On the Path to Least Principal Stress Prediction: Quantifying the Impact of Borehole Logs on the Prediction Model

Knowledge of the minimum horizontal principal stress (Shmin) is essential for geo-energy utilization. Shmin direct measurements are costly, involve high-risk operations, and provide only discrete values of the required quantity. Other methods were developed to interpret a continuous stress sequence from sonic logs. These methods usually require some 'horizontal tectonic stress' correction for calibration and rarely match sections characterized by stress profiling due to viscoelastic stress relaxation. Recently, several studies have tried to predict the stress profile by an empirical correlation corresponding to an average strain rate through geologic time or by using machine learning technologies. Here, we used the Bayesian Physics-Based Machine Learning framework to identify the relationships among the viscoelastic parameter distributions and to quantify statistical uncertainty. More specifically, we used well logs data and ISIP measurements to quantify the uncertainty of the viscoelastic-dependent stress profile model. Our results show that the linear regression approach suffers from higher uncertainty, and the Gaussian process regression Shmin prediction shows a relatively smaller uncertainty distribution. Extracting the lithology logs from the prediction model improves each method's uncertainty distribution. We show that the density and the porosity logs have a superior correlation to the viscoplastic stress relaxation behavior.