Machine learning application to predict in-situ stresses from logging data.

Determination of in-situ stresses is essential for subsurface planning and modeling, such as horizontal well planning and hydraulic fracture design. In-situ stresses consist of overburden stress (σ), minimum (σ), and maximum (σ) horizontal stresses. The σ and σ are difficult to determine, whereas the overburden stress can be determined directly from the density logs. The σ and σ can be estimated either from borehole injection tests or theoretical finite elements methods. However, these methods are complex, expensive, or need unavailable tectonic stress data. This study aims to apply different machine learning (ML) techniques, specifically, random forest (RF), functional network (FN), and adaptive neuro-fuzzy inference system (ANFIS), to predict the σ and σ using well-log data. The logging data includes gamma-ray (GR) log, formation bulk density (RHOB) log, compressional (DTC), and shear (DTS) wave transit-time log. A dataset of 2307 points from two wells (Well-1 and Well-2) was used to build the different ML models. The Well-1 data was used in training and testing the models, and the Well-2 data was used to validate the developed models. The obtained results show the capability of the three ML models to predict accurately the σh and σH using the well-log data. Comparing the results of RF, ANFIS, and FN models for minimum horizontal stress prediction showed that ANFIS outperforms the other two models with a correlation coefficient (R) for the validation dataset of 0.96 compared to 0.91 and 0.88 for RF, and FN, respectively. The three models showed similar results for predicting maximum horizontal stress with R values higher than 0.98 and an average absolute percentage error (AAPE) less than 0.3%. a index for the actual versus the predicted data showed that the three ML techniques were able to predict the horizontal stresses with a deviation less than 20% from the actual data. For the validation dataset, the RF, ANFIS, and FN models were able to capture all changes in the σ and σ trends with depth and accurately predict the σ and σ values. The outcomes of this study confirm the robust capability of ML to predict σ and σ from readily available logging data with no need for additional costs or site investigation.