Theory-Guided Deep Learning for Reservoir Characterization

Autor: O. Collet, T. Colwell, J. Downton, D. Hampson
Rok vydání: 2020
Předmět:
Zdroj: 82nd EAGE Annual Conference & Exhibition.
DOI: 10.3997/2214-4609.202011797
Popis: Summary Deep Neural Networks (DNNs) are used to predict reservoir properties from well and seismic data. Log properties from the well data serve as the targets while the seismic data serve as the input features. These DNNs require big data. In the reservoir context, we are unlikely to have enough well control in the project area to train the neural network adequately. To address this issue, we use the statistics from the well log data, knowledge of the possible range of reservoir properties and rock physics relationships to generate many pseudo wells and synthetic seismic data to train the DNN. The trained neural network is then applied to the real seismic data. This presentation describes the application of this methodology to a North Sea dataset to predict P-wave and S-wave impedances, porosity and fluid saturation. The elastic property estimates are superior to those obtained from deterministic inversion. The advantage of using augmented data in this deep learning workflow is twofold. First, the extra data allows for deeper networks with a wider range of architectures. Second, it expands the range of geologic scenarios that the DNN is trained on, increasing the range of situations that the DNN operator is appropriate for.
Databáze: OpenAIRE