Integration of Physics-based Models and Data-driven Models in Drilling Technology

Abstract

This review addresses the challenges in drilling technology, particularly as systems become more complex and the phenomena within wells are hard to predict or measure. For such application, pure physics models struggle with lack of fidelity and uncertainty, while data-driven models lack generalizability and require large datasets often unavailable in drilling. Various combinations of physics and machine learning are being developed to enhance the strengths and mitigate the weaknesses of each method. The trend can be categorized in two paradigms, Bayesian Filtering and physics-informed machine learning (PIML). This comprehensive review explores various techniques applied to drilling technologies, particularly two fields, ultra-deep drilling and early stuck pipe detection. Successfully applying these hybrids hinges on carefully selecting the model based on data availability and confidence in physical knowledge. Overall, this review emphasizes the importance of hybrid models in advancing drilling technology by combining physical insights with machine learning capabilities.