IMPROVED DESIGN SPACE EXPLORATION BY MACHINE LEARNING ESTIMATION FOR A PARAMETRIC TURBOFAN MODEL

In the preliminary design phase, selecting an architecture requires understanding the influence of noise parameters while exploring the design range. Given that architecture simulations are slow, the normal approach is to generate meta-models or fast running models that are sufficiently accurate. These metamodels are used to efficiently select robust designs by quickly estimating the impact of noise distributions for different design independents. The creation of these meta- models is efficiently accomplished by designed experiments. Predetermined runs are decided by factorial experiments, Latin hypercube designs, or sequential experimentation. These simulation runs have combinations of independents and noise parameters that are outside of previous experience. Without accurate initial estimates the simulations may not converge or have long simulation times. A method is presented to improve these simulations through more effective initial estimates. The estimates are created by a data-based, machine learning, model that is created from previous successful runs. This additional model is used to make accurate initial condition guesses. By using a chained linear regression model, the initial conditions for half the parameters are improved compared with using the previous run. The resulting meta-model can be used for additional design studies. It is also shown that improvement may not translate into significant run time differences as single parameter estimates may drive convergence.