Heterogeneous materials have been emerging and playing essential roles in various engineering and scientific fields. They usually include multiple phases of materials to create unique properties that are not accessible to their homogeneous counterparts. The traditional design approach in the material science community is to use trial-and-error iteratively, which is bounded by its high costs and time-consuming procedure of new materials synthesis. It also largely depends on the researchers’ experiences to choose new designs and thus becomes less effective when the design space is vast. There are three significant challenges in the design of heterogeneous materials system: design representation, design evaluation, and design synthesis. This dissertation addresses these issues associated with advanced heterogeneous materials design by developing a data-driven computational framework.', 'A spectral density function based (SDF) approach is established to characterize and reconstruct the complex quasi-random microstructures observed in heterogeneous materials systems, which allows convenient parameterization and reduction of the high dimensional structural design space. We covered three different methods for simulating microstructures based on SDF: the phase retrieval algorithm for 3D reconstruction of isotropic microstructure from 2D images; Gaussian random field (GRF) based approach for the fast generation of channel-type microstructures; the disk-packing algorithm for particle-based microstructures. The SDF based design representation method is the foundation for the computational design framework.', 'Machine learning (ML) models are necessary to mine existing materials related data for discovering structure-property relationship as well as reducing the dimensions of materials data. A novel structural equation model (SEM) based method is developed in this dissertation to identify essential microstructure descriptors and build an appropriate structure-property relationship for material properties prediction. The SEM-based method combines feature extraction and feature selection to discover underlying patterns of descriptors as well as selecting physically meaningful design variables. Additionally, the method successfully addresses the redundancy issue in existing importance ranking method.', 'In heterogeneous materials design, it is commonly needed to optimize quantitative design variables and qualitative material factors simultaneously. Therefore, we introduce a novel latent variable (LV) approach to Gaussian process (GP) models for handling both quantitative and qualitative input variables. Existing GP methods for handling this mainly assume a different response surface for each combination of levels of the qualitative factors and relate them via a multi-response cross-covariance matrix. The proposed method maps each qualitative factor to an underlying numerical latent variable (LV), with the mapped value for each level estimated similarly to the covariance length-scale parameters. This provides a parsimonious GP parameterization that treats qualitative factors the same as numerical variables and views them as affecting the response via similar physical mechanisms. We Observed superior predictive performance of the LV approach across a variety of examples. Moreover, the mapped LVs provide substantial design insight into the nature and effects of qualitative factors.', 'Finally, we propose a Bayesian optimization (BO) method for heterogeneous materials design, in which the SDF approach is used for representing design space of microstructures, and the latent variable GP model is employed as the surrogate model for further statistical inferences. With appropriate sequential sampling criteria, the BO material design flow effectively balances the exploitation of existing data and exploration of unknown feasible design space. Via two benchmark mixed-variable design problems: high-performance light absorption solar cell and searching the combinatorial space of hybrid organic-inorganic perovskite, we demonstrated that the proposed BO method consistently outperforms state-of-the-art BO method in dealing with design optimization problems involving both quantitative and qualitative input variables.

Last modified

• 10/28/2019

Creator

• Yichi Zhang

DOI

• https://doi.org/10.21985/n2-a5rq-4c31

Keyword

• Bayesian Optimization
• Materials Design
• Response Surface Model
• Machine Learning
• Mixed Variable

Resource type

• Dissertation

Rights statement

• In Copyright