Machine Learning Framework for Synthetic Well Log Generation Using Formation-Specific Models

Summary Characterizing rock mechanical properties for drilling optimization requires complete sonic and density logs, yet operational and economic constraints frequently result in incomplete datasets. This study presents a machine learning framework that generates synthetic well logs using offset data to address these gaps. The methodology combines drilling parameters and gamma ray measurements as input variables, employing sequential processing to preserve depth-related geological relationships. The framework offers flexibility to incorporate additional data types when available, enabling enhanced model performance. Model development progresses from whole-well approaches to formation-specific modeling that better captures geological heterogeneity. A leave-one-out cross-validation approach ensures robust model evaluation. Three progressive model configurations demonstrate systematic performance improvements: the baseline whole-well model shows moderate prediction capability; formation-specific spatial modeling significantly improves accuracy by capturing geological heterogeneity; and incorporating additional resistivity measurements further enhances prediction performance. Results confirm that formation-specific approaches with expanded input parameters substantially outperform conventional whole-well modeling methods. The framework’s adaptability enables deployment across different operational scenarios, from real-time drilling applications using readily available data to comprehensive post-drilling analysis with expanded datasets. This flexibility maintains technical standards required for geomechanical analysis while providing cost-effective alternatives to traditional logging programs, demonstrating practical value.