Engineering applications of metamodel-based optimization: genetic algorithms coupled with artificial neural networks

NADIA ROMAN

Buenos Aires

Otro; Machine Learning Summer School 2018; 2018

Universidad Torcuato Di Tella

The complexity of optimization problems in engineering leads to the necessity of developing methodologies to replace the usually time-costly computational models that simulation-based optimization requires. An alternative is to implement a metamodel (model of the model), for example artificial neural networks, to evaluate the objective function of the problem. This technique, called metamodel-based optimization, has become frequently used in recent years due to the fact that it can be implemented to solve large nonlinear problems, independently of the nature of the variables involved (continuous, discrete, binary, to name but a few). Artificial neural networks are, regardless of them being an approximation method, reliable and accurate models that can be implemented on several engineering problems.In this work, three engineering applications are introduced: the optimization of the composite laminate of horizontal wind turbine blades, the design and optimization of deflectors for a Savonious type vertical wind turbine and the multi-objective optimization of the energy efficiency in residential buildings. In the three cases, an artificial neural network metamodel was developed in order to replace the simulations required to evaluate the responses of the models. These metamodels were coupled to a genetic algorithm to solve the optimization problem.