CUDA Python code accompanying 'Thermal convection and jets produced by human exhalation dominate the indoor spreading of emitted $\mathrm{CO_2}$ and aerosols'
Description
GPU-based finite-difference simulation (CUDA Python / Numba) of the spreading of exhaled $\mathrm{CO_2}$ in a closed, non-ventilated room: compressible Navier-Stokes equations coupled to heat transport, $\mathrm{CO_2}$ transport and periodic nasal breathing of a seated person. The code is written in CUDA Python and optimized for an NVIDIA RTX 5070 Ti; a full run takes about 12 hours. It accompanies the paper "Thermal convection and jets produced by human exhalation dominate the indoor spreading of emitted $\mathrm{CO_2}$ and aerosols" by L. A. Natusch and R. R. Netz.
The code is currently set up to simulate Model III of the paper (thermal convection + periodic breathing), with the parameters of the paper (room 5.75 m x 2.55 m x 2.85 m, grid spacing 2.5 cm, 1800 s simulated time).
Files:
- run_simulation.py: parameters, setup, and main run
- initialization.py: grid, initial fields, body and no-slip masks
- equations.py: GPU kernels (momentum, density, temperature, $\mathrm{CO_2}$) and the time loop
Other models of the paper:
- Model II (thermal convection, no breathing):
1. run_simulation.py: set the tidal volume V_B = 0 (no breathing velocity).
2. equations.py: activate the line "c_diff += dc / grid_num" in update_navier_stokes_gpu ($\mathrm{CO_2}$ source in front of the nostril).
(For exactly the same no-slip mask as in the paper, also use the "Model II" variant of the mask in body() in initialization.py.)
- Model I (pure diffusion):
1. run_simulation.py: set V_B = 0 and, in init_const() of initialization.py, scale the diffusion coefficient D (the paper shows D, 500 D, and 1000 D of $\mathrm{CO_2}$ in air).
2. equations.py: in main_gpu(), comment out the call of update_velocity_gpu (no velocity field is computed), and activate the line "c_diff += dc / grid_num" in update_navier_stokes_gpu.
Requirements: NVIDIA GPU with CUDA support, Python 3, numpy, scipy, numba (with CUDA support). Plotting and evaluation scripts, and the experimental data, are available from the authors on request.