Data Centric Design for Microstructural Materials Systems

Iyer, Akshay

doi:https://doi.org/10.21985/n2-yq7y-s936

Work

Data Centric Design for Microstructural Materials Systems

Public

Download PDF

Download All Files (.zip)

Materials science has been central to human advancement since time immemorial. There has always been curiosity around studying the processes required to extract materials, examine their structure, and ultimately tailor their properties to meet human needs. Over the last few centuries, the ability to tailor material properties was driven by design rules identified via experimentation, theoretical analysis, and more recently computational capabilities. It is only over the last decade that we have realized the immense potential of data driven materials discovery. This dissertation further examines this new paradigm of material discovery through the lens of design engineering. We show that the decision made during material design process – design representation, design evaluation and design synthesis are informed by the process-structure-property knowledge contained in material databases. Through a variety of advanced material systems, we seek to address some challenges arising at the intersection of design engineering and material science.We investigate the design representation challenges arising in microstructure design. Spectral Density Function (SDF), a frequency domain microstructure approach is the focus of our study. We present a computational microstructure design framework for Organic Photovoltaic Cells (OPVC) using SDF and a novel structure-property simulation model. After identifying that there is a lack of microstructure representation and design methodologies for anisotropic microstructures, we demonstrate that SDF is capable of capturing the necessary information. Since design requires reconstruction of microstructures, we present a novel methodology to reconstruct isotropic and anisotropic microstructures. Our method is also computationally efficient than the existing ones. Finally, we show that this capability is useful for designing the active layer of OPVCs that outperform their isotropic contemporaries. The ability to design microstructures is useful for a wide variety of material systems, including polymer nanocomposites. In addition to the microstructure, the choice of constituents (polymer, filler and the filler’s surface modification) have a significant influence on behavior of nanocomposites. Consequently, we cast the nanocomposite design as a mixed variable, multicriteria optimization problem and leverage Latent Variable Gaussian Processes (LVGP) and Bayesian Optimization (BO) to identify Pareto optimal candidates for electric insulation. This design methodology involves usage of experimental datasets for calibrating physics models, training property prediction models as well identifying the bounds for design variables. The material properties are seldom determined completely by the composition. One such example is the metal insulator transition (MIT) compounds which display abrupt changes in their resistivity. To make them viable as next generation microelectronic devices, there is a growing interest in identifying compositions that simultaneously induce a large bandgap and high stability. We show that this combinatorial multicriteria optimization can be solved efficiently using LVGP and BO. LVGP allows us the circumvent the conventional feature engineering stage of design process which is extremely challenging for MIT due to limited understanding of the underlying physics. Although qualitative variables encountered in nanocomposite and MIT design have few levels, some material systems may involve high dimensional qualitative variables i.e., with many levels. This scenario poses a significant increase in computational cost of initiating BO since each level of every qualitative variable must be observed at least once for its latent variables to be estimated by LVGP. To this end, we develop a descriptor aided BO methodology that allows us to initiate BO with a small dataset (~O(1)) and parsimoniously predict latent variables for unobserved levels. The method is inspired by the belief that effect of qualitative variables is described by underlying numerical descriptors. Through a variety of examples, we outline the efficacy of our method in tackling several scenarios of partial and imperfect descriptor knowledge encountered in real world applications. While the critical role of microstructure in material design is acknowledged in the research community, the computational Microstructure Characterization and Reconstruction techniques are not easily accessible. To this end, we have developed eight webtools with friendly graphical user interface in NanoMine to allow users to analyze their microstructural images with only a few clicks of the button. Through a variety of material systems, this thesis exemplifies the strong confluence of material science with design engineering as outlined in the data centric design framework. With ever increasing focus on large scale data collection and analysis, we believe this framework serves as a guide to researchers for identifying critical tasks vis-à-vis data collection, method selection and method development required in the materials design process.

Creator