PREDICTING SUBSURFACE TEMPERATURE FROM WELL LOGS USING MACHINE LEARNING
Faculty
Department
Year of Publication
Keyword
Publication Type
Abstract
Predicting the subsurface temperature distribution within sedimentary and petroleum‐bearing formations is essential for accurate hydrocarbon maturation modeling, well‐bore stability, and drilling‐fluid design. Traditional approaches—relying on sparse bottom‐hole temperature (BHT) measurements and one‐dimensional conductive models—often misestimate true formation
temperatures by 5–10 °C in heterogeneous settings such as the Niger Delta. To address these limitations, this study develops a data‐driven workflow that employs machine learning algorithms trained on routinely acquired drilling logs to produce continuous, high‐resolution temperature profiles. First, we assembled a dataset from Niger Delta wells comprising wireline logs (gamma‐ray, four‐pad resistivities, density, neutron porosity, sonic velocity), drilling parameters (equivalent circulating density, rate of penetration, exposure time), and corrected BHT readings. After replacing sentinel values (–9999) with NaNs and depth‐referencing all curves to true vertical depth (TVD), each log was clipped to its 1st–99th percentile range to mitigate extreme outliers. Features were standardized to zero mean and unit variance. Derived attributes—such as depth‐derivatives and moving‐window averages—were also explored to enhance lithofacies and thermal signal detection. We compared three regression models: a multilayer perceptron neural network (ANN), a Random Forest (RF) ensemble, and Support Vector Regression (SVR). Hyperparameters were tuned via grid search with k-fold cross‐validation, and models were evaluated on a hold‐out subset (20 % of depths). Error metrics (RMSE, MAE, R²) and well‐log–style scatter and depth‐track plots quantified predictive performance and bias. The ANN achieved an RMSE of 0.16 °F and R² = 0.97, producing smooth temperature gradients. The RF delivered an RMSE of 0.25 °F and R² = 0.985, with feature‐importance analysis highlighting mid‐pad resistivities and drilling parameters as primary predictors. SVR, with an RMSE of 0.47 °F and R² = 0.955, was less competitive but still captured over 95 % of temperature variance. Well‐log plots demonstrated that both ANN and RF closely track actual temperature profiles across lithologic transitions. This end‐to‐end pipeline—from data cleaning and feature engineering to model training, validation, and interpretation—demonstrates that machine learning can offer accurate, cost‐effective alternatives to physics‐based thermal models. Future work will explore hybrid physics–ML models, temporal drilling‐data integration, expanded feature fusion (e.g; seismic attributes), and explainable‐AI techniques, with the goal of operationalizing these tools in real‐time drilling workflows
temperatures by 5–10 °C in heterogeneous settings such as the Niger Delta. To address these limitations, this study develops a data‐driven workflow that employs machine learning algorithms trained on routinely acquired drilling logs to produce continuous, high‐resolution temperature profiles. First, we assembled a dataset from Niger Delta wells comprising wireline logs (gamma‐ray, four‐pad resistivities, density, neutron porosity, sonic velocity), drilling parameters (equivalent circulating density, rate of penetration, exposure time), and corrected BHT readings. After replacing sentinel values (–9999) with NaNs and depth‐referencing all curves to true vertical depth (TVD), each log was clipped to its 1st–99th percentile range to mitigate extreme outliers. Features were standardized to zero mean and unit variance. Derived attributes—such as depth‐derivatives and moving‐window averages—were also explored to enhance lithofacies and thermal signal detection. We compared three regression models: a multilayer perceptron neural network (ANN), a Random Forest (RF) ensemble, and Support Vector Regression (SVR). Hyperparameters were tuned via grid search with k-fold cross‐validation, and models were evaluated on a hold‐out subset (20 % of depths). Error metrics (RMSE, MAE, R²) and well‐log–style scatter and depth‐track plots quantified predictive performance and bias. The ANN achieved an RMSE of 0.16 °F and R² = 0.97, producing smooth temperature gradients. The RF delivered an RMSE of 0.25 °F and R² = 0.985, with feature‐importance analysis highlighting mid‐pad resistivities and drilling parameters as primary predictors. SVR, with an RMSE of 0.47 °F and R² = 0.955, was less competitive but still captured over 95 % of temperature variance. Well‐log plots demonstrated that both ANN and RF closely track actual temperature profiles across lithologic transitions. This end‐to‐end pipeline—from data cleaning and feature engineering to model training, validation, and interpretation—demonstrates that machine learning can offer accurate, cost‐effective alternatives to physics‐based thermal models. Future work will explore hybrid physics–ML models, temporal drilling‐data integration, expanded feature fusion (e.g; seismic attributes), and explainable‐AI techniques, with the goal of operationalizing these tools in real‐time drilling workflows
Supervisor(s)
co-supervisor