Rayleigh-Bènard Convection Simulation

Setup

We use uv to manage python dependencies. Install virtual environment via:

uv sync
source .venv/bin/activate

The simulation is based on Julia. Please install according to their docs or via their standalone installer:

curl -fsSL https://install.julialang.org | sh

Get started

Run plain environment

Small example to run a 2D Rayleigh-Benard Convection simulation wrapped in a gym environment:

python example/run.py

Run in parallel utilizing vectorized envs

python example/run_vectorized.py

Run with wrappers

python example/run_wrappers.py

Gym environment

env = gym.make(
        "rbc_gym/RayleighBenardConvection2D-v0",
        render_mode="human",
    )

Roadmap

• Timing analysis

• vectorized environments

• Wrappers

• Checkpoints

• Make installable

• Normalize Reward Wrapper

• Reward Shaping Wrapper

• generate Checkpoints for train and test for all Ra

• Docs

• Make more parameters accessible to the gym env

• GPU compatibility

• Create Dataset from python gym

• PyTests

• 3D simulation

• Multi-Agent Environment via PettingZoo

Time efficiency

Time comparison vs a shenfun based simulation. Averaged over 1000 iterations on an apple sillicone based system.

Julia based simulation

Julia init time: 7.68 seconds
Average time to reset env: 0.02 seconds
Average time to step one timestep (dt=1): 0.12 seconds (sim dt=0.03)
Average time to step with rendering: 0.12 seconds

Shenfun based simulation

Shenfun init time: 0.00 seconds
Average time to reset env: 0.19 seconds
Average time to step one timestep (dt=1): 0.56 seconds (sim dt=0.03)