3$ m s$^{-1}$, a regime where traditional methods struggle.
• Traditional Baseline: A traditional approach using CCF-based activity indicators (CCF depth, FWHM, BIS SPAN) for detrending recovered only 9 planets and failed to recover any signals with $K < 0.5$ m s$^{-1}$.
• Sensitivity: Æstra demonstrated an order-of-magnitude improvement in detection completeness compared to the traditional baseline. While the traditional method could recover more signals if the false-alarm probability (FAP) threshold was relaxed (e.g., to 0.1%), this resulted in 322 spurious detections, highlighting the superior specificity of the generative approach.

Significance and Claims
The paper claims that Æstra II demonstrates the feasibility of distinguishing low-amplitude coherent Doppler signals from activity-driven, wavelength-dependent spectral variability in real data. The authors assert that:

• Spectrum-level generative modeling can exploit information lost in traditional RV-level reductions, enabling the detection of sub-meter-per-second signals.
• The method effectively disentangles telluric absorption, continuum fluctuations, and stellar line-shape changes without external templates.
• The approach provides a practical path toward detecting Earth-analog signals that are currently buried beneath the stellar activity noise floor.

Limitations and Scope
The authors note several limitations:

• The validation is based on the NEID solar feed, which offers exceptionally high signal-to-noise and cadence; performance on fainter, more active, or rapidly rotating stars remains to be tested.
• The current implementation assumes single-planet, circular orbits and relies on an initial candidate generation step from traditional searches; it cannot simultaneously recover multiple planets or signals missed by the initial candidate list.
• The method does not yet train a fully integrated neural RV estimator for computational scalability, separating the spectral decomposition from the RV inference steps.
• Degeneracies between long-lived active regions and planetary signals at similar periods remain a fundamental challenge.

The paper concludes that while the Sun serves as a controlled benchmark, the results suggest that spectrum-level generative modeling is a promising avenue for next-generation EPRV surveys targeting Earth-mass planets.