Title: Toward Robust and Efficient Atomistic Modeling of Solvent Effects

 Abstract:

Due to environmental and economic pressures, society has an ever-increasing need for renewable fuels and chemicals. Fulfilling this demand will require discovering and optimizing sustainable chemical processes. Computational screening can be used to explore chemical reactions before investing time and resources in experimental investigations. Atomistic modeling with quantum chemical methods is frequently used for gathering insight into reaction thermodynamics and kinetics. Accurate predictions of reaction pathways must account for crucial environmental effects from solvents and ions. This work investigates how to efficiently and reliably capture these effects without a priori information, such as experimental data. For example, molecular simulations with explicit solvent molecules are the most rigorous approach but bring high computational costs. Implicit solvent models are inexpensive, but their accuracy—especially for charged species or mixed solvents—is often unclear. Here, solvation schemes are discussed for mixed solvents. Implicit solvent models using only a homogeneous dielectric medium are generally inadequate, and frameworks that account for nonuniform solvent distributions are required. Then, sodium borohydride reduction of carbon dioxide is computationally investigated. This work systematically evaluates when different procedures are reliable by replacing solvent molecules from explicitly solvated structures with implicit solvent models. Implicit solvent models alone (i.e., no explicit solvent molecules) are insufficient, and the inner solvation shell is needed to model the reaction mechanism accurately. Thus, explicit solvent models are required for the reliable treatment of solvated reactions. However, the computational cost of numerous energy and force evaluations restricts their usage. Many-body gradient-domain machine learning (mbGDML) is introduced to accelerate molecular simulations based on quantum chemical data. These machine learning force fields are rapidly trained and demonstrate semi-quantitative agreement with water, acetonitrile, and methanol isomer rankings and experimental radial distribution functions.

Chair:

Dr. John A. Keith

Department of Chemical and Petroleum Engineering, University of Pittsburgh

Dr. J. Karl Johnson

Department of Chemical and Petroleum Engineering, University of Pittsburgh

Dr. Giannis Mpourmpakis

Department of Chemical and Petroleum Engineering, University of Pittsburgh

Dr. Kenneth D. Jordan

Department of Chemistry, University of Pittsburgh