Physics Augmented Correlations and Machine Learning Methods to Accurately Calculate Dead Oil Viscosity Based on the Available Inputs

Abstract

Summary Viscosity is one of the key parameters that is necessary to describe the fluid flow and pressure propagation in porous media as well as flow in pipes and various types of conduits. In this study, we have developed two different approaches to be able to estimate the viscosity seamlessly from a fraction to a million centipoise range. Our first approach was based on a trainable explicit expression that uses a reference viscosity measured at any temperature. This is a very common situation as the viscosity may have already been measured at reservoir temperature. Using the reference viscosity concept for many data sets, we also demonstrated the validity and tuneability of the proposed approach to predict the viscosities for the entire temperature range of interest. Our second method is developed for the cases that we did not have any viscosity measurements or no valid viscosity measurements. For such circumstances, we have utilized the extreme gradient boost (XGB) method algorithm (XGB Method) (Friedman 2001; Chen et al. 2015) using leave-one-out cross-validation (LOOCV) (Shao 1993), to be able to estimate the viscosity for the entire range of viscosities that are encountered for crudes. Some of the details of hybrid methods and other regression/machine learning methods and the role of the asphaltenes are already discussed by Sinha et al. (2019, 2020). The current approach uses a richer data set than any of the published papers with very limited input parameters: reservoir temperature, API gravity, and molecular weight, and the tuneability of the models is constrained with the molecular weight of the crudes.