Permeability, a key reservoir characteristic, governs the rate of fluid flow through reservoir rocks. Accurate permeability estimates are paramount for robust reservoir simulation, history matching, and production forecasting. Due to limited core data availability and intrinsic heterogeneity of permeability at different scales, establishing reliable permeability models can be challenging. This study aims to overcome these hurdles by predicting lab-measured core permeability from commonly acquired well logs, using various machine learning algorithms such as Support Vector Regression (SVR), Random Forest (RF), XGBoost, and LightGBM.
We examined two diverse datasets, representing a carbonate platform (Costa Field) and clastic formations (Volve Field). The Costa dataset, including 17 wells across a single reservoir, and the Volve dataset, comprising three wells across three different reservoirs, allowed for evaluating the robustness of our approach under different geological conditions. A critical part of our methodology is feature engineering, particularly incorporating vertical variability. We integrated measurements from adjacent well log readings into our models, recognizing the importance of spatial context and the smoothing effect of well logs over small-scale heterogeneities. This improved prediction accuracy by accounting for shared geological history and depositional environments in proximity.
In Costa Field, blind tests showed R2 scores up to 0.64, and validation R2 scores reached up to 0.8 using a leave-one-well-out cross-validation method. For the Volve Field, blind test R2 scores were up to 0.84, 0.76, and 0.78 for Hugin, Sleipner, and Skagerrak formations, respectively. These results, while satisfactory, underscore the potential of machine learning methods in accurately predicting permeability and highlight the need for effective feature engineering.
This work advocates that while machine learning holds promise for automated feature engineering, human intervention, specifically to incorporate spatial context, can still significantly enhance predictions. Future advancements may seek to internalize this spatial awareness within the machine learning algorithms themselves