 |
Center for Nonlinear Studies |
A primary goal in modeling complex systems is to modify already good parameters of a model so that the emergent properties of the system match known data. For example, we may have an approximate model of the behavior of a protein but it may have the wrong average radius or we may have a good model of water that doesn’t have exactly the right tetrahedral geometry. Repairing an approximately correct model is traditionally done through iterative cycles of simulation and re-parameterization. An alternative approach is to add terms to bias the underlying probability distribution of a system to force properties of the model to match known data directly. This process can be formulated to minimally change the model with a relative entropy argument, arriving at the smallest possible change to a model to improve it. I'll present two methods, one for when expected values are to be match and one when complete probability distributions are to be matched. I will use examples from chemical simulations. The first method, called experiment directed simulation, converges rapidly, can match many properties simultaneously, and minimally modifies the probability distribution of a model (maintains maximum entropy). The method has been tested on model systems and a three-component electrolyte simulation. Experiment directed simulations can only match expected values in molecular simulations. I'll also present a second method call targeted metadynamics, which can be used when a complete probability distribution of arbitrary properties of a model are known and we wish to modify the model parameters to match them. These two new methods open new ways to directly combine complex models and experiments.
Host: Gregory Voth