Improving the pulsed neutron-gamma density method with machine learning regression algorithms

Abstract

Formation bulk density is one of the critical parameters for formation evaluation. As a safe and environmentally friendly method, the pulsed neutron-gamma density (NGD) measurement is emerging as an alternative to the traditional chemical source-based gamma-gamma density (GGD) measurement. However, in the NGD measurement, the initial energy spectrum, source intensity, and spatial distribution of the secondary inelastic gamma-ray source vary with formation components. Moreover, the energy of inelastic gamma rays is at the MeV level, leading to a non-negligible pair production effect. All these factors affect the accuracy of the bulk-density measurement. In addition, the impact of borehole configuration on NGD is evident, and correction for borehole effects is essential. To improve density accuracy: first, we forgo exploring an explicit theoretical formula for density calculation and instead treat the NGD mathematically as a regression problem and introduce the machine learning regressor, a powerful and popular tool for solving regression problems, into the NGD for the first time; second, we select features less affected by changes in formation chemical composition as input features. Our results show that the final three tuned machine learning regressors selected from the 39 candidate regressors outperform the optimized polynomial model (an optimization of conventional density calculation models) in both accuracy and generalization ability. They complete density prediction and borehole correction in one step, avoiding complex exploration of borehole effects as in the optimized polynomial model and dramatically simplifying the borehole correction process. Moreover, the GaussianProcessRegressor can complete borehole correction without the standoff information and perform well, which is impossible for the optimized polynomial model, broadening application scenarios.