Groundwater management in Denmark: downscaling decision objectives and upscaling uncertainty models

Thursday, 27 July 2017: 9:25 AM
Paul Brest West (Munger Conference Center)
Jef Caers, Stanford Earth Sciences, Stanford, CA, United States
The real world requires decision making from messy data of various sources. These decisions are motivated by high-level decision objectives that impact local decision problems. The data however, is mostly local, and hence needs upscaling into a local and regional groundwater model on which decisions can be based. In this talk, I will present a practical case study in Denmark, involving re-allocating water wells that cause pollution and drain wetlands and streamflows. The data available are geophysical data (SkyTEM), head data, streamflow data, well-log data, conceptual geological data and information about the flow processes involved when pumping in a complex biogeochemical system. The strategy followed, termed Evidential Learning, avoids any explicit model inversion, instead, is based on learning from Monte Carlo runs through surrogate models. The surrogate model learns to make decisions directly from data, without calibrating groundwater models explicitly. This allows solving the decision problem effectively and account for all sources of uncertainty (lithology, boundary condition, rock physics, geological interpretation). The decision of re-allocation is currently being implemented by the local water district.