Welcome to Lightweaver’s documentation!

Contents:

Introduction

Lightweaver is a Non-LTE (NLTE) radiative transfer framework for constructing simulations of optically thick spectral lines (primarily solar). It is available under the MIT license. The flexible nature of its design allows it to be simply used in direct line synthesis or as a component of more involved frameworks (such as radiation hydrodynamics or inversion packages). Currently, one dimensional plane-parallel, and two-dimensional Cartesian atmospheres are supported, but three-dimensional setups could be added simply with a specific formal solver. It is well tested against RH and SNAPI.

The framework is managed from python (3.8+) whilst calling into an optimised C++ backend. Whilst following a similar methods, Lightweaver is often considerably faster than RH on non-trivial problems, whilst being slightly slower on trivial single-threaded problems due to the increased initialisation time the python layer brings with it. The backend can parallelise directly over as many threads as desired on single address space systems. MPI parallelisation across nodes can be accomplished manually on top of the framework using the packages available in python.

Whilst the core numerics are implented in C++, as much of the non-performance critical code as possible is implemented in Python, and the code currently only has a Python interface (provided through a Cython binding module). These bindings could be rewritten in another language, so long as the same information can be provided to the C++ core.

The aim of Lightweaver is to provide an NLTE Framework, rather than a “code”. That is to say, it should be more maleable, and provide easier access to experimentation, with most forms of experimentation (unless one wants to play with formal solvers or iteration schemes), being available directly from python. Formal solvers that comply with the interface defined in Lightweaver can be compiled into separate shared libraries and then loaded at runtime. The preceding concepts are inspired by the well-recieved machine learning frameworks such as PyTorch and Tensorflow.

Installation

The most recent release of Lightweaver will always be available on PyPI and is pre-compiled for Linux, Windows, and macOS (intel).

Lightweaver requires python 3.8+, and it is recommended to be run inside a virtual environment using conda.

In this case a new virtual environment can be created with:

conda create -n Lightweaver python=3.8

then, activate the environment:

conda activate Lightweaver

and Lightweaver can then be installed with

python -m pip install lightweaver

On other platforms you will need to compile the library, which will require a C++17 compiler on your path, at which point setuptools should handle everything. After downloading the release, or mainline version that you wish to install, you should be able to

python -m pip install .

from the Lightweaver directory created. You may also wish to do this to create a more optimised version for modern machines; the PyPI versions support Sandy Bridge CPUs, so newer machines may have wider instruction sets available. For doing a “development” installation for working inplace on the project, you may with to use python -m pip install -e ..

Indices and tables