The Woodruff School

SUMMARY

Microparticle composite thin films (MPCTFs) have applications in a diverse range of fields such as, water filtration, advanced energy applications, and medical devices, with the material composition and morphology needing to be tailored to their intended function. Due to the typical multistage processing to manufacture MPCTFs, it is difficult to create Process-Structure-Property (PSP) linkages that inform the effect of manufacturing changes or variability on the final morphology and properties. Material Informatics (MI) has been used to create predictive PSP linkages for a variety of material systems and processes, such as metals, alloys, and casting. However, MI has not been applied significantly to materials without a separation of length scales between the morphological features and the thickness which many workflows require. The objective of this work is to develop a machine learning workflow for MPCTFs, which have multiple length scales, by adapting the Material Knowledge System MI workflow to provide predictive PSP capabilities and to provide guidance of model parameters for implementation in a manufacturing setting. Microstructures generated from simulations and film dying models available in literature are used to create large datasets to train the model. To determine a microstructure property for training, flux through the resulting microstructures is calculated based on Darcy’s Law. It has been found that robust and bias free reduced order microstructure spatial models for MPCTFs that balance their ability to capture short-, mid-and long-range spatial microstructure features are key to training Property-Structure (PS) and Structure-Property (SP) models, which account for feature gradients. It has been demonstrated that the PSP workflow accounted for over 90% of the variance in the flux based upon the initial process parameters. A pathway for predicting MPCTFs structure and properties formed during slot die coating is possible with data science and spatial models, which minimizes the need for large image datasets while retaining structural information for future PSP modeling.

SUBJECT:	Ph.D. Dissertation Defense

BY:	Peter Griffiths

TIME:	Thursday, June 30, 2022, 11:00 a.m.

PLACE:	MRDC Building, 4211

TITLE:	Material Informatics Workflow for Property-Structural-Property Linkages of Particle Composite Thin Films

COMMITTEE:	Dr. Tequila Haris, Chair (Mechanical Engineering) Dr. Surya Kaladindi (Mechanical Engineering) Dr. Cyrus Aidun (Mechanical Engineering) Dr. Hamid Garmestani (Materials Science and Engineering) Dr. Isabel Escobar (Chemical and Materials Engineering, UK)

SUMMARY Microparticle composite thin films (MPCTFs) have applications in a diverse range of fields such as, water filtration, advanced energy applications, and medical devices, with the material composition and morphology needing to be tailored to their intended function. Due to the typical multistage processing to manufacture MPCTFs, it is difficult to create Process-Structure-Property (PSP) linkages that inform the effect of manufacturing changes or variability on the final morphology and properties. Material Informatics (MI) has been used to create predictive PSP linkages for a variety of material systems and processes, such as metals, alloys, and casting. However, MI has not been applied significantly to materials without a separation of length scales between the morphological features and the thickness which many workflows require. The objective of this work is to develop a machine learning workflow for MPCTFs, which have multiple length scales, by adapting the Material Knowledge System MI workflow to provide predictive PSP capabilities and to provide guidance of model parameters for implementation in a manufacturing setting. Microstructures generated from simulations and film dying models available in literature are used to create large datasets to train the model. To determine a microstructure property for training, flux through the resulting microstructures is calculated based on Darcy’s Law. It has been found that robust and bias free reduced order microstructure spatial models for MPCTFs that balance their ability to capture short-, mid-and long-range spatial microstructure features are key to training Property-Structure (PS) and Structure-Property (SP) models, which account for feature gradients. It has been demonstrated that the PSP workflow accounted for over 90% of the variance in the flux based upon the initial process parameters. A pathway for predicting MPCTFs structure and properties formed during slot die coating is possible with data science and spatial models, which minimizes the need for large image datasets while retaining structural information for future PSP modeling.