Predictive and Thermodynamically Transferable Coarse-Grained Models
Restricted (Penn State Only)
- Author:
- Szukalo, Ryan
- Graduate Program:
- Chemistry
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 15, 2023
- Committee Members:
- William Noid, Chair & Dissertation Advisor
Lasse Jensen, Major Field Member
Susan Sinnott, Outside Unit & Field Member
John Asbury, Major Field Member
Philip Bevilacqua, Program Head/Chair - Keywords:
- coarse-graining
transferability
representability
local density - Abstract:
- Computational methods have emerged as powerful tools for studying molecular condensed phase systems. Specifically, all-atom molecular dynamics has profoundly impacted the understanding and investigation of various molecular phenomena. However, the use of high-resolution simulations can be limited when modeling larger scales, longer times, and emergent collective behavior. Consequently, multiscale approaches have become crucial in chemical research, as exemplified by the 2013 Chemistry Nobel Prize. One such approach involves the development of bottom-up coarse-grained models, which are parameterized to replicate the structural ensemble generated by high-resolution models at a lower coarse-grained resolution. Bottom-up coarse-grained models simplify the representation of a system by grouping multiple atoms together into coarse-grained ``sites'' or ``beads'' that collectively retain the essential properties and interactions of the original system. This reduction in complexity decreases computational demand and enables efficient simulations at larger length- and time-scales. While there exists various coarse-graining methodologies, a rigorous bottom-up approach offers several advantages. By establishing a statistical mechanical link between the high- and low-resolution models, there exists a basis for parameterization and systematic improvement of the resulting coarse-grained models. Moreover, the resulting coarse-grained models are chemically specific, meaning they are directly related to a particular chemical system, as they are parameterized from a high-resolution model of the same (similar) system. Thus, these approaches not only offer increased efficiency relative to all-atom models but also can reveal chemically specific microscopic insights. The exact energy function for reproducing the high-resolution model at a given coarse resolution is the many-body potential of mean force. Unlike traditional atomistic potentials, the many-body potential of mean force depends on the thermodynamic state point. This dependence gives rise to two main issues that impede the widespread application of bottom-up coarse-grained models. First, coarse-grained models parameterized at one state point may not accurately represent other state points, leading to what is known as the transferability problem. Second, improper treatment of the thermodynamic dependence of the potential of mean force often results in poor modeling of thermodynamic properties, known as the representability problem. The transferability and representability problems stem from the fact that the removed atomistic degrees of freedom depend on the thermodynamic condition. Addressing these issues necessitates the development of new theories and methodologies. This thesis aims to further understand these issues for more complex approximate models, such as coarse-grained models with intramolecular degrees of freedom, across glass transitions, and for local density potentials. Initially, we demonstrate that the linear thermodynamic dependence observed for effective potentials of simple molecular liquids becomes more complex when considering a glass transition. Although the structure of a liquid and a glass are similar, the resulting coarse-grained potentials vary non-monotonically with temperature and density. Despite this complexity, we show that the potentials are more sensitive to density than temperature across a glass transition, as previously observed for other molecular liquids. Furthermore, the resulting coarse-grained models accurately reproduce the pair structure for both 1- and 3-site coarse-grained models of ortho-terphenyl. After quantifying how coarse-grained interaction potentials vary with temperature and density, we applied a ``dual approach'' that estimates the temperature variation (at constant density) of coarse-grained potentials and accurately evaluates energies at a given coarse-grained resolution. This approach enables the generation of predictive coarse-grained models, allowing for an analytical modification of a single coarse-grained potential to study state points for which it was not originally parameterized. Accurately predicting effective CG potentials across temperatures from a single state point provides an efficient alternative to reparametization of the CG potential, which require high resolution simulations at each temperature of interest. Specifically, we also extend the dual approach to incorporate both inter- and intramolecular degrees of freedom, expanding the range of systems that can be studied. Despite the success of the dual approach, it has only been applied for CG potentials that model all intermolecular interactions as pair-additive potentials that are a function of the distance between two particles. Pair-only models are often unable to capture complex many-body correlations that are introduced during the coarse-graining process, thus they are usually restricted to bulk, homogeneous systems. Therefore, we also investigate how coarse-grained potentials that depend on a molecule's local density vary with model parameters to begin developing transferable models for more complex systems. We demonstrate that the accuracy of these local density potentials depends on the length-scale used to define the local density. By employing an optimal length-scale, the resulting potentials exhibit reasonable transferability across a wide range of temperatures. Moreover, adopting a temperature-dependent length-scale leads to temperature-independent local density potentials, while the pair potential varies linearly across the temperature range. This provides a straightforward means of predicting potentials that accurately model new state points, in the NPT ensemble (varying density) without additional atomistic simulations. This research highlights the utility of extrapolative methods for transferring effective potentials across thermodynamic state points to fully realize the efficiency gains promised by bottom-up coarse-grained models. Developing such methods is crucial to reduce reliance on computationally expensive all-atom simulations. The dual potential approach and the temperature-dependent length-scale scaling demonstrate predictive accuracy beyond the originally parameterized state points, contributing novel and predictive methods to the field. Moreover, the implications of this research extend beyond the specific methods developed. By addressing the limitations of bottom-up coarse-grained models, this work contributes to the anticipated exploration of larger length- and time- scales, facilitating investigations into complex phenomena, such as glass transitions, intramolecular degrees of freedom in coarse-grained models, and the influence of local density potentials on thermodynamic properties.