GitHub

AgentMach.jl

docs | dev CI Codecov Coveralls License: MIT

AgentMach.jl is a playground for building high-performance computational fluid dynamics (CFD) solvers in Julia. The package assembles structured meshes, configurable physics kernels, and explicit time integration building blocks into a composable toolkit that targets both CPU and GPU execution and scales to distributed-memory systems.

Highlights

  • Second-order finite-volume discretisations on 2D structured meshes with periodic or Dirichlet boundary conditions.
  • Multiple equation sets: linear scalar advection for algorithm development and the compressible Euler equations as a stepping stone toward Navier–Stokes.
  • Explicit two-stage Runge–Kutta (Heun) time integration with CFL-aware step control helpers.
  • Portable kernels designed to run on Julia arrays today and on accelerators via KernelAbstractions.jl in forthcoming releases.
  • Simulation drivers with callback hooks, diagnostics, and CSV/visualisation utilities that streamline benchmarking and validation.

Architecture Overview

AgentMach keeps mesh generation, physics descriptions, flux evaluations, and time integration decoupled so the components can be recombined when experimenting with new physics or backends. The top-level module simply re-exports the most useful types and functions; you can construct meshes and PDE problems directly and pass them to the provided integrators or embed them within your own driver code.

For a quick tour of the API, see the Getting Started guide. Ready-to-run scripts that reproduce the figures in the repository live in examples/; the scripts report diagnostics so you can verify behaviour as you experiment with mesh resolution, CFL targets, or higher-order fluxes.

Roadmap

The current release focuses on CPU execution and correctness-tested kernels. Upcoming work centres on:

  • lifting the existing kernels onto KernelAbstractions for GPU execution;
  • wiring MPI-aware distributed arrays into the same API; and
  • extending the physics catalogue toward viscous Navier–Stokes flows backed by validation test cases.

If you would like to contribute or report an issue, please open a discussion on GitHub. The Authors page lists the current maintainers.

Provenance

This repository was created end-to-end using gpt-5-codex.