I am Matthijs, a postdoctoral researcher at Leiden University working on the application of machine learning techniques to high contrast imaging for direct imaging of exoplanets. Prior to this, I completed my PhD through the College for Doctoral Training in Data Intensive Science at UCL’s Mullard Space Science Lab, where I worked with Jason McEwen and Marta Betcke on developing innovative machine learning approaches for astronomical imaging.

My doctoral research focused on advancing image reconstruction methods for radio interferometry, a crucial tool in astronomy used to study phenomena from the epoch of reionisation to cosmic magnetism and distant galaxies. With next-generation telescopes like the Square Kilometre Array (SKA) set to generate unprecedented volumes of data, my work addressed the pressing need for efficient and scalable reconstruction techniques. I developed two novel approaches: a fully data-driven method and a hybrid approach combining data-driven learning with model-based optimization. These methods were initially developed for SPIDER, an optical interferometer concept, before being extended to radio interferometry.

A key challenge I tackled was adapting these learned reconstruction methods to handle the varying visibility coverages inherent in radio interferometry, developing robust training strategies that made the models coverage-agnostic. I also integrated these approaches into a generative framework capable of quantifying uncertainties in the reconstructions – a crucial feature for scientific interpretation. The resulting methods demonstrated significant improvements in both computational efficiency and reconstruction quality compared to traditional approaches, enabling real-time imaging for SPIDER and offering efficient, high-quality reconstructions with uncertainty quantification for radio interferometric telescopes.

During my PhD, I also collaborated on several other projects. I worked with Spotify to enhance their podcast dataset by precomputing audio features for the TREC Podcast track, a data challenge focused on information retrieval from podcasts. Additionally, I spent 6 months at the STFC Hartree Centre developing deep reinforcement learning methods for controlling plasma shape in nuclear fusion reactors, under the supervision of Adriano Agnello, George Holt, and Nicola Amorisco.

Publications

Talks and Posters