Speaker
Description
As the deadline to reach the Net Zero pledges outlined in the Paris Accords approaches, the need for each country to find economically viable renewable energy sources is a priority. Since 2018, the Netherlands has been involved in expanding its knowledge of its geothermal potential through the SCAN (Seismische Campagne Aardwarmte Nederland) project, which aims to expand the data coverage to areas that, historically, have been less of a focus for oil and gas development. Seismic inversion lies at the heart of a crucial energy challenge: bypassing the need to drill exploratory wells to reduce operational costs, time, and the human footprint on the environment. Currently, extracting reliable rock properties from full-stack seismic data is only possible by acquiring ground truth data from drilled wells. Attempts to bypass this limitation have been undertaken using machine learning algorithms without reliable success. Machine learning can effectively estimate rock properties, reducing the need to rely on conventional seismic inversion, expensive lab experiments, and well logging data.
In this study, we tried to apply the potential of computer vision algorithms to tackle this challenge on a geothermal porous media dataset from the Netherlands SCAN project. The goal of this work is to predict porous media properties using neural networks based first on the post-stack seismic data, and then on the pre-stack seismic data. The following programming libraries such as Equinor’s SegyIO for data loading, Pandas, Numpy, Matplotlib for data preprocessing, Empatches for image patch splitting, and Tensorflow Keras for model building have been used. Our neural network model architecture is based on the medical U-net as described in Ronneberger et al. (2015). The proposed U-net architecture was used to develop models that rely on a full-stack seismic dataset as inputs to the model. In this study, a total of 7 images representing 2D preprocessed (post-stack) seismic data of the SCAN dataset, along with associated ground truth properties derived from conventional seismic inversion, were used for training the models. The 2D input array of the training dataset is transformed into 256×256 patches to increase the size of the dataset to improve the robustness of the developed model. Each of the patches was also flipped horizontally for input into the model. The training/validation dataset is then split into an 80-20 ratio. Two neural network architectures are run three times, each for 60 epochs, with multiple output predictions for each of the desired rock properties. Mean squared error (MSE) with a regularization factor was considered as a loss function when training the model and Mean Absolute Error (MAE) to assess the performance of the model.
Results reveal an effective performance of the developed models in the estimation of rock properties with low MAE values ranging between 0.5-3%. This study demonstrates the potential of convolutional neural networks to predict rock properties from seismic data for efficient reservoir characterization. This paper will be helpful for geoscientists, reservoir engineers, and geophysicists who are dealing with field development plans related to geothermal reservoirs.
| References | Ronneberger, O., Fischer, P., and Brox, T. (2015). “U-net: Convolutional networks for biomedical image segmentation.” MICCAI. Springer, 234-241 |
|---|---|
| Country | India |
| Student Awards | I would like to submit this presentation into both awards |
| Acceptance of the Terms & Conditions | Click here to agree |
