Online Archive of University of Virginia Scholarship
Efficient Multistate Reweighting and Configurational Mapping Algorithms for Very Large Scale Thermodynamic Property Prediction from Molecular Simulations542 views
Author
Paliwal, Himanshu, Chemical Engineering - School of Engineering and Applied Science, University of Virginia
Advisors
Shirts, Michael, Department of Chemical Engineering, University of Virginia
Abstract
Thermodynamic property prediction using molecular simulation is a computationally expensive process and is the major bottleneck in large scale calculations where thermodynamic properties have to be evaluated multiple times. Examples of processes requiring multiple rounds of molecular simulation include molecular design, scanning of atomistic force field parameter space, simulation parameter space, and process design.
We present a reweighting technique which can accelerate property estimation by three orders of magnitude. We show the proof of this concept by doing a search in a combinatorially large simulation parameter space. The calculations in this exercise, if done without reweighting, would have taken 60 CPU years. However, with reweighting we were able to complete the search within one CPU month achieving an acceleration of 800 times. This search process makes error quantification of the simulation parameter space possible. The search also enabled us to choose a set of computationally inexpensive simulation parameters which gives statistically indistinguishable results compared to the most accurate but computationally most expensive set of parameters.
Estimating free energies between states having no overlap in the configurational space is very hard or sometimes impossible using standard free energy techniques. We have developed a multistate reweighting with configuration mapping algorithm which makes previously impossible problems trivially easy to solve. Free energy differences between rigid water models and free energy corresponding to change in the equilibrium
bond length of a dipole is estimated using the newly developed algorithm. The calculations are not only easier to perform but are also three to five orders of magnitude faster compared to standard techniques.
Finally, we use a reweighting and configuration mapping algorithm to accelerate a multidimensional, multiobjective parameterization of rigid water model. The parameterization also involved the design of an objective function which is able to simultaneously reduce error in all thermodynamic properties estimated using molecular simulations. The parameterization with standard techniques would have taken 1544 CPU years but with the application of the newly developed techniques the computational time was reduced to eight CPU weeks. The forcefield parameterization techniques based on pure fluid properties is extended to forcefield parameterization
based on mixture properties and proposed as the next step in this research program.
Paliwal, Himanshu. Efficient Multistate Reweighting and Configurational Mapping Algorithms for Very Large Scale Thermodynamic Property Prediction from Molecular Simulations. University of Virginia, Chemical Engineering - School of Engineering and Applied Science, PHD (Doctor of Philosophy), 2014-04-08, https://doi.org/10.18130/V3NZ30.
