Ensemble Deep Learning-Based Porosity Inversion From Seismic Attributes

Abstract

Underground porosity is important in many earth sciences and engineering fields, including hydrocarbon reservoir characterization and geothermal energy production. Popular methods largely rely on the analysis of lithological core data, well logs, and seismic inversion methods. While these methods are reliable, they are also time-consuming, expensive, and difficult to implement. In addition, seismic inversion has nonlinearity, data dimensionality, and non-uniqueness issues. However, deep learning (DL) can provide a more flexible, efficient, and accurate capability by mapping directly from seismic attributes to underground porosity. Therefore, we trained several DL models with different optimization functions. In the training steps, we labelled every seismic attribute data point with its corresponding porosity derived from the well-logs. In contrast to popular ensemble techniques, we proposed a weighted prediction approach based on the strengths of each model. Testing results showed a coefficient of determination ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text{R}^{2}$ </tex-math></inline-formula> ) of 0.94345 and a Pearson’s correlation coefficient of 0.9725 between the actual model and the model of the proposed approach, versus 0.9681 and 0.9716 for the best single and popular ensemble models, respectively. Further, we tested the effectiveness of our method using real seismic data from the North Sea. With a Pearson’s correlation value of 0.9743, the inverted model ranges from 27 to 35%, compared to the reference model, which has an overall range of 20 to 33%. These results provide insights into the potential of the proposed method and its applicability to any other seismic volume to determine spatially varying underground porosity from seismic attributes directly.