Relevance Based Transfer Learning for Reservoir Parameters Prediction with Logs

Abstract

Traditional methods for reservoir parameters prediction with logs are based on petrophysics knowledge, such as volumetric models and response functions (Darwin and Julian, 2007). The advantage of those methods is that the relationship between the logs and reservoir parameters is clear, there is a theoretical basis, and it can be explained; the disadvantage is that the response functions can only be constructed for the known physical relationship, and the unknown physical relationship may be ignored. Whereas using neural network to predict reservoir parameters, we can map the relationship between logs and reservoir parameters as long as building a suitable model and have a large number of training data. With the help of neural network, we can map the unknown physical relationship without much geological expertise. The existing research of reservoir parameter prediction neural network with logs only focuses on one kind of reservoir parameter prediction modeling, ignoring the relationship between reservoir parameters. In this paper, relevance transfer learning is introduced, which using the knowledge of petrophysics to improve the performance of neural network reservoir parameters prediction.