KH Machine Learning

A series of 105 Kelvin-Helmholtz simulations were preformed to train a machine learning algorithm.

Simulation Specs

In order to conserve computational resources, each case was started on a 2D x-z domain and was then remeshed to 3D after billow formation. Noise addition at the time of remeshing provides seeds for 3D instabilities. The output variables are normalized by half the velocity jump across the layer and the shear thickness. The total simulation time is 360 h/U, with 24 equally spaced 3D volume file outputs.

The following parameters were modified:

Richardson Number

0.0, 0.05, 0.1, 0.15

Reynolds Number

1000, 2000, 3000, 4000, 5000

Initial Velocity Perturbations

0.001, 0.01, 0.1

Duct

For applicable cases, namely Richardson number greater than zero, a duct was also imposed. The location and width of the duct were set the same as the shear layer: centered in the domain with a width of 1. The Reynolds number is taken at the center of the duct/shear layer. The duct ratio to the edge of the domain is 8.

Animations of vorticity magnitude in a xz plane

From left to right, the initial noise increases. Duct cases are shown directly below. Note: The time shown in animation does not have units (min) and is dimensionless.

Ri=0.00 Re=1000



Ri=0.00 Re=2000



Ri=0.00 Re=3000



Ri=0.00 Re=4000



Ri=0.00 Re=5000





Ri=0.05 Re=1000



Ri=0.05 Re=2000



Ri=0.05 Re=3000



Ri=0.05 Re=4000



Ri=0.05 Re=5000





Ri=0.10 Re=1000



Ri=0.10 Re=2000



Ri=0.10 Re=3000



Ri=0.10 Re=4000



Ri=0.10 Re=5000





Ri=0.15 Re=1000



Ri=0.15 Re=2000



Ri=0.15 Re=3000



Ri=0.15 Re=4000



Ri=0.15 Re=5000