Computational material engineering (CME) has increased the pace of material development many folds in recent years by taking advantage of computational tools such as using simulations as replacement to experiments, and building complicated processing-structure-property relationships using advanced machine learning tools. The importance of CME is underscored by new computation-based design frameworks such as material sensitive design and integrated computational material engineering. In these material design frameworks, the goal is to optimize a specific property, and since a material’s microstructure heavily affects its properties, there is an immense gain to study and analyze the microstructure accurately and efficiently. This leads to microstructure characterization which involves the reduction of the high-dimensional structural information of a material to a reduced form which effectively captures the most salient and relevant features of the structure at the appropriate length scale. To this end, the central theme of this dissertation is to supplant the current microstructure characterization techniques for the analysis and design of microstructural material systems. The overall contributions of this work are achieved in three major research tasks carried out on three different material systems and are briefly discussed below.Spectral Density Function (SDF) is a Fourier transform-based tool that can sufficiently characterize certain quasi-random microstructure material systems by decomposing the dominant microstructural features in the physical frequency space. SDF has also been used in the design of heterogenous materials. Until recently, the use of SDF for characterization was limited to isotropic microstructures. Even after its successful utilization, there are still some gaps in understanding the shape of SDF. My first task revolves on deconvolving the complexities of SDF function by finding its relationship with physical descriptors. This works also extends the use of SDF for reconstruction of three-phase material systems. By utilizing the knowledge of SDF gained, this thesis presents an SDF-based framework for the design of quasi-random material systems which centers upon design representation using SDF parameters. The framework is further modified to deal with special case scenarios where low fidelity and high fidelity simulations can be utilized to efficiently find the optimal processing conditions of a quasi-random material system. This multi-fidelity optimization scheme is only enabled by SDF frequency enhancement tool developed in this thesis. Quite different from the quasi-random microstructure, another interesting microstructural material system is that of granular materials. It consists of grains of different sizes, orientation, and phase angles. It encapsulates a much higher dimension information than quasi-random material systems previously mentioned, and thus cannot be characterized by SDF or similar techniques. Furthermore, these granular microstructures come with their own challenges such as differentiating between two microstructures quantitatively in model validation. This is a critical challenge in calibration of additive manufacturing simulations of alloys where the microstructure difference between ground truth and simulations needs to be quantified. To address this challenge, we present a novel difference metric to quantitatively differentiate between granular microstructures based on the combination of chord length distribution and earth mover’s distance. Building on this metric, this thesis extends its use to solve the main research problem of accurately calibrating simulations with respect to experiments. This is done by developing a framework which relies on the novel dissimilarity metric along with Bayesian optimization to calibrate simulation parameters efficiently and accurately for additive manufacturing simulations. Some microstructures do not fall in any general category and can be classified as complex microstructures. For such microstructures, it is almost impossible to numerically characterize them using the traditional tools. One possibility is that of using advanced machine learning tools such as neural networks. But the cost of such tools is that they require a huge amount of data. The more complex the microstructure, the more data it would require. Unfortunately, a big constraint of material design is that the scientists have generally very limited data to work with. An alternative to training such expensive models is to use an already trained model. This is known as Transfer Learning (TL). TL involves using pretrained model for a completely different purpose than for which they were trained, thus they need to be modified for their successful implementation. The last research task explores the possibility of modifying and using transfer learning from a pre-trained VGG-19 deep neural network model to characterize the microstructure and build process-structure-property relationships for complex microstructures.

Creator

• Ghumman, Umar Farooq

DOI

• https://doi.org/10.21985/n2-xmt1-cp33

Subject

• Design

Language

• en

Alternate Identifier

• etdadmin_upload_939364
• http://dissertations.umi.com/northwestern:16322

Date created

• 2022-01-01

Resource type

• Dissertation

Rights statement

• In Copyright