neutt

neutt is the differentiable JAX/GPU port of the neutrino-transport sector of BURST: N_eff and the neutrino spectral distortions through weak decoupling, with finite-temperature QED and the flavour-mixing (QKE) refinement, and (coupled to LINX) the light-element abundances.

Installation

pip install neutt                 # CPU
pip install "neutt[cuda12]"       # NVIDIA GPU (CUDA 12 build of jaxlib)

neutt is pure Python. It needs float64, which it switches on in JAX when imported, so import neutt before creating any JAX array.

Quickstart

from neutt.decoupling import make_decoupling_solver

solve, grid = make_decoupling_solver()    # builds the collision kernels
out = solve()                             # compiles on the first call, then solves
print(float(out["neff_final"]))           # 3.043322 (Standard Model, NLO QED)

Every solver has this shape: make_*_solver(...) builds once and returns a jitted solve, whose arguments are traced, so calling it again with new values does not recompile.

Documentation

• Reference manual: docs/burst_manual.pdf, covering every module, option and solver, with the physics and the numerical-design rationale.
• API reference: the Sphinx site built from docs/sphinx/ (pip install "neutt[docs]", then sphinx-build docs/sphinx docs/sphinx/_build).
• Quickstart: docs/USER_MANUAL.md.

Releasing

.github/workflows/release.yml builds the sdist and wheel on every v* tag and publishes them to PyPI by trusted publishing. Register this repository on PyPI as a trusted publisher for neutt (workflow release.yml, environment pypi), set the version in neutt/__init__.py, then push a tag such as v0.1.0.

License

MIT (see LICENSE).