A data-driven and machine learning-assisted interpretation of hydraulic fracturing experiments in various formations

Abstract

Hydraulic fracturing has evolved from enhancing oil and gas recovery to addressing challenges in groundwater hydraulics and geo-environmental engineering. This study consolidates data from 30 experimental studies, including both high-permeability formations (cohesionless sands and weakly cemented sandstones) and low-permeability formations (tight sandstones and shales). Machine learning (ML), specifically Random Forest models, is applied to identify the parameters influencing the fracture pressure. Key factors include the mean stress, stress differentials, rock properties, and fluid dynamics. The ML models demonstrate strong predictive performance, highlighting the critical role of stress states and flow conditions in the fracture mechanisms. In the case of weakly cemented rocks and cohesionless sands the stress state is by far the most influential parameter that determines the fracturing behaviour. The findings are contextualized using cavity expansion theory and conventional fracturing criteria, providing deeper insights into the fracture behaviour of the various unconventional rocks. The study identifies gaps in the literature, emphasizing the need for further experiments to refine predictive models. Recommendations for future research aim to improve experimental setups and parameter selection, enhancing the understanding of hydraulic fracturing in unconventional formations.