Toward a data-driven framework for hydrogeological uncertainty characterization

点击翻译按钮获取中文摘要

The relevance of data to subsurface hydrology, or hydrogeology, is particularly high due to the combination of highly-heterogeneous subsurface properties and the general scarcity of data. This scarcity is caused by the high costs often associated with subsurface exploration. As a result, widely available access to data sets on subsurface conditions should be paramount, since it facilitates the application of data-driven methods like machine learning or Bayesian statistics.Yet collecting these data and making them available to practitioners remains difficult. Currently, the largest database on geostatistical parameters of the subsurface is the World-Wide HYdrogeological Parameters DAtabase (WWHYPDA). It contains approximately 20,000 measurements from 150 sites worldwide. This represents only a small fraction of the total amount of available data, which puts some limit on the characterization of the parametric uncertainty found in the subsurface. Moreover, the WWHYPDA does not contain any information on spatial correlation structures, like correlation lengths or empirical variograms, which means that no information on structural uncertainty can be gleaned from it.In this project, I want to address this problem by increasing the number of data assets available to the community of scientists and practitioners of (stochastic) subsurface hydrology. Mainly two different types of data are going to be used. The first data type is geo-referenced measurements of conductivity and transmissivity fields. They provide the most direct way to estimate spatial correlation structure and are consequently a natural choice. In addition, estimates on statistics of spatial structures exist in the literature and can be used if properly collected and integrated into the database. Finally, pumping tests are going to be used. Pumping tests are a well-established technique for the characterization of subsurface systems. Their application has, however, historically been focussed on the inference of one-point statistics, like the mean value, only. Yet, more recent developments have made it possible to infer two-point statistics, like the correlation length, from pumping tests, as well. After these additional data have been gathered, the final step is to amend existing tools and potentially providing new tools for analyzing, testing and processing these data.If successfully finished, the results from this project would provide a database as well as a number of algorithms, which will facilitate practitioners for the first time to employ data-driven methods to characterize structural uncertainty of the subsurface. If this project succeeds, the process of collecting and making these data available will be streamlined, updated and greatly expanded. In addition, a number of tools will be made available which can help to analyze, process and test these data.