Abstract:
We introduce a novel parallel rejection scheme to give a simple but reliable way to parallelize the Metropolis-Hastings algorithm. The algorithm is demonstrated by an application of sampling the posterior distribution over eight parameters in a nonlinear numerical model of a geothermal field to achieve model ‘calibration’ from measured well-test data. We explore three scenarios using different training data subsets. Comparison across scenarios indicates model error. Comparison of one scenario with a previous least-squares estimate for the same model and data set shows that sample-based statistics give a more robust estimate than gradientbased least-squares, in less compute time.
