HYBRID SEARCH OF GLOBAL SOLUTIONS FOR HIGHLY NON-LINEAR PROBLEMS USING TREED GAUSSIAN PROCESS AND ADAPTIVE SAMPLING

Abstract

Numerical models of geothermal reservoirs are characterised by a strongly non-linear relationship between the input parameters and the model outputs, which makes the calibration step a difficult task. Some Bayesian global approaches rely on the construction of an emulator to provide reliable and fast approximations of the simulator output and hence time consuming tasks, such as sensitivity analysis or optimisation runs can be run in these fast approximations. This work presents some preliminary results on the linkage of a Gaussian process emulator with local and global optimisation routines in a synthetic geothermal reservoir with the aim of finding optimal input parameters that match some surface measurements.