Graph neural networks and high throughput quantum chemistry workflows for detailed kinetic modeling

Abstract

Detailed kinetic modeling describes a complex chemical system in terms of thousands of intermediate elementary reaction steps occurring between hundreds of short lived species. Many chemical systems in combustion, atmospheric chemistry, and heterogeneous catalysis are described using detailed kinetic models. Building detailed kinetic models manually is error prone and requires a significant amount of human time. Computer aided automatic reaction mechanism generation accelerates the process of building detailed kinetic models with little to no human input, enabling faster research iteration cycles and experimentation. Due to the vastness of chemical space and short time scales, only a handful of thermochemical or kinetic parameters can be determined experimentally. Quantum chemistry methods, such as Density Functional Theory (DFT), allow for the determination of thermochemical properties of chemical species in-silico with sufficient accuracy, but come at expensive computational cost which hinders the rapid development of detailed kinetic models. Fast estimation methods, such as Benson's group additivity (GA), are at the forefront of detailed kinetic modeling of hydrocarbons containing C,H,O, and N atoms. However, GA estimation methods are difficult to scale or to extend to new chemical systems when an extra set of atoms, such as F, Cl, and Br, are added because of the combinatorial nature of complex chemical systems.This thesis presents data-driven machine learning methods for detailed kinetic modeling and high throughput workflows to generate high quality thermokinetic data. Over the past decade, data-driven machine learning (ML) methods, specifically neural network based deep learning (DL) methods, have outperformed traditional rule-based methods in computer vision, natural language processing, and machine translation. Graph neural networks (GNN) are an obvious choice for chemistry because chemical species can be modeled as "nodes" of individual atoms connected by "edges" representing the bonds or interactions between neighboring atoms. The nodes and edges are initialized with chemically relevant features and the information is shared between nodes by iteratively aggregating and updating features on nodes using a procedure called message-passing. This thesis evaluates several message-passing GNN architectures in the context of developing thermochemistry estimators for automatically building detailed kinetic models. Also, this thesis introduces Attn MPNN, a message passing neural network with multi head attention mechanism, which outperforms previously published GNNs in predicting thermochemistry by ~25% and GA estimations by ~70%. Furthermore, a high-throughput automated quantum chemistry workflow is developed to calculate high quality thermochemistry to further enhance the applicability of GNN estimators to novel chemical systems. By utilizing the automated thermochemistry workflow, this thesis presents a large-scale high quality dataset of approximately 67k unique chemical species, mostly comprised of halocarbons, calculated using high quality DFT methods. Kinetic parameters of elementary reactions in a detailed kinetic model are typically calculated using Transition State Theory (TST). However, identification and optimization of a transition state (TS) is a major computational bottleneck in the determination of kinetics. To that end, this thesis introduces an ML accelerated method to identify and optimize a TS given an initial starting geometry. Finally, we introduce OrbNet Denali, a large scale GNN model, as a drop-in replacement for ground state DFT energy calculations. OrbNet Denali uses features from low-cost, semi empirical quantum chemistry methods to predict the electronic energy of a given geometry. OrbNet Denali is benchmarked using several well established datasets and shown to provide accuracy on par with modern DFT methods but with up to three orders of magnitude in speedups.--Author's abstract