Automatic Estimation of Reservoir Properties Using 3D Machine Learning Workflow Independent of Well Alignment in the Presence of Geological Complexes

Abstract

Abstract One of the most important steps for reservoir characterization is estimating its properties. The machine learning workflows developed in recent years to automate the process mostly capture 2D context of seismic and are capable of using information from vertical wells only. In this work we propose a new machine learning workflow to automatically estimate the properties that not only uses 3D seismic context in the presence of geological complexes but also is applicable to both vertical and deviated wells. Our proposed workflow consists of a machine learning model that estimates the property at each point of the survey by learning the 3D context around it. To build this model, 3D cubes of seismic and salt bodies are extracted around each point of the wells and fed as inputs into a multi-layered convolutional neural network architecture, with the corresponding properties at those well points as outputs. Once the model has been trained, it is used for estimating all properties at any point on the survey simultaneously by feeding it 3D cubes of seismic and salt bodies extracted around that point. To evaluate our method, we use the publicly available dataset from Groningen gas field. We train our model on 35 wells and use 4 other wells as test wells. We used seismic and salt as inputs and sonic and gamma ray logs as output. We observe that the predicted property values by our machine learning workflow for every test well match with the true property values of the wells, thus concluding that our workflow is capable of successfully estimating the properties automatically from seismic and salt information. We also observe that our workflow is capable of predicting properties at the well location that has no properties measured. As our workflow predicts properties at points, it can be used for both vertical and non-vertical wells. We also used the model to predict entire slices which pass through the test wells and observed that the distinct patterns followed by the predicted values coincide with the true values of the wells and the inputs, seismic and salt. The novelty of our approach involves two aspects. The first aspect is using the 3D context of seismic and salt as input. And the second aspect is estimating the properties only at points, thus generalizing the workflow across wells with various alignments – vertical or non-vertical.