Mitigating climate change necessitates a rapid transition away from fossil fuels and toward renewable carbon-neutral energy sources such as wind and solar. This doctoral research addresses fundamental limitations of first-principles computational methods for the design and discovery of new processes and materials to accelerate industrial decarbonization and the global transition to a clean and sustainable energy economy by developing practical methods that build on thermodynamics and leverage foundational advances in machine learning.
Recent breakthroughs in artificial intelligence for materials design and discovery aim to screen entire material libraries for desirable properties and to predict novel materials with target properties. Because of the scarcity of available thermodynamic data, designing materials for thermodynamic conditions far away from absolute zero temperature and pressure has proven particularly challenging. In principle, machine learning can speed up materials modeling by providing surrogate models, learning the relationship between structure/composition features and material properties, and training the model to predict desired properties. Due to a lack of experimental data, these models rely heavily on synthetic data from first-principles approaches such as electronic density functional theory. Designing high-temperature processes is also problematic because of the intrinsic limitations of conventional density functional theory calculations, which are strictly correct only at zero temperature. To overcome these data and methodological limitations, I integrated thermodynamic relationships with machine learning models to augment results from first-principles calculations. Additionally, I identified materials descriptor spaces that provide natural representations of structures and compositions for materials discovery.
Chapter 1 introduces in more detail the motivation for this doctoral research and for the combination of computational materials thermodynamics and machine learning. Chapter 2 reviews computational materials science methods that I employed. Chapter 3 showcases how the melting temperatures of materials can be predicted with a combination of electronic structure theory and machine learning. In Chapter 4, our approach for Gibbs free energy predictions is discussed. Chapter 5 deals with the representation learning of materials, dimensionality reduction, quantifying the information content of materials representation spaces, and constructing property-aware materials descriptors. I conclude the thesis with a summary and a discussion of future directions.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/1qnv-6142 |
| Date | January 2024 |
| Creators | Gharakhanyan, Vahe |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0017 seconds