Synthetic well log generation based on SBOA-XGBoost and SHAP

Abstract Well logging plays a critical role in characterizing reservoir petrophysical properties and is essential for hydrocarbon exploration and development. Nevertheless, the absence of well logs is a common issue during development, which can lead to a decrease in identification accuracy. A novel method for synthesizing missing well logs based on SBOA-XGBoost and Shapley Additive Explanations (SHAP) was proposed, where the missing well logs were reconstructed using available conventional well logs. The method can efficiently learn subsurface patterns and temporal trends from conventional well logs. SBOA-XGBoost was applied to synthesize natural gamma ray (GR), gamma ray without uranium (KTH), photoelectric (PE), compensated density (DEN), compensated neutron (CNCF), and compressional wave velocity (DTC) logs. For comparison, SBOA-XGBoost was compared with Fully Connected Neural Networks (FCNN), Bidirectional Long Short-Term Memory Networks (BiLSTM), and Convolutional Bidirectional Long Short-Term Memory Networks (CNN-BiLSTM) to evaluate its performance. Mean absolute error, root mean square error, and the coefficient of determination (R2) were utilized to objectively evaluate the predictive capability of the models. Synthesis accuracy was the highest for GR, KTH, DEN, and DTC, as indicated by their R2 values of 0.95, 0.92, 0.93, and 0.90. Reasonable accuracy was achieved for PE and CNCF, with R2 values of 0.82 and 0.83, respectively. SBOA-XGBoost surpassed the other methods in synthesizing all types of well logs. An interpretability analysis of the SBOA-XGBoost model was performed using SHAP, demonstrating that the model’s decision-making mechanism aligns with domain knowledge.