Estimation of Dynamic Petrophysical Properties from Multiple Well Logs Using Machine Learning and Unsupervised Rock Classification
- Mohamed Bennis (The University of Texas at Austin) | Carlos Torres-Verdín (The University of Texas at Austin)
- Document ID
- Society of Petrophysicists and Well-Log Analysts
- SPWLA 60th Annual Logging Symposium, 15-19 June, The Woodlands, Texas, USA
- Publication Date
- Document Type
- Conference Paper
- 2019. held jointly by the Society of Petrophysicists and Well Log Analysts (SPWLA) and the submitting authors
- 27 in the last 30 days
- 217 since 2007
- Show more detail
The process of mud-filtrate invasion gives indirect information about the dynamic petrophysical properties of reservoir rocks, which are essential to predict ultimate hydrocarbon recovery and to optimally design hydrocarbon recovery procedures.
Our focus is on the interpretation of conventional well logs such as density, neutron porosity, and apparent resistivity to estimate dynamic properties of rocks such as permeability and residual hydrocarbon saturation using a machine-learning (ML) algorithm. The inversion problem is mathematically posed as a minimization of a cost function. There exist various approaches to solve this problem such as gradient-based and statistical methods. However, these methods can be computationally expensive, and the process needs to be repeated for each new set of measurements. In this study, we investigate ML methods to automatically detect and extract complex features present in the training dataset.
The dataset is synthetically generated using fast numerical simulations implemented with The University of Texas at Austin Petrophysical and Well-log Simulator (3D UTAPWeLS). We assume a vertical well, horizontal layers, and axial symmetric invasion. Dynamic petrophysical properties of layers are randomly generated using biased distributions. We simulate the process of mud-filtrate invasion and obtain the radial distribution of water saturation and salt concentration. Corresponding well logs such as density, neutron porosity, and apparent resistivity are numerically simulated. This process is repeated until thousands of examples are generated to serve as a database using parallel computations. The performance of the ML model is optimized through the adjustment of hyperparameters or by adding or removing features.
Several synthetic examples describe the successful application of ML methods on the estimation of dynamic petrophysical properties of hydrocarbon-bearing rocks invaded with water-base mud and water-bearing rocks invaded with oil-base mud. The estimated dynamic petrophysical properties are then used as an input for unsupervised rock classification using hierarchical clustering.
Mud-filtrate invasion takes place under overbalanced drilling conditions. In the case of invasion with oil-base mud (OBM), the invasion process involves immiscible fluid displacement between OBM-filtrate and water, whereby no change in water salinity takes place. In the case of invasion with water-base mud (WBM) filtrate, it is necessary to account for the effect of salt mixing between connate water and mud filtrate.
|File Size||1 MB||Number of Pages||11|