The Investigation Of Data Mining Based Macro-Meso Mechanics Model For Composites

In recent decades, composite material has become an important material in various industrial fields, especially in the fields of aeronautics and astronautics. Different from classical material, a composite is a material made up of two or more components in special designed structure at mesoscale. Therefore, the macroscale property of a composite is depend on the properties of components and designed structure at mesoscale. Meanwhile, the deformation and failure of composite is determined by mechanisms at both mesoscale and macroscale. Thus, the investigation of composite’s mechanical behavior at macroscale and mesoscale is a significant problem for both material design and analysis of deformation and failure. The key issue in macro-mesoscale mechanics is the relationship between macroscale parameters and mesoscale parameters. However, with the increase of design parameters, classical methods are hard to analyze these macroscale parameters and mesoscale parameters effectively. In order to solve this problem, a data mining based composite mechanics approach is proposed in this thesis. This approach can avoid complex mathematical derivation and hypothesis. Meanwhile, it can be widely used for various composite materials. Moreover, this data mining based approach can discover potential unknown knowledge from database. Therefore, the data mining macro-mesoscale model provides a novel approach in composite mechanics. The main contents included:(1) In order to generate representative volume element (RVE) model, a reconstruction program for particle reinforced composite is developed. Both random distribution reconstruction and image-based reconstruction are included in this program. Meanwhile, modules for microstructure, finite element mesh, periodic boundary conditions and cohesive elements are included in this program. Moreover, a smooth reconstruction algorithm is proposed and used in the image-based reconstruction program. Compared with classical reconstruction approach, the smooth algorithm can eliminate the rough mesh at material interface, and then provide more accurate prediction in the simulations of interface stress and crack propagation.(2) Based on the reconstruction program, the representative volume element (RVE) model for fiber reinforced composite is developed. Meanwhile, a temperature and strain rate dependent elastic-plastic constitutive law is proposed for resin matrix. The matrix constitutive law is validated by experiment with various temperature, strain rate. Furthermore, the RVE model with this constitutive law also agrees well with experimental results. Therefore, the RVE model is reliable to predict the macroscale property for the composite when the mescoscale reconstruction and constitutive laws for components are accurate.(3) A composite property prediction method based on data mining technique is proposed. In this method, the database is generated from representative volume element (RVE) models with different mesoscale design parameters. A tensor decomposition approach is adopted to reduce the dimension of database, and a curve fitting approach is utilized at lower dimension to get the quantitive relationship between mesoscale design parameters and macroscale properties. Furthermore, the relationship between multi mesoscale parameters and macroscale elastic modulus, strength and residue modulus is investigated for particle reinforced composites. Meanwhile, the effect of decomposed parameters on the results is discussed.(4) The data mining approach is extended, and a data mining based hierarchical multiscale model is proposed. In this model, non-orthogonal periodic boundary conditions and real-time homogenization method are utilized to generate database for tangent modulus at mesoscale. Meanwhile, the constitutive law at macroscale is updated according to the law discovered from the mesoscale database. The numerical results show that the data mining based hierarchical multiscale model agree well with direct numerical simulation (DNS) results, and correct classical hierarchical multiscale model.(5) The data mining based material property prediction approach is combined with dynamic piecewise exponential model for fracture problem in functionally graded composite. Using the present model, the material design parameter can be applied as input parameter directly. Moreover, the influence of graded ratio, graded form, crack length and crack location on the dynamic stress intensity factor are analyzed.