ORION-B & DAOISM Related Free Software

python   BEETROOTS (BayEsian infErence with spaTial Regularization of nOisy multi-line ObservaTion mapS) — Pierre Palud   pypi-beetroots   docs-beetroots   repo-beetroots

This generic package reconstructs maps of physical conditions (e.g., temperature, gas density) from multispectral-structured cubes (e.g., maps of integrated intensities of emission lines). Built as an all-in-one tool, it has many features that make it state-of-the-art:

  • it relies on a realistic statistical modelling of the noise, that can combine additive Gaussian noise, multiplicative lognormal noise and upper limits on observations,

  • it has an already implemented interface with NNBMA to accelerate inference,

  • it runs an efficient sampling algorithm that escapes from local minima. This algorithm can be converted to optimization to get results faster,

  • it exploits spatial regularization to produce physically consistent reconstructions. This regularization is critical in low signal-to-noise regions,

  • using Bayesian sampling methods, it quantifies the uncertainties associated with the estimations which is critical to draw conclusions in the absence of ground truth,

  • it quantitatively tests the ability of the astrophysical model, the noise model and the spatial regularization to reproduce the observations. This is essential to assess the relevance of the reconstruction results, and can provide a valuable feedback for future development of astrophysical models,

  • it is well-suited for both Galactic and extragalactic observations.

The documentation provides a user-friendly series of examples to understand how the sampling algorithm works, how to interpret the inputs and outputs, and how to use this package with your own observation maps and astrophysical model.

Associated publication:   paper   Palud et al (2023)

python   InfoVar (Informative Variables) — Lucas Einig   pypi-infovar   docs-infovar   repo-infovar

This package provides tools to efficiently study the usefulness of variables on data of interest.

Associated publication:   paper   Einig, Palud et al (2024)

python   NNBMA (Neural Network-Based Model Approximation) — Lucas Einig, Pierre Palud   pypi-nnbma   docs-nnbma   repo-nnbma

This Python package offers a simple API to build, train and manipulate deep neural network in a very NumPy way. Its primary purpose is to handle numerical models approximation, but its scope of application is much broader. In particular, it was originally designed and used to derive an approximation of the Meudon PDR code, a complex astrophysical numerical code.

This package relies on PyTorch to build multilayer perceptron-based neural networks. No prior knowledge of deep learning frameworks is required to obtain results. This package offers a wide range of features for use in research:

  • it implements variations of the multilayer perceptron architecture (see this paper), suitable for approximating numerical simulations such as the Meudon PDR code,

  • it offers a turnkey solution for approximating models from anomaly-prone data,

  • it enables efficient calculation of the derivatives of model outputs with respect to inputs (gradient, Hessian matrix, or even higher-order derivatives), which allows its use in optimization (e.g., Bayesian inference procedures),

  • it proposes solutions for optimized calculation of a small number of outputs for models approximating simulations with a large number of outputs.

Associated publication:   paper   Palud, Einig et al (2023)

python   RANCH (Radio AstroNomy Cubes Handler for data science) — Lucas Einig   pypi-ranch   docs-ranch   repo-ranch

An API to easily handle radio astronomy FITS files with Python.

python   IRAM 30-meter EMIR observations informativity — Lucas Einig   docs-infobs   repo-infobs

This package implements tools to quantitatively estimate the usefulness of spectral line observations for estimating physical conditions. It provides a tool for simply reproducing observations made at IRAM 30-meter millimeter-wave telescope coupled with the EMIR receiver. Other instruments can also be simulated.

Line intensity predictions are made using a neural network emulation of the Meudon PDR code. This emulator enables a thousand predictions for more than 5 000 lines to be made in around 10 ms on a laptop, with an average error of less than 5%.

Associated publication:   paper   Einig, Palud et al (2024)

python   BeehiveNet — Lucas Einig   docs-beehivenet   repo-beehivenet

Python package for defining, training and making predictions with a machine learning model composed of several specialist artificial neural networks. This type of model was originally designed to infer astrophysical parameters in a robust and interpretable way, using neural networks dedicated to specific physical regimes.

python   LaTeX ISM emission lines — Lucas Einig, Pierre Palud   repo-latexlines

This code implements a conversion of emission lines formatted according to the convention of the Meudon PDR code as well as a number of user-friendly methods for manipulating these lines.