Search

We present a generic method for automatically calibrating a computer code to an experiment, with uncertainty, for a given “training” set of computer code runs. The calibration technique is general and probabilistic, meaning the calibration uncertainty is represented in the form of a probability distribution. We demonstrate the calibration method by calibrating a combined Finite-Discrete Element Method (FDEM) to a Split Hopkinson Pressure Bar (SHPB) experiment with a granite sample. The probabilistic calibration method combines runs of a FDEM computer simulation for a range of “training” settings and experimental uncertainty to develop a statistical emulator. The process allows for calibration of input parameters and produces output quantities with uncertainty estimates for settings where simulation results are desired. Input calibration and FDEM fitted results are presented. We find that the maximum shear strength σ t max and to a lesser extent maximum tensile strength σ n max govern the behavior of the stress-time curve before and around the peak, while the specific energy in Mode II (shear) E t largely governs the post-peak behavior of the stress-time curve. Good agreement is found between the calibrated FDEM and the SHPB experiment. Interestingly, we find the SHPB experiment to be rather uninformative for calibrating the softening-curve shape parameters ( a , b , and c ). This work stands as a successful demonstration of how a general probabilistic calibration framework can automatically calibrate FDEM parameters to an experiment. [ABSTRACT FROM AUTHOR]

Data collected from a network of detectors and analyzed together has the opportunity to provide a more complete picture than when the data from each individual detector are analyzed independently. However, even with a dense array of detectors, the network data will not provide a complete picture and constraints will need to be added to the model in order to maximize the usefulness of the conclusions that can be drawn. In this work, we demonstrate this concept by considering the task of tracking a moving radioactive source of special nuclear material in a structured environment with data from a network of radiation detectors. Our approach uses a Bayesian model and analysis that naturally provides uncertainty in the estimate of the source's dynamic location. We find that adding domain aware constraints to a Bayesian model (e.g., the location of the road) can improve both location inference and do so with diminished uncertainty even though the fit to gamma count data is largely unchanged. [ABSTRACT FROM AUTHOR]