This study used production data and a novel machine learning approach utilizing Factor Contribution Analysis (FCA) to highlight geologic sweet spots for multiple US on-shore basins. Each model result was validated against key geologic parameters to establish if the geologic conditions exist for the modeled sweet spots. Further analysis shows how geologic production drivers can change across each play.
Geologic assessments rely primarily on parameters related to tectonic/depositional settings, reservoir storage, saturations, hydrocarbon phase, and wellbore deliverability to define resource play outlines. These same parameters are often used to identify geologic sweet spots and help explain production drivers. Available data resolution varies widely across plays depending on maturity of the play and/or complexity of subsurface relationships. Using only publicly available production and well completion data, XGBoost and SHAP machine learning approaches were used to identify play sweet spots and prepare reservoir quality maps. The focus of this study was on validating the results obtained from machine learning of production variables by using geological information. These geological data were derived from multiple sources including regional interpretations and incorporating geologic parameter cutoffs traditionally used for highlighting geologically favorable areas. Regional play data was provided through public data sources, technical publications, and investor presentations. Parameter cutoffs were overlayed with model results to validate the process.
The machine learning methodology utilizing FCA was used to highlight production sweet spots across multiple US on-shore basins. This study has validated the production-based machine learning results through geologic analysis. The result was a strong correlation between key geologic parameters and model results. Specific relationships are established between the geology and model results that allow for deeper insights to be uncovered regarding changing geologic production drivers across the play.
This analysis has corroborated independently that machine learning of production variables does result in a reliable characterization of reservoir rock quality. This type of analysis has been applied successfully to several unconventional resource plays, and provides significant impetus for intelligent use of explainable machine learning modeling. Moving forward, application of similar approaches can not only validate model results, but also highlight key geologic production drivers. Validation of the machine learning methodology allows users to better answer questions related to completion effectiveness, well evaluations, and development strategies.