(290d) Bayesian Parameter Estimation for Connecting Theory & Predictions to Experiments

A frequent goal of research is to find the correct physical parameters for chemicals and materials (energetics, lattice constant, etc.). Bayesian parameter estimation (BPE) is intended to give the most probable set of parameters for a given model and a given data set. However, this same methodology can be applied to multiple types of problems in different ways. Broadly speaking, application of BPE can be divided into two categories. One way, the more 'modern' usage of BPE is obtaining parameters which are adequate for creating predictive surrogates (this is often what BPE is used for in machine learning, and is the philosophy behind much of the application of Gaussian Processes to sparse data sets). This first category can be considered analogous to making a calibration curve with fitting. A second way, and more 'traditional' application of BPE is for finding the parameters which have the highest probability of being correct, for a given model. Additionally, this second category of BPE can also be used for model selection (which model is more probable given the data). This second category of BPE is particularly useful for comparing theoretical predictions to experiment. Importantly, practitioners using the first category of BPE to make theoretical predictions can then use the second category of BPE to compare with experiment. In this talk, the usefulness of this concept will be presented, as well as how it is implemented in the free, open source, and user friendly CheKiPEUQ software. This software was developed with U.S. Department of Energy funding, and has been designed for science and engineering applications.