A New Computer Model Calibration Framework Using Deep Learning

• Activity Number: 266 - Recent Advances in Spatial-Temporal Modeling and Its Applications
• Type: Topic Contributed
• Date/Time: Tuesday, August 4, 2020 : 1:00 PM to 2:50 PM
• Sponsor: Korean International Statistical Society
• Author(s): Saumya Bhatnagar* and Won Chang and Seonjin Kim and Jiali Wang
• Companies: Univ of Cincinnati - Cincinnati, OH and University of Cincinnati and Miami University and Argonne National Laboratory
• Keywords: Computer model Calibration; CNN; LSTM; Spatial Data; Temporal Data

Abstract

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. Here we have formulated a new calibration framework based on deep learning methods which are capable of handling complex data-model discrepancy structures, one of the long-standing issues. By directly modeling the inverse relationship between the model output and their corresponding input parameters using deep neural network frameworks, our approach can efficiently handle highly complex large-scale spatial and temporal data. For uncertainty quantification, the Monte Carlo dropout method based on Variational Bayes has been used. We have tested our framework by calibrating UVic ESCM and WRF-hydro models and confirmed that it can successfully estimate the input parameter settings even with substantial data-model discrepancy. We have used already established deep learning methods such as LSTM or CNN in these problems.