Paper submitted to URTeC 2025, by Charles Connell, Kyle LaMotta (PetroAI), Jizhou Li (ExxonMobil) and Matt Mayer (TGS)
Summary
This paper presents an automated modeling system designed to predict well performance in unconventional oil development, incorporating geological and geomechanical property variations in the Delaware Basin, particularly in Lea and Eddy Counties. The approach addresses the challenge of quantifying spacing degradation effects within heterogeneous reservoirs, where traditional analysis methods often reveal noisy relationships between well spacing and performance.
The methodology integrates diverse datasets including well logs, directional surveys, and production data, enabling data-driven analysis that minimizes interpretation biases. The process consists of multiple steps: first, generating machine learning ready well logs, then interpolating those well logs to create property maps, then developing a spatial clustering framework to identify distinct regimes within the Wolfcamp formation. Further analysis is done to quantify the critical distances at which well interference significantly impacts production in each cluster.
The research culminates in a spacing-aware machine learning model that achieved 78.5% accuracy in predicting normalized production outcomes. Sensitivity testing across varying well densities (1-12 wells per section) in different geological settings demonstrated variations in production degradation patterns across clusters. The findings reveal between 10% and 25% degradation at five wells per section, compared to a standalone well depending on which cluster the wells are placed. This approach equips engineers with quantitative tools to optimize well spacing decisions, prioritizing location-specific geological characteristics over basin-wide trends.