APPLICATION OF POST-STACK SEISMIC ATTRIBUTES, WELL DYNAMIC DATA, AND MACHINE LEARNING FOR CARBONATE RESERVOIR FACIES AND PRODUCTIVITY PREDICTION FOR UPSTREAM OPTIMIZATION

Abstract

Even almost 20 years after the discovery of the prolific pre-salt oil province in SE Brazil offshore, seismic characterization of these complex reservoirs is still a challenging task. In the oil and gas fields, where dozens of wells are available, the seismic inversion-based workflows are proven to be the best for reservoir characterization. However, in exploratory stage areas the reduced number of wells makes seismic inversion non-viable. As a result, solely seismic amplitude-based workflows are strongly affected by ambiguity pitfalls, which delay the appraisal phase, postpone commercial production, and, thus, erode the economic value of the project. As a solution, I propose a machine learning approach that leverages the big amount of dynamic reservoir data in a known field (Mero field in Santos Basin) and easy-to-obtain post-sack 3D seismic attributes, which are common to both production and exploration areas and are available as soon as the seismic processing is done, without requiring wells for calibration like seismic inversion does. Well dynamic data, mainly drill stem tests, are the most reliable information on reservoir productivity before the commercial production and are available during the upstream phase. In this sense, my doctoral research was structured in three main projects: the first one (Chapter 2) on qualitative reservoir modelling using post-stack seismic attributes and unsupervised learning techniques. I used combinations of ten different seismic attributes (interval velocity, amplitude versus offset, complex traces, geometric and voxel-based texture) as inputs for self-organizing maps, generative topographic mapping and k-means clustering algorithms for seismic facies discrimination in the Barra Velha reservoir of the Mero field. The seismic facies models were validated using information from 13 wells which were not used for any training. The best models were obtained with self-organizing maps. The second project (Chapter 3) is on quantitative interpretation using the same attributes and drill stem test data to train supervised learning algorithms for prediction of reservoir productivity of the Barra Velha carbonates in the Mero field. I tested classic supervised learning algorithms (shallow learning) of random forest, support vector machines and K-nearest neighbors for supervised regression of flow capacity and productivity index. I also tested a deep learning algorithm (multi-layer perceptron) to compare its cost-effectiveness to the shallow learning algorithms. For the validation of my predictive models, I used information from 20 blind test wells. The best results were obtained with random forest regression of flow capacity (85% blind test performance). In the third project, I studied how transfer learning techniques can leverage the machine learning training using Mero field data to accurately predict reservoir facies and productivity in other seismic surveys 200 km distant from Mero field: the Bacalhau and Lapa fields, which have production data to validate my predictive models. Using self-organizing maps and random forest algorithms trained with Mero field data, I could accurately predict (80% average performance) the facies distribution and flow capacity values observed in the blind test wells in Bacalhau and Lapa fields. Transfer learning using Mero training proved effective for reservoir de-risking and upstream optimization even when working with multiple seismic surveys. As I used post-stack seismic attributes, well test productivity, and injectivity data from subsurface reservoirs to train our models, this approach can be used in any kind of project such as CCUS, geothermal, and hydrogen storage projects in the context of the energy transition.