Phd thesis Title : Robust Preliminary Design Optimization
Supervisor name : Jean Bigeon
Laboratory supervisor : G-SCOP
Co-Supervisor name : Gilles Foucault
Doctoral School : IMEP2
Start Date : September-October 2014
Financing – Context – Partnerships :
Description of the subject: The advances in computer-aided engineering (CAE) tools of the past two decades have changed the engineering design process. Products can be easily synthetized in virtual environment. Computational models are the key of this simulation-based engineering design paradigm. Non-linear Optimization is needed as soon as possible in the process and must also take into account uncertainties (material and geometrical parameters), trying to minimize some parameters under nonlinear constraints. Despite the impressive progress in numerical optimization algorithms and the availability of several software packages that provide platforms for linking them to CAE tools (e.g., Matlab, Isight, modeFrontier …), fundamental obstacles remains avoiding full adoption of automated optimization practice particularly:
Successful use of gradient-based optimization algorithms depends highly on the quality of the required gradient approximations
Evolutionary algorithms led to a very high number of model evaluations and are too dependent of random operators.
Simulation based engineering problems must handle inherent uncertainties (geometric and material parameters). Quite all design engineers utilizes a probabilistic approach. But quite all techniques required gradients.
Some Derivative free optimization techniques recently developed have been introduced but handle with difficulties nonlinear constraints. So we want to explore derivative free algorithms adding capacity to handle nonlinear constraints. We will formulate and solve multi-objective optimization problems to quantify tradeoffs between optimality (i.e., maximizing engineering system performance) and robustness (i.e., sensitivity of performance to variations), as opposed to single-objective formulations that maximize performance subject to probabilistic constraints that address uncertainty by means of arbitrarily-defined reliability bounds. One is an electromechanical coming from an earlier O2M-MOVEO project and has been validated. Second is coming from a collaborative project with ETS/Quebec/CANADA based on inspection of flexible parts in aeronautics by high resolution camera with software tools (which include optimization). We’ll also compare with several preliminary design models (industrials models). For these models (analytic) we can get exact gradients and we can compare with optimization techniques we have successfully developed in our team (global interval methods and PSO evolutionary methods).