High-resolution estimation of porosity with deep-learning-based interpretation of well logging and core data for gas-hydrate-bearing sediments in the Republic of Korea

This study proposes a deep-learning-based algorithm to estimate a high-resolution porosity from well-logging and core data. It is challenging for porosity estimation to integrate well-logging and core data due to their intrinsic differences in data quantity, quality, and resolution. The proposed approach allows direct comparison of well-logging-based porosity with core porosity, thereby overcoming the resolution disparity of previous works. We design a deep neural network that inputs well-logging data (e.g., neutron porosity log, density log, gamma-ray log, and sonic log) and outputs high-resolution porosity through two training stages. The primary training estimates high-resolution porosity, same as that of core porosity, using well-logging data as inputs and empirical porosity calculated using Lee and Colletts formula as outputs of the deep neural network. The trained model is re-trained by correcting the porosity estimates against core porosity. Smoothing is applied to preserve geological continuity. The validity of the proposed algorithm is tested with application to gas-hydrate-bearing sediments in the Ulleung Basin, Republic of Korea. The corrected porosity is not only in better agreement with core porosity than the empirical porosity but also coincides with the empirical porosity overall.