Flame3D.jl

FLuid And Machine learning Engine 3D

Code for compressible flow simulation with neural network for real gas and chemical reaction.

Use CUDA.jl, AMDGPU.jl and MPI.jl for multi-GPU parallelization
3D, with high order scheme (up to 7th order)
LES modeling for complex flows
Use Lux.jl trained neural network model for chemical reaction prediction
GPU kernel for fast chemical reaction evaluation
Call Cantera to parse the input file, easy to use
Call Cantera for chemical reaction evaluation, either Python or C++ interface
Use HDF5 and VTK for I/O
Use matplotlib, Makie.jl or Paraview for post-processing
Use immersed boundary method on curvilinear coordinates to represent complex geometry (work in progress)
Be able to simulate incompressible flow efficiently (work in progress)

Usage

Generate mesh and metrics with

julia parse_mesh.jl

It will compute metrics and store it to metrics.h5, if you want to visualize the mesh, make sure const vis::Bool = true, then the mesh.vts file can be opened with Paraview, etc.

Run the simulation with

julia run.jl

Or to use multi-GPU, make Nprocs=<nprocs> and run it with mpiexecjl -n <nprocs> julia run.jl.

For more details, checkout the comments in run.jl.

For post-processing, a sample code is provided in Utils/plot.jl. It can be used to visualize and analyze the result. Alternatively, Paraview or Makie.jl can also be used for 3D rendering.

Validation cases

Taylor Green Vortex

$Re=1600$ vortex breakdown with 7th order upwind scheme for advection and 4th order central scheme for viscous term. Results are compared with reference spectral method code.

0D reaction

$H_2-O_2$ reaction at $10$ atm, $1500$ K, use Li-Dryer mechanism and 1st order implicit method for ODE integration. Results are compared with cantera.