Decoding Speckles, Discovering Treasures

Over the past decade, machine learning has become an important part of many fields in the physical sciences. While the success of these methods is undeniable, they are often used as "black boxes" which limits their interpretability. This thesis explores the use of computational methods and machine learning in exoplanet high-contrast imaging (HCI), a technique that directly detects exoplanets by resolving their light from that of their host star. This thesis focuses on removing and quantifying speckle noise, a type of systematic noise caused by imperfections in telescope optics and atmospheric turbulence. The methods developed in this thesis not only achieve better results but also contribute to a better understanding of the data and the underlying physics. The first contribution is a new statistical framework for the robust quantification of HCI detection limits. The method is based on parametric bootstrapping and generalizes the commonly used standard to account for non-Gaussian speckle noise. By comparing detection limits under different noise assumptions, we find that non-Gaussian noise can bias detection limits by approximately one magnitude. The second contribution is the introduction of 4S (Signal-Safe Speckle Subtraction), an explainable machine learning algorithm for speckle subtraction. 4S not only outperforms the commonly used baseline methods, but also explores new ways to incorporate domain knowledge into the algorithm. Using saliency maps, 4S provides insight into the underlying noise structures, revealing a physical correspondence with known speckle behavior. The improvement provided by \fours is largest at small separations from the star. This enhancement enables the detection of the exoplanet AF Lep b in archival data from 2011, over a decade before its subsequent discovery. This additional astrometric data point helps us to significantly improve the constraints on the orbit and mass of the companion. The third contribution is the first uniform reanalysis of the entire NaCo L'-band archive. Using 4S on these data, we identified four additional known companions in archival data taken before their official discovery, as well as sixteen new companion candidates. Future observations with ERIS will confirm or refute whether the candidates are bona fide companions. A quantitative comparison of the detection limits of coronagraphic and non-coronagraphic datasets shows that the vortex coronagraph in \naco, yields shallower detection limits than the non-coronagraphic data. This performance loss is partially due to the effectiveness of the data post-processing -- a result that underlines the importance of considering algorithm-instrument synergies during instrument design. Overall, this thesis presents a foundation for more transparent, physically interpretable, and statistically sound exoplanet imaging analyses. It paves the way for both deeper detection limits in existing data and more effective use of upcoming facilities, such as METIS and PCS, at the ELT.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.