[Paper Review] Machine Learning for Gas and Oil Exploration

Vito Alexander Nordloh, Anna Roubícková+1 more|arXiv (Cornell University)|Oct 4, 2020

Reservoir Engineering and Simulation Methods3 references4 citations

TL;DR

This paper proposes a machine learning approach to predict missing values in oil and gas well logs using Gradient Boosting and neural networks, leveraging geologically proximate wells as training data. The key contribution is demonstrating that local training on nearby wells improves prediction accuracy, with Gradient Boosting showing more stable performance than neural networks despite similar error metrics.

ABSTRACT

Drilling boreholes for gas and oil extraction is an expensive process and profitability strongly depends on characteristics of the subsurface. As profitability is a key success factor, companies in the industry utilise well logs to explore the subsurface beforehand. These well logs contain various characteristics of the rock around the borehole, which allow petrophysicists to determine the expected amount of contained hydrocarbon. However, these logs are often incomplete and, as a consequence, the subsequent analyses cannot exploit the full potential of the well logs. In this paper we demonstrate that Machine Learning can be applied to \emph{fill in the gaps} and estimate missing values. We investigate how the amount of training data influences the accuracy of prediction and how to best design regression models (Gradient Boosting and neural network) to obtain optimal results. We then explore the models' predictions both quantitatively, tracking the prediction error, and qualitatively, capturing the evolution of the measured and predicted values for a given property with depth. Combining the findings has enabled us to develop a predictive model that completes the well logs, increasing their quality and potential commercial value.

Motivation & Objective

• Address the challenge of incomplete and noisy well log data in oil and gas exploration, which limits accurate hydrocarbon yield estimation.
• Improve the quality and commercial value of well logs by filling in missing values using machine learning.
• Investigate how training data size and geographical proximity affect model performance in predicting missing well log properties.
• Compare the predictive accuracy and stability of Gradient Boosting and neural network models for well log imputation.
• Demonstrate that quantitative error metrics alone may be misleading and advocate for qualitative visual inspection of predictions.

Proposed method

• Use supervised machine learning to predict missing values in well logs by treating one physical property as the target variable and others as input features.
• Train models on subsets of well log data where the target property is measured, using a disjoint test set for evaluation to avoid data leakage.
• Apply two regression models: Gradient Boosting (XGBoost-like) and feedforward neural networks, with hyperparameter tuning for optimal performance.
• Evaluate model performance using standard metrics: Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE), and supplement with visual analysis of predicted vs. measured values across depth.
• Test local training strategies by incrementally adding the nearest geographically proximate wells to the training set to assess impact on prediction error.
• Use correlation heatmaps to analyze feature relationships across wells and assess whether global correlations mask local geological variations.

Experimental results

Research questions

• RQ1 How does the amount of training data, particularly from geographically proximate wells, affect the accuracy of missing value prediction in well logs?
• RQ2 Which machine learning model—Gradient Boosting or neural networks—yields more accurate and stable predictions for missing well log values?
• RQ3 To what extent do standard error metrics like MSE and MAPE accurately reflect the quality of predictions in well log imputation?
• RQ4 Can local training on nearby wells improve model performance compared to training on a single well or the entire dataset?
• RQ5 How do geological variations at small spatial scales affect the generalization and reliability of machine learning models in well log prediction?

Key findings

• Training on all wells simultaneously does not improve model accuracy and may introduce spurious correlations due to geological heterogeneity.
• The addition of nearby wells to the training set does not consistently reduce prediction error, as geological differences at small spatial scales confuse model learning.
• Gradient Boosting models exhibit more stable performance across different wells and hyperparameter settings compared to neural networks, which show high variability in results across runs.
• For most well log properties, Gradient Boosting models achieve over 90% accuracy in predicting gaps of average size, indicating strong predictive capability.
• Quantitative metrics like MAPE and MSE can be misleading; visual inspection of predictions across depth reveals that models may appear accurate in error metrics but fail to capture meaningful geological trends.
• The study concludes that improving model accuracy requires better evaluation metrics and higher-quality input data rather than more complex models.