Simulating Geophysical Logging Data Set for Machine Learning

Abstract

Summary Intelligent geophysical logging inversion based on data-driven machine learning can efficiently realize traditional logging interpretation and formation evaluation, and has broad application prospects. Compared with commercial artificial intelligence, intelligent logging faces the problem of small samples and low-quality labels. It is necessary to study the representative labeled logging data sets for the application scenario of intelligent logging. However, for geophysical logging, the observed downhole strata cannot be seen or touched, coupled with the multi-solution of inversion and the heterogeneity of strata, the labeled system constructed from the measured data set is not only a small quantity but also doubtful in reliability. Based on geophysical knowledge and response functions, we construct the geophysical logging machine learning data set through the forward modeling method, to meet the needs of generating a large number of labeled training data. The generated data set is based on the virtual stratum, borehole, and instrument conditions, effectively realizing data privacy protection. The experimental results show that the deep learning model trained by generated data set has great performance. The data set generation method has a remarkable effect on developing data-driven and data-model hybrid driving methods and improving the effect of logging reservoir evaluation parameter prediction model.