The main challenge in exoplanet high-contrast imaging (HCI) is to separate the faint planetary signal from the bright glare of its host star. For ground-based observations, speckle noise from atmospheric turbulence and optical imperfections exacerbates this problem. Post-processing techniques aim to suppress this speckle noise, but they often remove part of the signal in the process.
We employ explainable machine learning to investigate why one of the most commonly used methods, principal component analysis (PCA), tends to over-subtract the planet. We demonstrate that PCA learns a representation of the noise which overlaps with the characteristic shape of the planetary signal. Building on these insights, we introduce a new algorithm called 4S, which constrains the noise model to preserve the planet’s signal. 4S achieves up to 1.5 magnitudes deeper contrast at small angular separations (<4lambda/D), opening up a new discovery space for exoplanet imaging.
To exploit this gain, we uniformly reprocessed the entire VLT/NACO archive obtained in L’-band pupil tracking mode between 2009 and 2019. The complete sample comprises 560 stars and incorporates extensive surveys such as NaCo-ISPY, making it one of the largest archival searches ever conducted for exoplanet imaging. The enhanced detection capabilities of 4S allow for five pre-discovery recoveries, including the exoplanet AF Lep b in data from 2011— eleven years before its reported discovery — and identify sixteen new companion candidates, which are now being followed up in a dedicated 50-hour VLT/ERIS program.
Finally, we derive detection limits for coronagraphic and non-coronagraphic observations with VLT/NaCo and VLT/ERIS. Surprisingly, non-coronagraphic VLT/NaCo data outperformed coronagraphic observations in both contrast- and background-limited regimes by approximately one magnitude. A direct comparison of the detection limits provides practical guidance for future L′-band surveys with VLT/ERIS and the coronagraph design of ELT/METRIS.
