A Data-Driven Approach for Stylolite Detection

Abstract

Summary Stylolite is a specific geopattern that can occur in both sedimentary rocks and deformed zones, which could change the porosity of the reservoir, modify the permeability, and even result in horizontal permeability barriers. Though there are many related studies to characterize this geopattern, most of them focus on experimental methods. In this work, we investigated a new approach for recovering geometrical information of the stylolite zone (including its size and location) based on neural network architectures including convolutional neural network (CNN), long short-term memory (LSTM), and transformer encoder, which could serve as a data-driven solution to the problem. To our knowledge, this paper is the first to exclusively use well testing data for deducing field geopatterns, whereas other studies have relied on additional data sources such as well logging data. To simplify the problem, we first conducted simulation by building 3D multilayer reservoir models with one stylolite zone. We considered both simplified cases with only a few homogeneous layers and cases with heterogeneous layers to generalize our work. For the heterogeneous case, we extracted the permeability from SPE10 Model 2, a commonly used public resource (SPE 10 Benchmark, Model 2 2008). Producing and observing wells in the model are at different locations and provide pressure and production rate data as inputs for the deep learning models, in the form of multivariant time series data. For homogeneous cases, after zero-padding and standardizing our inputs to tackle different-length data and features with different scales, we applied a CNN-LSTM model to our data set, leveraging the CNN’s ability to capture short-term local features and the LSTM’s capacity to extract long-term dependencies. This combination improves the extraction of important information from time series data related to stylolites. The two subnetworks are connected in parallel to combine the advantages of CNN in extracting local temporal features and the strengths of LSTM in capturing long-time dependency via self-loops. Our work also separately covers the two subnetworks of the CNN-LSTM model as baseline models. For heterogeneous cases, a CNN-based model U-Net and an attention-based model set function for time series (SeFT) were introduced to make the predictions. In a more realistic scenario featuring fluid pathways with irregular shapes within a heterogeneous reservoir, we employed a transformer encoder model to predict both the shape and location of the fluid pathways. On the homogeneous data set, our CNN-LSTM model achieved a satisfactory performance and could predict the locations and sizes of the stylolite zone and outperformed the two baseline models. On the more challenging heterogeneous data set, our baseline and CNN-LSTM models failed to deliver meaningful results. In contrast, SeFT and U-Net performed better in the sense that we could successfully predict the locations of the stylolite zones. In a more realistic scenario involving irregularly-shaped fluid pathways, our transformer encoder model achieved accurate predictions of their locations.