2018/12/19 by Thomas Wutzler, Wutzler, Thomas
Computer Science · Engineering · Chemistry · #Gaussian Processes and Bayesian Inference #Fault Detection and Control Systems #Spectroscopy and Chemometric Analyses
paper · pdf · doi:10.48550/arxiv.1812.07801
Mechanistic simulation models are inverted against observations in order to\ngain inference on modeled processes. However, with the increasing ability to\ncollect high resolution observations, these observations represent more\npatterns of detailed processes that are not part of a modeling purpose. This\nmismatch results in model discrepancies, i.e. systematic differences between\nobservations and model predictions. When discrepancies are not accounted for\nproperly, posterior uncertainty is underestimated. Furthermore parameters are\ninferred so that model discrepancies appear with observation data stream with\nfew records instead of data streams corresponding to the weak model parts. This\nimpedes the identification of weak process formulations that need to be\nimproved. Therefore, we developed an efficient formulation to account for model\ndiscrepancy by the statistical model of Gaussian processes (GP). This paper\npresents a new Bayesian sampling scheme for model parameters and discrepancies,\nexplains the effects of its application on inference by a basic example, and\ndemonstrates applicability to a real world model-data integration study.\n The GP approach correctly identified model discrepancy in rich data streams.\nInnovations in sampling allowed successful application to observation data\nstreams of several thousand records. Moreover, the proposed new formulation\ncould be combined with gradient-based optimization. As a consequence, model\ninversion studies should acknowledge model discrepancies, especially when using\nmultiple imbalanced data streams. To this end, studies can use the proposed GP\napproach to improve inference on model parameters and modeled processes.\n