CEASIOMpy

SMTrain

Categories: Optimisation, Surrogate Modeling

SMTrain is a module to train a surrogate model, with data computed using CEASIOMpy’s Modules.

Notion of fidelity in CFD: - 1st-level: potential flow - 2nd-level: euler flow - 3d-level: rans flow

The computed forces and moments are getting more and more accurate through each levels (in practice).

In CEASIOMpy currently you can use: - AVL for the 1st-level - SU2 Euler for the 2nd-level - Nothing for the 3rd-level (but there is going to be an update soon with CPACS2GMSH)

Inputs

SMTrain takes as input a CPACS file.

Workflow

SMTrain’s workflow in a nutshell:

1. DATA for training:
   • uniform sampling (n_samples parameter)
   • ceasiompy.db data [optional]
2. Training:
   • Trains on 1st-level
     • generates data using PyAVL module with the uniform-sampled parameters
     • augments the dataset with data from ceasiompy.db
     • trains solely on AVL results
   • Trains on 2nd-level:
     • using SU2Run module
     • trains on data with high-variance points from the 1st-level in a loop until the rmse error is small enough
   • Trains on 3rd-level (Not yet implemented due to CPACS2GMSH status)
3. Saves model and all results in an aeromap

Installation or requirements

SMTrain is a native CEASIOMpy module, hence it is available and installed by default. It is base on the SMT: Surrogate Modeling Toolbox

Limitations

1. Can not choose the range for the Hyper-parameters of the surrogate model

More information

• SMT: Surrogate Modeling Toolbox

This site is open source. Improve this page.