RESEARCH ON PARAMETER IDENTIFICATION OF COMPOSITE MATERIALS BY COMBINATION OF SELF-CONSISTENT CLUSTER ANALYSIS, NEURAL NETWORK AND BAYESIAN OPTIMIZATION

At present, the forward and inverse identification of heterogeneous composite materials is still faced with the dilemmas of high forward calculation cost and low universality of inverse identification. On the one hand, the data-driven computational homogenization method utilizes advanced algorithms of data science to reduce the number of variables in the governing equation, and on the other hand, it establishes the connection between the designed structure and the equivalent parameters, thereby significantly improving the computational efficiency and exploring the internal correlation between parameters. In this paper, a data-driven self-consistent clustering analysis (SCA) method is developed to classify clusters according to the strain concentration tensor of each grid point, and the discrete Lippmann-Schwinger equation is solved on the cluster region. The calculation degree of freedom is greatly reduced, and the equivalent Youngs’ modulus, thermal expansion coefficient, thermal conductivity and other parameters are efficiently obtained. However, the efficiency of SCA method is slightly insufficient when dealing with a large number of different structural conditions, so this paper further utilizes artificial neural network (ANN) as a proxy model to accelerate the calculation and achieve rapid prediction of equivalent parameters under different conditions. For the inverse problem of identify materials and structures, Bayesian optimization is combined to identify the most optimized materials and geometric structures under required equivalent parameters, forming a joint recognition framework of self-consistent clustering analysis-artificial neural networks-Bayesian optimization. In this paper, superconducting EAS strands and particle reinforced composites are taken as examples, and comparative analyses of the joint identification framework with existing experimental and numerical results are conducted. Then, the aspects of computational accuracy, computational efficiency, model hyperparameter selection, sensitivity analysis and inverse validation of the presented framework are discussed, which help us understand its advantages and shortcomings. Finally, the presented framework could provide ideas and guidelines for the development of higher precision and wider application of composite material parameter identification methods.