Hybrid modeling of deep neural networks and unsupervised machine learning algorithms for missing well log prediction based on geological lithofacies similarities

Well logging plays a vital role in formation characterization and resource evaluation in oil and gas exploration. However, acquiring well logging data through conventional field methods is often costly and time-consuming, highlighting the need for accurate and cost-effective predictive solutions. To address these challenges, this study introduces a novel hybrid modeling framework that integrates Self-Organizing Maps (SOM), Multilayer Perceptron (MLP), and Social Ski-Driver (SSD) optimization. SOM is employed for unsupervised lithofacies classification based on similarities in well log responses, and these lithofacies, along with other well log inputs, serve as key features for the supervised MLP model. The SSD algorithm optimizes the MLP's weights and biases, further enhancing its performance. The results demonstrate that the hybrid SOM-MLP-SSD model significantly improves the accuracy of missing log predictions, particularly in lithologically complex and hydrocarbon-bearing zones where traditional methods, such as the Greenberg-Castagna equation, fall short. To benchmark this approach, a SOM-SVM model was also tested to evaluate how another established machine learning algorithm performs with the same facies-guided structure. While SOM-SVM produced reasonable results, the SOM-MLP model consistently achieved more reliable and accurate predictions. The model also incorporates uncertainty quantification using least-squares estimation, increasing prediction robustness. This methodology offers a significant advancement in subsurface characterization by combining geological insights with advanced machine learning and optimization techniques. The hybrid approach enhances prediction accuracy, providing valuable insights for geomechanical analysis, reservoir evaluation, and decision-making in hydrocarbon exploration. The proposed model represents a promising step forward in improving the accuracy of missing log predictions and optimizing resource extraction strategies in complex reservoirs, facilitating more efficient and cost-effective exploration and development.