Abstract: Various scientific and engineering areas rely on computer model experiments to study and make future projections of important physical processes. The success of a computer model experiment is heavily determined by the way the relevant uncertainties are properly quantified and constrained. One important source of uncertainties surrounding a computer model experiment is the parametric uncertainty, which is due to incomplete knowledge of the “correct” values for model input parameters that control important physical characteristics of the simulated processes.
Computer model calibration is a statistical framework for combining information from computer model runs and the corresponding real-world observations to quantify and reduce parametric uncertainties. The existing calibration framework is subject to various issues including nonidentifiablity between input parameter effects and data-model discrepancy effects and computational challenges due to large spatial or temporal model output.
Our aim is to develop a new class of statistical calibration framework based on deep neural network models that do not suffer from the major shortcomings of the existing approach. The central idea is to model the inverse relationship between the model output and input parameters directly by using deep learning network models. By utilizing the feature extraction ability of deep networks, the proposal approach can filter out the signal from data-model discrepancy and accurately estimate the true input parameter values. We will apply the developed method to the problem of calibrating WRF-hydro model to demonstrate the scientific utility of our approach.