Machine Learning Data Generation Dynamics Emulator

Generate the netcdf file

Use the following const.h, really just make sure ord and tord are both 3.

...
int constexpr ord      = 3;
int constexpr tord     = 3;
...

Use the following input.txt:

// Input parameters for the model
// Legal syntax is key : value
// You can end input key:value lines with comments

nx         : 200        // Number of cells in the x-direction
ny         : 1          // Number of cells in the y-direction
nz         : 100        // Number of cells in the z-direction

xlen       : 20000.     // Domain length in the x-direction (m)
ylen       : 20000.     // Domain length in the y-direction (m)
zlen       : 10000.     // Domain length in the z-direction (m)

cfl        : 0.8        // CFL value to use
simLength  : 2000.      // How long to simulation (s)

parNx      : 1          // Number of parallel tasks in the x-direction
parNy      : 1          // Number of parallel tasks in the y-direction

outFreq    : 0.0001     // Frequency for output in seconds

doWeno     : 1          // Whether to do WENO

strakaVisc : 0          // Whether to apply Straka 2nd-order viscosity

// Initial data to use
//   thermal  : Rising thermal (neutral atmosphere)
//   collision: Colliding cold and warm bubbles (neutral atmosphere)
//   straka   : Falling cold bubble
dataInit : collision

// Temporal discretization to use:
//   ADER: ADER with Differential Transforms  (max cfl = 1)
//   SSPRK3: optimal 3-stage, 3rd order SSPRK (max cfl = 1)
timeMethod : ADER

// Equation set to use:
//   theta_cons: q={rho,rho*u,rho*v,rho*w,rho*theta}, conservative
//   theta_prim: q={rho,u,v,w,theta}, non-conservative
eqnSet : theta_cons

• make -j
• ./cloudFV
• # Install NCO
• ncks -d t,7500,8500 output.nc ml_training.nc

Process Training Data in python

• Read in density, u, w, and theta using the python netCDF4 package
• Randomly choose a set of N cases to train, each from a random cell of the domain [1,198] x [1,98] (to avoid the edge cells) and a random time step of the domain [0,nt-1].
  • For this sample:
    • Save density, u, w, and theta in a 3x3-cell stencil centered on the cell in question as input
    • Save density, u, w, and theta at the next time step only in the center cell as output