Hybrid neural method for radiative transfer modelling in astrophysics

Supervision

Pierre Ocvirk (OBAS, Strasbourg)

Laboratory and team

OBAS, Strasbourg - Team "CDS"

Subject description

This project develops a hybrid approach combining classical numerical methods with neural networks to efficiently simulate radiative transfer in astrophysical simulations, particularly for the Epoch of Reionization. Current models like M1 compromise between precision and computational cost but have physical fidelity limitations. The project extends the Micro-Macro methodology developed for the M1 model using Physics-Informed Neural Networks (PINNs). It aims to design an optimal neural network architecture that captures radiative transfer physics by exploiting symmetries and physical constraints. The approach will be systematically evaluated against traditional methods and applied to classical tests from the literature. The three-year plan includes: (1) mathematical formulation of the micro-macro problem and implementation in 2D using the SCIMBA framework; (2) extension to 3D and development of physical regularization terms; (3) exploration of coupling strategies with Dyablo (https://github.com/Dyablo-HPC/Dyablo) and parallelization for massive simulations. The project emphasizes computational efficiency, documentation, and usability to ensure adoption by the scientific community.

Related mathematical skills

The candidate must be proficient in partial differential equations and their numerical analysis. Skills in machine learning, particularly neural networks with PyTorch/TensorFlow libraries, are essential. Familiarity with kinetic equations and radiative transfer models will be appreciated. Experience in scientific computing and parallel programming is desirable. Basic knowledge of numerical astrophysics, although not essential, would be an asset. The candidate should demonstrate the ability to combine mathematical theory with efficient numerical implementation.

IRMIA++ is one of the 15 Interdisciplinary Thematic Institute (ITI) of the University of Strasbourg. It brings together a research cluster and a master-doctorate training program, relying on 12 research teams and 9 master tracks.

It encompasses all mathematicians at Université de Strasbourg, with partners in computer science and physics. ITI IRMIA++ builds on the internationally renowned research in mathematics in Strasbourg, and its well-established links with the socio-economic environment. It promotes interdisciplinary academic collaborations and industrial partnerships.

A core part of the IRMIA++ mission is to realize high-level training through integrated master-PhD tracks over 5 years, with common actions fostering an interdisciplinary culture, such as joint projects, new courses and workshops around mathematics and its interactions.

Selection will rely on the professional project of the candidate, his/her interest for interdisciplinarity and academic results.