Dr. Tianfang Xu, Postdoctoral Research Associate at Michigan State U., presents "Bridging models and data: error-explicit Bayesian uncertainty quantification of groundwater models".

"Physically-based subsurface and integrated hydrologic models are powerful quantitative tools to support water resources decision making under varying hydrologic, climatic and anthropogenic development conditions. These models are inherently subject to errors in input data (inputs such as groundwater withdrawal rate are often unknown or estimated) and model structure (due to simplification and/or misrepresentation of the “true” hydrologic systems). First, the degrading effects of inputs and model structural errors on model predictions are discussed. Then, we present an error-explicit Bayesian uncertainty quantification framework that accounts for various sources of uncertainties. The framework constructs error models to statistically describe errors in model structure and input data, and the error models are inferred in an inductive, data-driven way through learning from observations of groundwater system response. The error-explicit framework is illustrated using synthetic and real-world groundwater modeling case studies. In all case studies, it is found that the error-explicit framework achieves substantially more accurate prediction with more robust prediction interval than conventional uncertainty quantification techniques, which neglect input and model structural errors and therefore yield biased and overconfident prediction. We also show that Bayesian inference can be greatly facilitated using high performance computing and fast surrogates based on machine learning. Finally, we discuss future directions for our work on model-data fusion and its applications in hydrologic modeling."