Optimisation de la conception et de la fabrication des materiaux composites par des algorithmes d'evolution et d'apprentissage (French text).

摘要

This research project is mainly concerned with solving optimal design and fabrication problems occurring in the field of composite materials. To tackle these issues, stochastic optimization techniques broadly inspired from natural evolution, i.e., evolutionary algorithms, are applied. Firstly, theoretical aspects of evolutionary and multi-criteria optimization are briefly exposed. Secondly, two specific multi-objective problems are considered, namely: optimal design of helical composite springs and gate location optimization in liquid composite moulding. In the case of spring design, the functions evaluation is highly prohibitive since it requires three-dimensional finite element analyses, which bear a significant computational cost. Consequently, methods borrowed from the fields of statistics and machine learning are implemented in order to have at hand a computationally lighter model, thus providing a faster convergence. Kriging, a tool developed within the field of Geostatistics, and active learning, a recent paradigm in artificial intelligence, are used for this purpose. Thereafter, NSGA-II (Non-dominated Sorting Genetic Algorithm-II) is applied on the data structures that are modelled. The analysis of the Pareto fronts reveals the optimal compromise between the chosen criteria, i.e., stiffness and mass, and the influence of constraints on this relation. In the case of optimal gate location, a fast filling algorithm is used. The latter allows predicting the filling time and the last point of the mould to be filled in a very short time compared to conventional numerical simulation. Also, a re-meshing technique is used to adapt the mesh to the presence of holes on the geometry. Different variation operators are assessed on various composite parts and an optimization module for gate location is built on the basis of the EO (Evolving Objects) library.