A Data-driven Approach To Real-time Structural Topology Optimization

Structural topology optimization is an innovative structural design method,which has made major breakthroughs in the fields of aerospace,automobile and ship.However,when dealing with practical engineering problems,the traditional implicit topology optimization method has the disadvantages of a large number of structural design variables,fuzzy structural boundaries,difficult production processing,and large calculations.Therefore,using data-driven methods to do real-time topology optimization design has become a development trend in the field of structural topology optimization in the future.Currently,machine learning methods are developing rapidly,and their training efficiency and prediction accuracy have been greatly improved.This research thesis combines machine learning algorithm with the moving morphable component framework.This method can realize the real-time structural topology optimization design by extracting the potentially decisive features in the data.The results predicted by machine learning can be used as the initial condition of the topology optimization method,and achieve the effect of rapid convergence.The specific research contents are as follows:First,manually extract structural features.Although the topology optimization method based on moving morphable component framework has the natural advantage of fewer structural design variables,the layout and arrangement between the components are extremely complicated,and there may be the situations such as coverage,overlap,and juxtaposition between components.Therefore,the actual number of design variables is smaller than the number of design variables required by moving morphable component framework.By sequentially drawing the position of each component in the design domain,the arrangement between the components can be judged,and the design variables of the covered component are eliminated to reduce the design variables.Second,extract structural features by the principal component analysis method.Compared with the manual method of extracting the features of the structure,the principal component analysis method has a strict mathematical foundation,which reduces the data set to a suitable dimension through principal component analysis,thereby reducing the complexity of the machine learning algorithm and improving the prediction accuracy.Third,extract structural features by the auto encoder.Auto encoder is a very basic deep learning algorithm.In terms of structure,an auto encoder is a fully connected neural network with a symmetry structure,which is mainly used for feature extraction in data.Due to the particularity of its structure,the features extracted by the auto encoder are learned by the neural network algorithm itself,and there is no longer any subjective participation.The experimental results show that the features extracted by the auto encoder are effective,and the optimal topology structure predicted on the test set has a huge improvement in performance compared to the structure predicted by the features extracted by the above two methods.