“Sharing Private” Multi-task Learning for Petrophysical Parameters Prediction with Logs

Abstract

Using neural network to map the relation between logging data and petrophysical parameters has been studied actively in recent years (Korjani, 2016). The results show that neural network can predict petrophysical parameters based on logging data with higher efficiency and accuracy than traditional model-driven methods (Korjani, 2016). The existing study, however, single predict neural network were used, that is, for a neural network one petrophysical parameter can be predicted, such as porosity (POR) or water saturation (SW) with a set of logging data. We propose a multi-task machine learning method for petrophysical parameter prediction with logs, which can improve the efficiency, simplify the process and reduce the mean absolute error compared with single predict neural network.