Accelerating the Discovery of Materials for Energy Capture and Energy Storage Using Quantum Chemistry and Machine Learning
出版項
Ann Arbor : ProQuest Dissertations & Theses, 2023
說明
145 p
附註
Source: Dissertations Abstracts International, Volume: 84-09, Section: B
Advisor: Sargent, Ted
Thesis (Ph.D.)--University of Toronto (Canada), 2023
This item must not be sold to any third party vendors
In order to reduce the global dependence on fossil fuels it is critical that we develop new materials to efficiently capture and store energy. The enormous size of the material space poses a formidable challenge to the discovery of energy materials. In this thesis, I develop machine learning (ML) approaches and search algorithms that enable accelerated discovery of energy materials.First, I develop a material representation that enables relaxed crystal properties prediction from unrelaxed crystal structures using ML. The results show that the approach accelerates the discovery of mixed halide perovskites by a factor of 105 over conventional computational screening approaches based on Density Functional Theory (DFT). I apply the method to two material discovery problems: the search for photovoltaic materials with small bandgaps; and UV and IR light-emitting materials.Next, I generalize the above-mentioned approach to tackle one of the main criticisms of ML-based approaches: their black-box nature i.e., lack of interpretability. I do this by training accurate ML models that can predict energy above hull, bandgaps, and the nature of bandgaps among different material classes using unrelaxed crystal structures. With the combination of statistical techniques and simpler supervised learning methods, I develop a method to extract scientific insights from trained ML models and use these generated insights to design new, stable UV light emitting materials and direct bandgap semiconductors.Finally, I develop a method to efficiently search the chemical space. The proposed search method efficiently identifies the material with the desirable set of properties an order of magnitude faster than genetic algorithms and bayesian optimizations. Applying this to the discovery of oxygen evolution reaction (OER) catalysts, I predict several previously unreported promising candidates that are then validated in the lab. The best candidate Ru0.58Cr0.25Mn0.08Sb0.07O2 shows a mass activity 8 times higher than the current state-of-art, RuO2, and maintains performance for 180 hours while operating at 10mA/cm2 in acidic 0.5M H2SO4 electrolyte.The studies in this thesis introduce and demonstrate the application of data-driven approaches to accelerate the discovery of energy materials through the development of generalizable ML models, interpretability analysis and efficient search algorithms