Efficient Sensitivity Computations For Automatic Geothermal Model Calibration

Abstract

The TOUGH2 family of codes is used extensively around the world to model geothermal reservoirs. The efficacy of the TOUGH2 simulator, as a tool for describing system behaviour, relies heavily on the sound choice of the input variables. Increasingly, automated model calibration is used to determine suitable model parameters, with gradient-based inverse methods commonly used as part of the automatic calibration process. These methods currently require multiple time-consuming forward runs of the geothermal simulator (TOUGH2) in order to calculate derivatives of the model response with respect to each model parameter. Progressive refinement and revision of the current inversion methods are required to overcome this computational cost and to accommodate the geothermal industry\'s need for more sophisticated models with an ever-greater number of model parameters. This paper describes analytical approaches for improving the computational efficiency of the inversion process. We describe the motivation and basic theory of the adjoint and direct simulation methods for calculating derivatives of model outputs with respect to model parameters. Finally, the implementation of the methods is discussed for large-scale geothermal simulators such as TOUGH2.