Efficient Calibration of Computationally Intensive Groundwater Models through Surrogate Modelling with Lower Levels of Fidelity
Abstract
Many sophisticated groundwater models tend to be computationally intensive as they rigorously represent detailed scientific knowledge about the groundwater systems. Calibration (model inversion), which is a vital step of groundwater model development, can require hundreds or thousands of model evaluations (runs) for different sets of parameters and as such demand prohibitively large computational time and resources. One common strategy to circumvent this computational burden is surrogate modelling which is concerned with developing and utilizing fast-to-run surrogates of the original computationally intensive models (also called fine models). Surrogates can be either based on statistical and data-driven models such as kriging and neural networks or simplified physically-based models with lower fidelity to the original system (also called coarse models). Fidelity in this context refers to the degree of the realism of a simulation model. This research initially investigates different strategies for developing lower-fidelity surrogates of a fine groundwater model and their combinations. These strategies include coarsening the fine model, relaxing the numerical convergence criteria, and simplifying the model geological conceptualisation. Trade-offs between model efficiency and fidelity (accuracy) are of special interest. A methodological framework is developed for coordinating the original fine model with its lower-fidelity surrogates with the objective of efficiently calibrating the parameters of the original model. This framework is capable of mapping the original model parameters to the corresponding surrogate model parameters and also mapping the surrogate model response for the given parameters to the original model response. This framework is general in that it can be used with different optimization and/or uncertainty analysis techniques available for groundwater model calibration and parameter/predictive uncertainty assessment. A real-world computationally intensive groundwater modelling case study developed with the FEFLOW software is used to evaluate the proposed methodology. Multiple surrogates of this computationally intensive model with different levels of fidelity are created and applied. Dynamically dimensioned search (DDS) optimization algorithm is used as the search engine in the calibration framework enabled with surrogate models. Results show that this framework can substantially reduce the number of original model evaluations required for calibration by intelligently utilizing faster-to-run surrogates in the course of optimization. Results also demonstrate that the compromise between efficiency (reduced run time) and fidelity of a surrogate model is critically important to the success of the framework, as a surrogate with unreasonably low fidelity, despite being fast, might be quite misleading in calibration of the original model of interest.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2012
- Bibcode:
- 2012AGUFM.H43B1330R
- Keywords:
-
- 0550 COMPUTATIONAL GEOPHYSICS / Model verification and validation;
- 1805 HYDROLOGY / Computational hydrology;
- 1846 HYDROLOGY / Model calibration;
- 1847 HYDROLOGY / Modeling