Uncertainty Quantification for Basin Scale Heat Flow Models with a Physics-Based Machine Learning Approach.

In order to determine suitable locations for geothermal exploration, reliable predictions of theearth’s subsurface temperature field are essential. For these predictions, it is necessary toconsider the uncertainties of the involved parameters. However, with the currentstate-of-the-art simulations standard uncertainty quantification methods, such as MarkovChain Monte Carlo are computationally intractable for basin-scale models at high resolution.We thus require numerical methods that considerably accelerate the forward simulation toenable the use of uncertainty quantification approaches that can easily require up to a millionforward simulations. For this purpose, we introduce the reduced basis method, a physics-based machinelearning approach. Our previous studies show that we obtain speed-ups of four to six ordersof magnitude in comparison to standard finite element simulations. One main advantage of the reduced basis method in contrast to other surrogate modelsis that we obtain temperature values at every point in the model and not only atthe observation points. Consequently, we can generate uncertainty maps of thetemperatures at the target depth of the geothermal wells for the entire extent of thebasin. We use the Brandenburg (Germany) model to illustrate the application and benefits of thereduced basis method for large-scale geological models. The numerical simulations arerealized within the DwarfElephant package, an open-source high-performance applicationbased on the Multiphysics Object Oriented Simulation Environment (MOOSE)developed by the Idaho National Laboratory. The DwarfElephant package offers aphysics-independent and user-friendly access to the reduced basis method within ahigh-performance finite element library, allowing computations of spatially highdimensional models. In addition, we present how the method can be used for other inverseprocesses, such as automated model calibrations. Inverse problems are becoming rapidlyextremely expensive computationally even without including all major sources ofuncertainty. In that regard, the reduced basis method is very promising because it allows asignificant reduction in computation time without introducing additional physicaluncertainties. [ABSTRACT FROM AUTHOR]