Radiative Transfer as a Bayesian Linear Regression problem

Abstract

Electromagnetic radiation plays a crucial role in various physical and chemical processes. Hence, almost all astrophysical simulations require some form of radiative transfer model. Despite many innovations in radiative transfer algorithms and their implementation, realistic radiative transfer models remain very computationally expensive, such that one often has to resort to approximate descriptions. The complexity of these models makes it difficult to assess the validity of any approximation and to quantify uncertainties on the model results. This impedes scientific rigour, in particular, when comparing models to observations, or when using their results as input for other models. We present a probabilistic numerical approach to address these issues by treating radiative transfer as a Bayesian linear regression problem. This allows us to model uncertainties on the input and output of the model with the variances of the associated probability distributions. Furthermore, this approach naturally allows us to create reduced-order radiative transfer models with a quantifiable accuracy. These are approximate solutions to exact radiative transfer models, in contrast to the exact solutions to approximate models that are often used. As a first demonstration, we derive a probabilistic version of the method of characteristics, a commonly-used technique to solve radiative transfer problems.

arXiv comment16 pages, 1 figure, accepted to Monthly Notices of the Royal Astronomical Society (MNRAS)

References in corpus (19)

• Artificial Neural Networks for Solving Ordinary and Partial Differential Equations
• The UMIST database for astrochemistry 2012
• MPI-AMRVAC 2.0 for Solar and Astrophysical Applications
• Physics Informed Neural Networks for Simulating Radiative Transfer
• Probabilistic Numerics and Uncertainty in Computations
• (Sub)stellar companions shape the winds of evolved stars
• Monte Carlo Radiative Transfer
• Bayesian Probabilistic Numerical Methods
• The statistical finite element method (statFEM) for coherent synthesis of observation data and model predictions
• Probabilistic Numerical Methods for PDE-constrained Bayesian Inverse Problems
• A Modern Retrospective on Probabilistic Numerics
• Multi-Dimensional Astrophysical Structural and Dynamical Analysis I. Development of a Nonlinear Finite Element Approach
• Reducing the complexity of chemical networks via interpretable autoencoders
• The Role of Internal Photons on the Chemistry of the Circumstellar Envelopes of AGB Stars
• Magritte, a modern software library for 3D radiative transfer: I. Non-LTE atomic and molecular line modelling
• Chemulator: Fast, accurate thermochemistry for dynamical models through emulation
• Radiation-Hydrodynamics with MPI-AMRVAC: Flux-Limited Diffusion
• Incorporating Astrochemistry into Molecular Line Modelling via Emulation
• Magritte, a modern software library for 3D radiative transfer: II. Adaptive ray-tracing, mesh construction and reduction