University of Minnesota
Bridging Accuracy and Efficiency in Multireference Calculations: Cost-Effective Electronic Structure Methods and Machine-Learned Potential Energy Surfaces
Abstract
dc:description.abstractAccurately modeling photochemical reactions and strongly correlated molecular systems remain a central challenge in computational quantum chemistry due to the fundamental trade-off between accuracy and computational cost. This thesis addresses these challenges by developing, validating, and integrating cost-effective strategies for describing multireference electronic structure problems and constructing global potential energy surfaces (PESs) suitable for modeling nonadiabatic dynamics. The thesis begins by establishing the complexity of photochemical PESs through a case study of methyl radical photodissociation. Using multireference methods including SA-CASSCF, CMS-PDFT, L-PDFT, and XMS-CASPT2, we characterize the low-lying excited states and dissociation channels, demonstrating both the structural and topological difficulty of accurately characterizing these surfaces as well as the vast computational resources required for its description. The technical complexity involved in describing PESs of photochemical processes indicate the need for efficient analytic representations that aid in modeling nonadiabatic dynamics. To address this challenge, we introduce the Companion Matrix Neural Network (CMNN), a machine-learned approach for fitting coupled adiabatic PESs directly from electronic structure energies at different molecular geometries. CMNN leverages the smoothness of the characteristic polynomial with respect to the molecular geometry and a modified companion matrix to enforce real, smooth eigenvalues, eliminating the need for confusing, labor-intensive diabatization schemes. This method provides accurate analytic fits of the potential energy surfaces for systems ranging from LiH, and H₃ to larger organic molecules like phenol and thiophenol. Building on this foundation, the thesis evaluates the accuracy of multiconfiguration pair-density functional theory (MC-PDFT) as a cost-effective alternative to multireference perturbation theory in describing multireference systems. Benchmarking the recently developed MC23 on-top functional against high-quality excitation-energy datasets reveal that MC-PDFT often matches or surpasses CASPT2 in accuracy while offering far lower computational cost per calculation. This establishes MC-PDFT as a practical electronic structure method for excited-state applications for more realistic systems. To further reduce computational cost, we assess two-point complete basis set extrapolation schemes using reduced-diffuse jun-cc-pVXZ and jul-cc-pVXZ basis sets. These truncated basis sets reproduce aug-cc-pVXZ extrapolations with significantly lower cost, providing an efficient strategy for high-accuracy energy calculations. Finally, the thesis tackles the automerization of cyclobutadiene, a prototypical strongly correlated system where KS-DFT fails. By combining MC-PDFT with the optimized basis set strategy, we obtain highly accurate barrier heights, outperforming unrestricted KS-DFT and approaching (at times surpassing) CASPT2 quality results at a fraction of the computational expense. The work presented in this thesis provides an efficient, accurate, and scalable framework for modeling multireference systems and photochemical processes. The methods developed and validated in this work advance the broader goal of enabling predictive, high-fidelity simulations for complex molecular systems at a practical computational cost.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Parameswaran, Aiswarya M.
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Parameswaran, Aiswarya M.. (2026). Bridging Accuracy and Efficiency in Multireference Calculations: Cost-Effective Electronic Structure Methods and Machine-Learned Potential Energy Surfaces. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/283136.