🔭 astrophysics

Comparing telluric removal methods in their capability to recover injected exoplanet atmosphere signals with high resolution emission spectroscopy

This study compares three telluric removal methods (PCA, Molecfit, and Astroclimes) using CARMENES observations of $\tau$ Bootis b, finding that while PCA can yield higher signal-to-noise ratios, it often degrades the signal more than model-based approaches, and ultimately reports non-detections of the previously claimed water signal in the planet's atmosphere.

Original authors: Marcelo Aron Fetzner Keniger, Matteo Brogi, David Armstrong, Siddharth Gandhi

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marcelo Aron Fetzner Keniger, Matteo Brogi, David Armstrong, Siddharth Gandhi

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: Listening for a Whisper in a Storm

Imagine you are trying to listen to a very quiet whisper (the atmosphere of a distant planet) coming from a stage. However, the room is incredibly noisy. There are two main sources of noise:

The "Room" Noise (Telluric Lines): The air in the room itself is making sounds (water vapor and oxygen in Earth's atmosphere).
The "Stage" Noise (Stellar Lines): The person on stage (the host star) is shouting so loudly that their voice drowns out the whisper.

Astronomers use powerful telescopes to try to hear that planetary whisper. But to do it, they have to use "noise-canceling" software to remove the Earth's atmosphere and the star's light. The problem is, there is no agreement on which "noise-canceling" method works best. Some methods might be too gentle and leave the noise in; others might be too aggressive and accidentally delete the whisper along with the noise.

This paper is a "taste test" to see which of three different noise-canceling methods works best at keeping the planetary signal safe while removing the noise.

The Three Contenders

The researchers tested three different algorithms (computer programs) designed to clean up the data:

PCA (Principal Component Analysis): Think of this as a statistical eraser. It looks at all the data and says, "What patterns appear in almost every single photo?" It assumes that if a pattern is there all the time, it's noise (like the Earth's atmosphere or the star). It then erases those patterns.
- The Risk: If the whisper (the planet) moves around just a little bit, the eraser might think the whisper is part of the background noise and erase it too.
Molecfit: This is a model-based builder. Instead of just erasing, it tries to build a perfect 3D model of what the Earth's atmosphere should look like at that exact moment, based on temperature and humidity. It then subtracts that model from the data.
Astroclimes: This is the researchers' own new builder. It works similarly to Molecfit but uses a different, more flexible way of calculating the Earth's atmosphere. It also includes the star's voice in its model to make sure it doesn't get confused.

The Experiment: The "Fake Whisper" Test

To test these methods fairly, the researchers didn't just look at real data; they created a simulation.

They took real telescope data of a planet called Tau Bootis b.
They secretly injected a "fake" water signal (a whisper) into the data.
They then ran the three different cleaning methods on the data to see if they could find the fake whisper.

They asked two main questions:

Did they find it? (How loud is the signal after cleaning?)
Did they break it? (Did the cleaning process change the signal's location or make it look weaker than it really is?)

The Results: What They Found

1. The "Aggressive Eraser" (PCA)
PCA is very good at making the data look clean. Sometimes, it even makes the signal look louder (higher Signal-to-Noise ratio) than the other methods.

The Catch: It's like a heavy-handed editor. To get that clean look, it often deletes parts of the actual whisper. If the planet is moving slowly (low orbital velocity), PCA gets confused and deletes a huge chunk of the signal. It also tends to shift the location of the signal, making it look like the planet is in a slightly different place than it actually is.

2. The "Model Builders" (Astroclimes and Molecfit)
These two methods were much gentler. They didn't make the signal look quite as "loud" as PCA sometimes did, but they preserved the signal much better.

They kept more of the original "whisper" intact.
They were less likely to move the signal to the wrong location.
They struggled a bit more when the planet was moving slowly, but not nearly as badly as PCA did.

3. The "Deep Lines" Problem
The researchers also tested what happens if they ignore the deepest, darkest parts of the noise (the "deep lines"). They found that while ignoring these lines can sometimes make the signal look slightly louder, it doesn't actually change the fundamental results. It's mostly a matter of preference.

The Final Twist: The Real Mystery

After testing their methods with fake signals, the researchers tried to find the real water signal that other scientists had previously claimed to see in Tau Bootis b.

The Result: They could not find it.
Using their new methods (Astroclimes) and the standard method (PCA), they looked at the same data used by the previous study, but they found no evidence of water.
They tried many different settings and models, but the signal remained silent. This suggests that the previous detection might have been a false alarm caused by the noise-canceling method used back then, or that the water signal is much weaker than previously thought.

The Takeaway

The paper concludes that how you clean your data matters just as much as the data itself.

If you use a method that is too aggressive (like PCA with too many settings), you might "find" a signal that isn't really there, or you might distort the signal so much that you get the wrong answer about what the planet is made of.
Methods that build models of the atmosphere (like Astroclimes and Molecfit) seem to be safer because they are less likely to accidentally delete the planet's voice.

In short: To hear the universe's whispers clearly, you need a noise-canceling tool that is smart enough to know the difference between the noise and the whisper, without accidentally deleting the whisper in the process.

Technical Summary: Comparing Telluric Removal Methods for High-Resolution Emission Spectroscopy

Problem Statement
Ground-based high-resolution spectroscopy (HRS) is a critical tool for characterizing exoplanet atmospheres, offering resolving powers ( $R > 25,000$ ) that space-based telescopes often cannot match. However, a major bottleneck in this analysis is the removal of contamination from Earth's atmosphere (telluric lines) and the host star (stellar lines). While Principal Component Analysis (PCA) and its variants (e.g., SYSREM) are widely used to blindly detrend data by removing stationary spectral features, they risk removing or distorting the planetary signal, particularly for planets with low orbital velocities. Conversely, forward-modeling approaches like Molecfit attempt to physically model the telluric spectrum but have historically faced challenges in infrared wavelength ranges and with specific instrument configurations. There is currently no consensus in the literature on the optimal method for telluric correction, and the biases introduced by different detrending techniques on recovered planetary signals remain insufficiently quantified.

Methodology
The authors utilize near-infrared observations of the non-transiting hot Jupiter $\tau$ Bootis b obtained with the CARMENES spectrograph. The study focuses on the detection of water ( $H_2O$ ) signatures, which are heavily affected by telluric variability. The authors conduct a series of injection and recovery tests using a synthetic planetary signal with a volume mixing ratio of $VMR_{H_2O} = 10^{-3.0}$ , injected into the observational data cubes.

Three distinct detrending methods are compared:

Astroclimes: A newly developed synthetic telluric transmission algorithm (Fetzner Keniger et al. 2025) that uses Markov-Chain Monte Carlo (MCMC) optimization with GGG2020 atmospheric profiles to fit telluric abundances. It incorporates a stellar model (from the Göttingen Spectral Library) to account for stellar lines, dividing the data by a "master stellar spectrum" after telluric removal.
Molecfit: A standard synthetic transmission algorithm using Levenberg-Marquardt least-squares fitting with GDAS and MIPAS atmospheric profiles. An experimental version was utilized to support the CARMENES instrument.
PCA: A blind detrending method using Singular Value Decomposition (SVD) to identify and remove common modes (telluric and stellar lines) across the time series. The study tests various numbers of principal components ( $N_C$ ) to assess signal degradation.

The performance of each method is evaluated by measuring the Signal-to-Noise Ratio (SNR) of the recovered signal and quantifying signal degradation via a scaling factor ( $\delta$ ) and deviations in retrieved orbital parameters ( $K_p$ and $V_{sys}$ ). The analysis is performed on multiple observing nights with varying weather conditions (Precipitable Water Vapor) and orbital velocities ( $K_p = 50, 80, 110$ km/s).

Key Results

Signal Degradation vs. SNR: While PCA with an optimal number of components ( $N_C=3$ for Night 1) achieved the highest SNR, it came at the cost of significant signal degradation. PCA removed approximately 60–80% of the injected signal magnitude (scaling factors $\delta \approx 0.17–0.40$ ), whereas Astroclimes and Molecfit preserved significantly more of the signal ( $\delta \approx 0.64–0.70$ ).
Orbital Velocity Dependence: All methods struggled more with injected signals at lower orbital velocities ( $K_p = 50$ km/s). However, PCA was disproportionately affected, showing severe signal loss and significant biases in the retrieved $K_p$ and $V_{sys}$ values. Astroclimes and Molecfit maintained better signal retention, though they also exhibited increased bias in parameter retrieval for low-velocity cases.
Signal Distortion: The detrending process altered not only the magnitude of the signal but also its location in velocity space. For lower orbital velocities, the retrieved parameters deviated significantly from the injected values across all methods, highlighting the potential for biases in characterizing real detections.
Deep Line Masking: The study explored the impact of masking deep telluric lines (transmission < 0.2). While higher thresholds occasionally improved SNR by removing poorly corrected artifacts, they did not significantly alter the retrieved MCMC parameters.
Real Signal Detection: When attempting to detect the previously claimed water signal in $\tau$ Bootis b (Webb et al. 2022) using the same dataset but different detrending methods (Astroclimes and PCA) without model reprocessing, the authors found no significant detection. This result aligns with non-detections reported by Brogi et al. (2012) and Panwar et al. (2024), suggesting the previous detection may be sensitive to specific analysis choices, such as model reprocessing or orbital solutions.

Significance and Conclusions
The paper demonstrates that while PCA can yield high SNR, it is often overly aggressive, degrading the planetary signal and introducing biases in retrieved atmospheric parameters, particularly for planets with low orbital velocities. In contrast, forward-modeling approaches like Astroclimes and Molecfit offer superior signal preservation, making them more robust for atmospheric retrieval where accurate abundance measurements are critical.

The authors emphasize that the choice of detrending method fundamentally alters the nature of the recovered exoplanet signal. They argue that understanding these effects is essential to avoid biases in characterizing real detections. Furthermore, the study highlights that Astroclimes offers a flexible alternative to Molecfit, capable of handling non-ESO instruments and integrating directly with atmospheric retrieval frameworks, though further work is needed to eliminate systematic biases and improve telluric artifact removal in specific conditions. Ultimately, the work underscores the necessity of refining ground-based HRS cleaning methods to ensure reliable exoplanet atmosphere characterization alongside space-based observations.