Machine Learning and the Analysis of High-Power Electromagnetic Interaction with Subsurface Matter

Abstract

Abstract This work is an ongoing effort to design a numerical platform based on machine learning algorithms to characterize, predict, optimize and guide the interaction of [high power] electromagnetic (HPEM) sources (laser, microwave, RF, etc.) with subsurface matter (e.g. rocks, oils, brines, etc.). Advanced statistical analysis routines are essential to identify key variables and relations in the thermal- mechanical-electromagnetic coupling in heterogeneous and anisotropic materials. Advanced statistical analysis and machine learning have been recently used to evince relations in complex environments and physical dynamics; e.g. fluid dynamics, P&ID analytics, and drill cuttings classification, to cite a few. The methods make use of sophisticated algorithms to classify and model problems in multiple areas, from image processing to certain optimization problems. In the realm of subsurface photonics, and in particular for high power electromagnetic (HPEM) interaction with subsurface matter, these routines could become essential to identify key variables, assess the environment and process, and evince models to predict the outcome of an inherently multiphysics and multi-dimensional problem. Numerical models that capture the interaction between HPEM sources and subsurface matter are essential to predict, optimize, adapt, and evaluate the process prior to, and during, deployment in subsurface. These models can come as the solution to a set of coupled partial differential equations that fully describe the physical dynamics, or as the result of supervised-learning algorithms and analysis of experimental and field data. The former is highly sensitive to dynamic material properties, environmental conditions, and source parameters. In addition, it can be challenging to characterize the properties of subsurface materials over the wide range of temperatures and pressures observed in the process. Thus, a machine learning method could provide an ever-improving alternative that learns from the available data to build a numerical platform that can predict, optimize, and guide the process. Machine learning and advanced statistics provide a compelling alternative to build numerical tools to predict, optimize, and control physical processes. This work introduces a variety of numerical approaches to identify essential variables, predict their impact, and optimize the outcome for subsurface applications. Combined, the methods described in this work can help guide the control of the governing dynamics and parameters for use in multiple applications. This numerical platform can be extended to other applications, enhance experimental prototypes, and advance the design of a comprehensive numerical tool for downhole HPEM operations.