Multiple-point geostatistical simulation using enriched pattern databases

Abstract

This study presents a new approach of generating a set of alternative training images (ATI) to use in patch-based multiple-point simulation. The purpose of using ATI is to improve both the conditioning capabilities of the patch-based methods to hard data and the continuity of the conditionally simulated images. The ATIs are produced as a series of unconditional patch-based simulations using unilateral path with weighting and decoupage to improve continuity. A simple strategy is described to control objectively the ATI generation and keep only the few ATIs most useful to ensure hard data conditioning. Hundreds of ATIs are generated, their statistics are compared with that of the original TI and finally an ensemble of ATIs is selected in a pre-simulation step. The CPU time is kept overall at a quite reasonable level over large 2D and 3D grids by the use of fast distance computation by convolutions and FFT. Different examples are considered: categorical or continuous, with small or large TIs. In 2D, the richest database obtained by adding the ATIs enables to ensure 100 % hard data conditioning in all realizations for the categorical examples tested and a very strong correlation coefficient (r = 0.999) in the continuous case. In 3D, the hard data reproduction rate in the simulation is increased. Different possible improvements to the method are discussed.