Quantitative prediction of median particle size of conventional logged sandstone based on machine learning algorithm

The average rock size is the main identification basis for identifying sedimentary phases, and is an extremely important parameter for reservoir evaluation. However, there is no petrophysical method in the logging data that can directly evaluate the average rock size, and previous work has been done to calculate the average rock size (Md) based on linear and statistical analysis methods using natural gamma logging data for medium to high permeability sandstone reservoirs. However, for low-permeability sandstone reservoirs, the error of the fitted average rock size using linear multiple regression methods is too large for the application of the calculated results. Therefore, the calculation of the median particle size for low-permeability sandstone reservoirs is a difficult problem that needs to be solved. In this paper, for low-permeability sandstone reservoirs, we combine the characteristics of logging data with the learning characteristics of nonlinear mapping of BP neural networks, and select the corresponding loss and activation functions to ensure that overfitting occurs. Using the preferred sensitivity logging parameters, the average rock size is modeled and neighboring wells are predicted by shallow and deep neural networks, and the relative errors of modeling are 7.09% and 0.58%, and the relative errors of prediction are 14.44% and 8.53%, respectively. Its quantitative prediction relative error meets the application requirements. The method takes into account the nonlinear mapping relationship between logging data and the fitting problem of small sample data, and provides a systematic way of thinking for predicting rock size of low-permeability sandstone from logging curves.