Integrated Computational Materials Engineering (ICME) has emerged as a transformative paradigm that enables the co-design of products, materials, and their manufacturing processes. A gamut of computational tools, such as manufacturing process simulations and multiscale materials analyses, have been developed henceforth by their respective research communities. However, it remains challenging to seamlessly integrate these models from multiple domains, efficiently evaluate the integrated models, and perform design optimization under uncertainties arising from the stochasticity of material microstructures and lack of simulation or experiment data. Using carbon fiber reinforced polymer (CFRP) composites as examples, an ICME framework is developed in this dissertation to realize integrated material and structural design with heterogeneous microstructures. The major contributions include image-based nonhomogeneous random CFRP microstructure reconstruction algorithms for uncertainty quantification, data-driven nonlinear material models for arbitrary microstructures and loading cases for efficient uncertainty propagation, integrated predictive simulation models linking stochastic microstructures to uncertain material properties, and a concurrent process and material selection and optimization framework combining latent-variable Gaussian processes (LVGPs) and Bayesian Optimization (BO). These developments pave the way for accomplishing concurrent materials, product, and manufacturing process design in a truly integrated yet efficient manner. We start from image-based microstructural uncertainty quantification (UQ) and propagation framework to study the variability of the nonhomogeneous CFRP microstructures and their impact on part performance. Unlike traditional approaches that model the descriptors of a microstructure as scalars, we directly describe the complete three-dimensional microstructures via random fields by coupling one- and two-dimensional random processes. The microstructure features are modeled with information characterized from microscopic images, and jointly sampled with their correlations considered. Our statistical models are nonparametric and do not require presumed parametric probability distributions such as Gaussian or lognormal. Realistic microstructure samples are randomly generated from the random field and they are assigned to finite element models for structural performance simulations. Stochastic nonlinear and failure behavior are studied under loading. With the simulated structural performance data, real-time failure probability prediction and failure forecast are achieved by inferring from uncertain microstructures, which are further inferred from structure deformation measurements. After stochastic characterization and reconstruction of microstructures are accomplished, microstructure-guided deep material network (MGDMN), a deep-learning-based material modeling framework, is established to efficiently model the nonlinear mechanical behavior of composite materials considering its constituent plasticity for any microstructure phase configurations described by descriptors such as volume fraction and orientation. Our work is built upon the deep material network (DMN), a machine learning model for nonlinear material modeling for a specific microstructure. To generalize the DMN method from considering only a given microstructure, we develop the MGDMN approach that generates a DMN model’s parameters for any new microstructure by directly interpolating those from a few carefully selected trained DMNs. The creation of new microstructures’ DMNs does not require new data collection and optimization-based model training; therefore, significant savings of computational cost can be achieved when more microstructures are to be simulated. Our framework enables new studies of nonlinear material behavior of complex microstructures. We apply MGDMN to short fiber reinforced polymer (SFRP) composites modeling with matrix plasticity and show that it is efficient enough to enable a Monte-Carlo-based UQ of material properties from stochastic microstructure morphologies and uncertain constituent properties. To account for the manufacturing-induced microstructure heterogeneity in structural components, process simulations, material models, and finite-element-based structural analyses are seamlessly integrated for ICME predictive modeling. This multi-domain integration enables the mapping between processing, structure, properties, and performance of materials and structures. Latent-variable Gaussian processes (LVGP) are applied to emulate the integrated simulations so that we can efficiently explore the resulting design space spanned by process and material design variables. This mixed-variable surrogate model maps designs from a discontinuous space, a common case in ICME due to the inclusion of ordinal and categorical variables such as material selections, to any performance metric of our interest. We demonstrate the benefits of this integration and emulation via two examples involving multi-material/multi-component and multiple process conditions design, respectively. Finally, we actualize the mixed-variable and multi-objective process and material design by incorporating LVGP and constrained optimization methods into Bayesian optimization, an efficient global optimization approach. The resulting cBO-LVGP (constrained Bayesian optimization with LVGP) method allows us to obtain the optimal selection and design of materials and processes in ICME towards multiple (and often conflicting) objectives by sequentially sampling promising designs as estimated by LVGP. The two exemplary integrated simulations are utilized to demonstrate how cBO-LVGP effectively searches for optimal solutions in the mixed-variable and constrained design space and realizes integrated material and process selection and design as a result. Product design with enhanced performance, improved manufacturability, and accelerated material development cycle is achieved with our approaches developed for ICME.

Creator

• Huang, Tianyu

DOI

• https://doi.org/10.21985/n2-a7qy-sk68

Subject

• Mechanical engineering

Language

• en

Alternate Identifier

• etdadmin_upload_781131
• http://dissertations.umi.com/northwestern:15424

Date created

• 2020-01-01

Resource type

• Dissertation

Rights statement

• In Copyright