A frequent challenge is the absence of sonic log data for various reasons, which calls for effective solutions. One practical approach to address this issue is the estimation or prediction of missing well logs, commonly referred to as soft logging, which can significantly lower exploration costs in the oil and gas industry.

     Given the structural complexity and heterogeneity of geological reservoirs, there are often pronounced nonlinear relationships among different well logs. This study introduces a substitution method that leverages the Random Forest machine learning algorithm for sonic log prediction, to create a dependable model for predicting sonic logs using the available log data.

     The target log type in this research is the Differential Time (DT) log, which indicates the velocity of wave propagation through a geological formation. The features used for training the model include resistivity, density (RHOB), porosity (NPHI), gamma-ray (GR), and the sonic log (DT).

     This study incorporates ILD [Deep induction (Resistivity Deep)] logs and Micro-resistivity (MSFL) logs, with the DT log being blinded in well A-5 for both the training and testing phases, serving as the prediction target within the same well.

     Experimental results indicate that the proposed method (Random Forest) provides a more accurate estimation of missing logs than traditional techniques, demonstrating notable performance. The effectiveness of the sonic log (DT) model prediction is largely attributed to the Random Forest Algorithm, as evidenced by reductions in MSE, RMSE, and MAE to 4.783, 2.187, and 1.351, respectively, alongside an increase in R² to 0.893. Furthermore, the correlation coefficient between the actual and predicted DT logs reached r = 0.99.