Detection and characterization of galaxies using Machine Learning on a massive radio-astronomical dataset
david.cornu@observatoiredeparis.psl.eu (Webpage) philippe.salome@observatoiredeparis.psl.eu (Webpage) cyril.tasse@observatoiredeparis.psl.eu
This Master 1 internship is to join the MINERVA project (Machine Learning for Radioastronomy at the Observatoire de Paris) during the spring or summer of 2023.
Context – In the perspective of the future SKA (Square Kilometer Array) radio-telescope, radio-astronomy is experiencing a rebirth in its low-frequency domain. However, this new age is also facing an explosion of volumes of observational data. This is particularly true with the development of giant interferometers (LOFAR, ALMA, NenuFAR, SKA). These instruments produce very large and highly dimensional datasets. In this context, traditional methods of source detection and classification reach their limits. In parallel, Machine Learning methods have undergone algorithmic developments that bring them to a high level of maturity for these tasks.
Artificial Intelligence – The MINERVA team has recently won the second SKA data challenge that was focussed on source detection and characterization in a large 3D synthetic cube. The team developed a specialized Deep Learning detection method to achieve this goal. In the meantime, we also applied these method to the dataset of the previous edition of the SKA data challenge that was focussed on source detection and classification in synthetic 2D images. In both contexts, the approach demonstrated state-of-the-art detection performance. The team now seeks to use this methodology for various real observed datasets in order to construct new source catalogs and is exploring new ways of combining information from multiple surveys.
Galaxy survey and Observations – The MINERVA team is also involved in LoTSS (the LOFAR Two-metre Sky Survey: a high-resolution 120-168 MHz survey of the entire northern sky). This survey opens an unprecedented view of the radio sky and its population of radio-galaxies. However, source detection and classification is still an on-going work. In particular, the cross-matching with galaxy catalogs in the optical domain still encounters problems because of the different morphology of the objects at different wavelengths.
Internship mission – This proposed internship follows the work done in a previous one on applying the detection method developed by the MINERVA team to real observational data, namely the LoTSS fields. We already demonstrated that direct use of the detector trained on SDC1 data can efficiently identify most of the previously identified sources in the LoTSS survey. Still, many differences remain between the simulated SDC1 dataset on which the network was trained and the LoTSS on which it is applied, leading to suboptimal results and false detections in the presence of imaging artifacts. The present internship will aim to construct a complementary training sample specifically for the LoTSS survey by pruning source catalogs obtained by other detection methods (e.g PyBDSF). The objective will be to re-train the detector and assess its new prediction performances on LoTSS fields. To achieve this goal, the internship fellow will have access to computing resources (GPUs) dedicated to the MINERVA project.
We encourage applications from candidates with some knowledge of Machine Learning and Neural Networks methods, but it is not mandatory and can be learned during the internship. Skills with manipulating astronomical images and data cubes will also be considered. A good level in Python (and ideally in C/C++) programming language is strongly advised.