Porosity prediction from pre-stack seismic data via a data-driven approach

Abstract

Porosity estimation plays an important role in geophysical exploration and reservoir development, which is a highly complicated and extremely challenging problem. Usually, the model-driven inversion for reservoir physical parameters consists of the rock-physics modelling and the probability statistical inversion. However, the rock-physics modelling usually is intricate and the efficiency of inversion algorithm is low. Besides, the relationship between seismic elastic properties and petrophysical parameters is highly non-linear. Therefore, the deep learning algorithm is introduced for two important aspects, rock-physics modelling and reservoir parameters estimation. We build two neural network models, the PorNet model and the BlstmNet model. The former is based on classical convolutional neural networks (CNNs), and the latter is based on the variant network of traditional recurrent neural networks (RNNs). CNNs extract and compress the spatial characteristics between input data and labeled data using multi-layer convolutions, while the bidirectional long short-term memory (BLSTM) neural network conducts porosity prediction by processing the internal correlation of the labeled data. First, we conduct the experiment on synthetic data set for evaluating the performance of these two neural network models on the prediction accuracy and the convergence speed. The experimental results show that the BlstmNet model provide a more precise prediction result and the PorNet model has a more stable convergence process. To further evaluate the generalization ability for two neural network models, an example for porosity prediction is carried out on oil field data. The prediction results demonstrate that two neural network models have ability to estimate reservoir properties with high accuracy and lower artificial cost. The data-driven inversion of reservoir physical parameters has certain feasibility and application prospects.