Selecting the best material to deliver optimum performance in real-world applications is one of the most significant challenges in engineering. Hundreds of thousands of computationally-predicted, but experimentally unexplored materials exist today in the public inorganic material databases as candidates for consideration. This thesis discusses three projects in the domain of materials selection and discovery, and in each of them, one or more materials with a desired set of properties are identified from a large pool of candidates.The first work describes the computational discovery of three high-dielectric, high-bandgap materials within 17 selections from a set of more than 11,000 candidate materials obtained from the Open Quantum Materials Database (OQMD). We built statistical machine learning (ML) models from a sub-dataset of the Materials Project high throughput database to predict dielectric values along with the associated model uncertainty. The final material selections are made using a statistical optimization algorithm, and the final validations are done using expensive first-principles calculations to compute the dielectric properties. The second project details the identification of a new bridge material for MoS2-based 2D electronic inks that acts as an adhesive between the 2D ink nanoparticles without interfering with the ink’s electronic properties. This project uses a sequential selection workflow incorporating machine learning-aided high-throughput heuristic modeling to select the best material from a candidate set of more than 2000 materials, and subsequent estimation of the charge-transport properties from expensive atomistic simulations. In the third project, we created a machine learning model that can identify the semiconductors and insulators which are misclassified from lower-accuracy Density Functional Theory (DFT) calculations to be metallic. The accuracy of bandgaps computed using DFT is dependent on the functionals chosen to describe the exchange-correlation energy of electron interactions. The PBE functional results in less accurate, but significantly cheaper estimations of the bandgaps compared to using the HSE hybrid functional. Our ML model predicts the bandgaps at an accuracy level of DFT-HSE at the cost of doing a cheaper DFT-PBE calculation. The reliability of ML predictions is analyzed from quantified model uncertainties and extensive literature surveys.

Creator

• Gopakumar, Abhijith Mini

DOI

• https://doi.org/10.21985/n2-yrkw-zw66

Subject

• Materials Science

Language

• en

Alternate Identifier

• etdadmin_upload_947056
• http://dissertations.umi.com/northwestern:16368

Keyword

• statistical modeling
• materials design
• uncertainty quantification
• materials selection
• computational materials
• materials informatics

Date created

• 2022-01-01

Resource type

• Dissertation

Rights statement

• In Copyright