Extracting IP parameters of rock samples using machine learning

Abstract

SUMMARY The induced polarization (IP) phenomenon describes the variation of resistivity with frequency and this geophysical method has been widely used to classify rock properties. Various experimental models have been developed to describe the mechanism of IP effect. The model based on the generalized effective-medium theory of induced polarization (GEMTIP) is an extension of the classic Cole–Cole model by considering electromagnetic inductions. Compared to the Cole–Cole model, the GEMTIP model can effectively incorporate rock composition and morphology to provide a more precise complex resistivity response. However, adding one more grain type to the GEMTIP model will result in adding three more parameters to the inversion process. Even for the GEMTIP model with only three-phase spherical grains, seven parameters need to be inverted. In general, the inversion of GEMTIP model with more than three phases is characterized by high uncertainty. Traditional geophysical inversion methods such as global optimization and least-squares optimization have certain limitations in inverting IP parameters. The effectiveness of least-squares optimization algorithm depends on the setting of initial inversion parameters. Efficiently processing massive amounts of data is challenging for both least-squares optimization and global optimization algorithms. Considering the powerful search and generalization capabilities of neural network (NN), we develop a machine learning (ML) approach for efficient inversion of IP parameters based on long short-term memory (LSTM) networks with physical constraints. Compared with traditional methods, ML method can significantly reduce the computational cost and produce reliable inversion results. To validate the effectiveness of ML approach, we first conduct multiple synthetic model studies based on the Cole–Cole and the GEMTIP models, respectively. We then further demonstrate the effectiveness of NNs in recovering IP parameters from rock samples using measured data from man-made and field rock samples. The results show that the developed method can effectively obtain accurate inversion results based on Cole–Cole and GEMTIP models.