Machine Learning a Linear Gibbs Free Energy Function Towards Accelerated Green Catalysis Computation
Date of Award
Spring 2023
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Chemistry
First Advisor
Batista, Victor
Abstract
At a fundamental level, the field of chemistry studies the identities and properties of molecular and atomic substances and how they react. The energetics that guide those reactions determine which mechanisms are performed and which products are formed in what proportion. Intriguingly, reactions may take place between not just reactants that are transformed into products, but also with the help of molecules known as catalysts that may interact with the reactants to lower the activation energies of the reaction steps and then return to their original state by the end of the reaction. Catalysts therefore play an important role in making many reactions faster than they would be without the presence and reaction of the catalyst. Within the area of energy production and use, catalysts are especially beneficial due to their abilities to both lower energy costs and increase the efficiency of energy producing reactions. However, knowing exactly how a catalyst will react before experimentation requires calculation of the energetics. Currently, to produce reliable energies requires utilization of quantum ab initio techniques. While these methods work and are well understood, they take times often on the scales of hours to weeks for single calculations to complete, especially for catalysts which are often larger molecules. Solutions to decrease the time requirements have been introduced, from assumptions to force fields. However, these alterations often exchange time for accuracy, making them useful for speeding up initial optimizations, but requiring additional higher level calculations to validate their results. Despite this, Hohenberg and Kohn's first theorem suggests that there exists a universal functional that solves the Schrodinger equation exactly for energy. The universality of this functional means that it is molecular size agnostic, and therefore would be able to find the energy of dramatically variant molecules. Inspired by this theorem, one may be reminded also of neural networks within the field of machine learning, which are commonly dubbed universal function machines because of their ability to, in the limit of an infinitely sized model, represent any function. Of course, once a function is determined, the benefit is that the size of the function is set and output values can be determined in constant time. Pairing these two theories together then presents an opportunity to improve the way in which energetics of individual molecular systems are calculated. Herein, we introduce catalytic systems for ammonia conversion and explore their mechanisms the using traditional ab initio calculations as a way to highlight the importance of such studies and also highlight the needs of any machine learning system that wishes to replace those systems. This will be followed by three machine learning studies that iteratively build upon the findings of each other to determine new ways to calculate energetics of molecular and catalytic systems. This work culminates in the development of a simple input representation for molecules that requires no prior ab initio computations to generate, and a molecular size agnostic functional form that calculates Gibb's free energies in constant time and has a meaningful form from which chemical understanding can be easily extracted. Finally, points that are not yet handled under the current methods that Gibbs free energy and catalyst problems need for a robust machine learning system that outperforms current ab initio methods are laid out. Sprinkled next to relevant research chapters are reviews and tutorials related to each topic.
Recommended Citation
Freeze, Jessica Gene-Caitlyn, "Machine Learning a Linear Gibbs Free Energy Function Towards Accelerated Green Catalysis Computation" (2023). Yale Graduate School of Arts and Sciences Dissertations. 925.
https://elischolar.library.yale.edu/gsas_dissertations/925
