Title: Assessing and Improving the Accuracy of Chemical Property Predictions Using Electronic Structure Theory and Machine Learning Approaches

 Abstract: Accurate prediction of chemical properties is crucial for advancing catalysis, materials design, and reaction discovery. While density functional theory (DFT) is widely used, it has two major limitations: (1) Self-interaction errors (SIE), where an electron erroneously interacts with itself, leading to inaccurate predictions of properties such as band gaps, reaction barriers, electron affinity, and others, and (2) high computational cost, which restricts atomistic simulations to relatively small systems. My thesis addresses both issues through improved first-principles methods and machine learning integrations.

The first part of my work focuses on overcoming SIE in DFT, particularly in systems with stretched bonds, which often occur in transition states. My work addresses these limitations by applying self-interaction corrections (SIC) using the Perdew-Zunger (PZSIC) and locally scaled (LSIC) methods, implemented within the Fermi-Löwdin orbital SIC (FLOSIC) scheme that uses orbital-by-orbital corrections. We found that both PZSIC and LSIC significantly improved chemical reaction barrier predictions for the BH76 benchmark dataset, with SIE most pronounced in stretched bond orbitals. The BH76 benchmark dataset is comprised of only gas phase reactions. However, since most catalytic reactions occur on transition-metal surfaces, the impact of SIC on transition metal energetics remains unexplored.

To address this, we evaluated the performance of DFT, PZSIC, and LSIC in describing the energetics of 3d transition metals. We introduced a new measure called “sd energy imbalance” to assess the errors in s- and d- electron second ionization energies of the 3d atoms, avoiding the use of excited state energies. While LDA, PBE, and r2SCAN provided a reasonable balance between s and d states, PZSIC introduced significant errors of up to 2 eV.  This was attributed to an energy penalty associated with noded 3d orbitals. This penalty was most pronounced when an electron was removed from half-filled (3d5) or fully filled (3d10) subshells. LSIC mitigated these errors. Understanding and reducing the error associated with describing 3d orbitals is critical for accurately modeling redox reactions and oxidation-state changes in transition metal systems.

Building on these insights, we investigated SIC performance for NOx reduction on a Cu-SSZ-13 zeolite cluster model. We benchmarked the performance of DFT (LDA, PBE, and r2SCAN), PZSIC, and LSIC methods against CCSD(T) reference energies. Among LDA-based functionals, LDA overestimated adsorption energies and underestimated reaction barriers due to SIE. While PZSIC improved the description of reaction barriers in some cases, it introduced large errors in systems involving changes in the Cu oxidation state, with adsorption energy errors exceeding 3 eV. These errors resulted from the SIC energy penalty associated with the 3d10 configuration of Cu. The energy penalty also caused PZSIC to favor 3d9-like FOD configurations in singlet systems, giving rise to errors in the PZSIC densities. Thus, LDA@PZSIC often performed worse than uncorrected LDA because of these density errors. LSIC mitigated these energy penalties and provided a balanced description of both adsorption and reaction energetics. Overall, LSIC outperformed LDA and PZSIC.

To address DFT’s computational limitations, we developed an active learning workflow that combined DeePMD-based machine learning interatomic potentials (MLIPs) with transition-state search algorithms. This workflow enabled efficient, unbiased exploration of reactive potential energy surfaces without requiring prior knowledge of products or intermediates. Our workflow identified new pathways and accurate reaction barriers with near-DFT accuracy for solution-phase methanimine hydrolysis. This work provides methodology to sample reactive intermediates effectively allowing new avenues for automated reaction discovery in catalysis and small-molecule chemistry

Dissertation Chair:

Dr. J. Karl Johnson, Department of Chemical and Petroleum Engineering, University of Pittsburgh

Committee Members:

Dr. John Keith, Department of Chemical and Petroleum Engineering, University of Pittsburgh

Dr. Kenneth Jordan, Department of Chemical and Petroleum Engineering and Department of Chemistry, University of Pittsburgh

Dr. Koblar Alan Jackson, Department of Physics, Central Michigan University