Neuro-evolutionary topology optimization of structures by utilizing local state features

Abstract

In this paper we propose a novel method for the topology optimization of mechanical structures, based on a hybrid combination of a neuro-evolution with a gradient-based optimizer. Conventional gradient-based topology optimization requires problem-specific sensitivity information, however this is not available in the general case. The proposed method substitutes the analytical gradient by an artificial neural network approximation model, whose parameters are learned by an evolutionary algorithm. Advantageous is that the number of parameters in the evolutionary search is not directly coupled to the mesh of the discretized design, potentially enabling the optimization of fine discretizations. Concretely, the network maps features, obtained for each element of the discretized design, to an update signal, that is used to determine a new design. A new network is learned for every iteration of the topology optimization. The proposed method is evaluated on the minimum compliance design problem, with two different sets of features. Feasible designs are obtained, showing that the neural network is able to successfully replace analytical sensitivity information. In concluding remarks, we discuss the significant improvement that is achieved when including the strain energy as feature.