Digital Rock Physics Combined with Machine Learning for Rock Mechanical Properties Characterization

Saad, Bilal (BHGE) | Negara, Ardi (BHGE) | Syed Ali, Shujath (BHGE)

OnePetro 

Abstract

Rock mechanical properties is essential for several geomechanical applications such as wellbore stability analysis, hydraulic fracturing design, and sand production management. These are often reliably determined from laboratory tests by using cores extracted from wells under simulated reservoir conditions. Unfortunately, most wells have limited core data. On the other hand, wells typically have log data, which can be used to extend the knowledge of core-based mechanical properties to the entire field. Core to log integration of rock mechanical properties and its interpretation is limited by our current understanding of rock physics. The gap is clearly evident where approximations such as empirical relationship between dynamic and static mechanical properties are used for rock failure estimation. This paper presents a hybrid framework that combines advances in digital rock physics (DRP) and machine learning (ML) to predict rock mechanical propertiy (e.g., Young's modulus) from rock mineralogy and texture to improve the accuracy of mechanical properties determined from log data.

In this study, mineralogy, density, and porosity data are obtained from routine core analysis and rock mechanical property from triaxial compression tests. In our methodology, we utilized DRP models which were calibrated against core data and then generate rock mechanical property, for intervals for which triaxial measurements were not available. Mineralogy and texture data are used to create DRP models by numerically simulating rock-forming geological process including sedimentation, compaction, and cementation. Rock mechanical properties derived from DRP are used to enhance the set of training data for the ML algorithm to establish a correlation between rock mineralogy, texture, and mechanical property and construct the ML-based rock mechanical property model. The ML model generates Young's modulus predictions and are compared with the laboratory measurements.

The predicted Young's modulus of rock models from the combined approach has a good agreement with the laboratory measurements. Two quantitative measures for estimation accuracy are calculated and examined including the correlation coefficient and the mean absolute percentage error. Cross-correlation plots between the Young's modulus predicted from the ML model and experimental results show high correlation coefficients and small error. The results of the study show that DRP model can be used to feed the ML model with reliable data so that the prediction accuracy can be improved. The results of this work will provide an avenue of learning from the formation lithology and using the knowledge to predict rock mechanical properties.