MACHINE LEARNING-BASED DESIGN OPTIMIZATION OF VARIABLE STIFFNESS FIBER REINFORCED COMPOSITES TO MINIMIZE STRUCTURAL COMPLIANCE

The variable stiffness design optimization of fiber-reinforced composite laminates optimizes the designability of fiber laying angles point by point to match the spatial variation of stress states in the structure and more efficiently exert the directionality of fiber-reinforced composite laminates in strength and stiffness performance, providing the designers with a broader design space and design flexibility. However, the traditional variable stiffness design optimization of composite material based on gradient algorithms inevitably faces large-scale computational challenges in structural and sensitivity analysis due to its large number of design variables. At the same time, there is a problem with the randomness of load conditions in the conceptual design stage of the structure, and how to formulate an efficient design scheme for random load conditions in the initial conceptual design stage has important engineering value. In recent years, with the rapid development of artificial intelligence and high-performance computing, it has become possible to build end-to-end machine learning models based on the datasets obtained by traditional optimization, which provides the possibility of achieving real-time variable stiffness optimization of composite material. In this paper, the back propagation (BP) neural network algorithm is used to establish a variable stiffness design optimization method for fiber-reinforced composites based on machine learning. Firstly, based on the normal distribution fiber optimization (NDFO) interpolation scheme, a composite material variable stiffness design optimization model is constructed with minimizing structural compliance as an objective function, and the sample datasets required for neural network model training are obtained by considering the randomness of load magnitude and direction. Secondly, the means square error (MSE) was used as the objective function to train the sample dataset using the BP neural network model. Finally, a model evaluation system based on the Pearson correlation coefficient and MSE is established to evaluate the generated neural network model. Numerical examples discuss the variable stiffness design optimization of MBB beam with round holes and C-type cantilever beam, elaborate the implementation process of the variable stiffness design optimization of composite based on machine learning, and systematically compare the differences between the variable stiffness design optimization of composite based on machine learning and the traditional variable stiffness design optimization results of composite based on NDFO in fiber laying trajectory and objective function, and verify the effectiveness of the proposed method.