Image based Observations

Note

This functionalty has been reworked and the documentation is not up to date. Please check the schema for the renamed/relocated options and functionality.

Note

The solver used in this example is not openly available. The functionality after creating the image should work with any solver.

The framework has now the ability to not only use the spline coordinates as the observations but it can also use image based observations. These use flow field information directly from the solver to create a false color image of the flow field. This means that this feature is currently only available in conjunction with the XNS solver, from CATS RWTH Aachen. Other solver would need to either do one of these options:

  1. export the results to the same location and in the same format

  2. the image exporter in the framework would need to be extended to this solver specifically.

  3. add a SPOR step which takes the solver results and

converts them into the XNS format and to the correct location

Here I will assume the framework can read the results correctly.

Activating of the options

To use the image based observations two options need to be changed in the json file. But these can have some consequential changes. Please think through them.

The two main changes are:

  1. environment.use_cnn_observations = true

  2. agent.policy = [“CnnPolicy”, “MultiInputPolicy”]

The second change is either the on evalue or the other depending on wheter only image based observations are present (CnnPolicy) or also other observations are present (MultiInputPolicy).

In the channel test case, used while developing and validating this framework, only the image observations were necessary. A second test case using the geometry of a t-junction at least a second value had to be also added. Therefore the MultiInputPolicy was needed in the second case. Please check the SPORObjects you have defined if they produce additional observations. (A nice hack in this regard is that if the additional observations are set to 0, the framework does not check if observations are set, even if there are some they will be ignored)

Flow Field Image creation

The way the flow field image is created is by combining the gray scale contour representations of the x-,y- velocity field and that of the pressure field. Each gray scale image is used as one color channel of the resulting image.

A respresentation of this is shown in the following image.

asd