Offre en lien avec l’Action/le Réseau : BigData4Astro/– — –

Laboratoire/Entreprise : LESIA, Observatoire de Paris-PSL
Durée : 18 to 24 months
Contact : baptiste.cecconi@obspm.fr
Date limite de publication : 2024-01-31

Contexte :
The EXTRACT project (EU, https://extract-project.eu) is currently conducting the design of an edge-to-cloud solution for heavy data processing based on Deep learning methodologies. One of the use case is the project Transient Astrophysics using SKA pathfinders (TASKA) that covers the processing of dynamical astronomical imaging data in radio using deep learning.

Radio astronomy imaging involves the inversion of set of Fourier domain samples acquired by an interferometer, observing a certain direction of the sky at a specific temporal and spectral rate.

The transformation from the recorded data (set of sparse and incomplete Fourier samples) to a multidimensional image cube containing scientific information, is a strong and ill-posed inverse problem.

For decades, « classical » radio interferometric imaging usually involved the production of single 2D image from the averaging (in time and frequency) of Fourier samples and solving the deconvolution problem to remove the instrumental impulse response. The target was to obtain a static image of the sky in radio. The classical CLEAN algorithm (1974) and CLEAN derivatives methods were historically the most widespread methods used to solve for the problem, mainly in the image space.

Sujet :
When the observed sky is steady, the accumulation of Fourier samples helps getting better images with improved signal-to-noise ratio (SNR) and image fidelity. However, if an astrophysical event (a.k.a. a radio « transient ») occurs during the observation, a long time integration can average out and prevent the detection of such short-lived event. Fast snapshot imaging using the same methods trying to follow fast variations of the sky provide a very limited SNR that would limit the detection level to only powerful astrophysical transients. In addition, extended time-variable emission (e.g., Solar flares, planetary emissions, etc.) are only poorly imaged using classical 2D imaging at a higher rate.

With the development of machine learning and deep learning (ML/DL) methodologies, solving for the imaging and deconvolution can be revisited to produce images cubes with the lowest possible bias while maintaining the integrity of the physical information measured from the sky. The imaging problem is analog to a video restoration problem where identified features are restored and tracked in time and spectral domains.

Hopefully, the astrophysical transients usually have a smooth behavior in time and spectral domains and can be located in a region of the sky. Therefore, the approach of this project is to model the varying source as a 4D structured signal that could be detected and restored with the appropriate approach of the data.

The developed networks will use trainings sets composed of simulated data as well as real data.

Profil du candidat :
(M/F) Post-Doc OR Research Engineer degree (>= 2 years experience in image processing)

Application domain: Applicant profile can come from various image processing and image restoration fields, such as medical imaging, astronomy, video restoration, industry, etc. Applications with a general scientific or signal processing background will be favoured.

Academic background: PhD or MsC degree in Computer Science or similar.

Formation et compétences requises :
We require candidates to have an moderate or advanced level of expertise in the following fields:
– Python/C++
– General knowledge in DL networks (e.g. CNN, GAN, Unet, mainly focused on image processing, image restoration or time series analysis etc.
– Knowledge of DL Frameworks (e.g. TensorFlow, keras, PyTorch
– (Optional) Inverse problem formulation and resolution
– (Optional) Fourier analysis and Fourier sampling
– (Optional) Data workflow management systems
– (Optional) Knowledge on edge/cloud technologies

Adresse d’emploi :
LESIA, Observatoire de Paris,
5 Place Jules Janssen,
92190 Meudon

Document attaché : 202312061248_Profil TASKA D-E-2.pdf