Multiple-point statistical simulation for hydrogeological models: 3D training image development and conditioning strategies

Most studies about the application of geostatistical simulations based on multiple-point statistics (MPS) to hydrogeological modelling focus on relatively fine-scale models and concentrate on the estimation of facies-level, structural uncertainty. Much less attention is paid to the use of input data and optimal construction of training images. For instance, even though the training image should capture a set of spatial geological characteristics to guide the simulations, the majority of the research still relies on 2D or quasi-3D training images. In the present study, we demonstrate a novel strategy for 3D MPS modelling characterized by: (i) realistic 3D training images, and (ii) an effective workflow for incorporating a diverse group of geological and geophysical data sets. The study covers an area of 2810 km2 in the southern part of Denmark. MPS simulations are performed on a subset of the geological succession (the lower to middle Miocene sediments) which is characterized by relatively uniform structures and dominated by sand and clay. The simulated domain is large and each of the geostatistical realizations contains approximately 45 million voxels with size 100 m × 100 m × 5 m. Data used for the modelling include water well logs, high-resolution seismic data, and a previously published 3D geological model. We apply a series of different strategies for the simulations based on data quality, and develop a novel method to effectively create observed sand/clay spatial trends. The training image is constructed as a small 3D voxel model covering an area of 90 km2. We use an iterative training image development strategy and find that even slight modifications in the training image create significant changes in simulations. Thus, the study underlines that it is important to consider both the geological environment, and the type and quality of input information in order to achieve optimal results from MPS modelling. In this study we present a possible workflow to build the training image and effectively handle different types of input information to perform large-scale geostatistical modelling.