Predicting Heavy Oil Combustion Kinetics with Machine Learning

Heavy oil resources are becoming increasingly important for the global oil supply, and consequently there has been renewed interest in techniques for extracting heavy oil. Among these, in-situ combustion (ISC) has tremendous potential for late-stage heavy oil fields, as well as high viscosity, very deep, or other unconventional reservoirs. A critical step in evaluating the use of ISC in a potential project is developing an accurate chemical reaction model to employ for larger-scale simulations. Such models can be difficult to calibrate, however, that in turn can lead to large errors in upscaled simulations. Data-driven models of ISC kinetics overcome these issues by foregoing the calibration step and predicting kinetics directly from laboratory data. In this work, we introduce the Non-Arrhenius Machine Learning Approach (NAMLA). NAMLA is a machine learning-based method for predicting O2 consumption in heavy oil combustion directly from ramped temperature oxidation (RTO) experimental data. Our model treats the O2 consumption as a function of only temperature and total O2 conversion and uses a locally-weighted linear regression model to predict the conversion rate at a query point. We apply this method to simulated and experimental data from heavy oil samples and compare its ability to predict O2 consumption curves with a previously proposed interpolation-based method. Results show that the presented method has better performance than previously proposed interpolation models when the available experimental data is very sparse or the query point lies outside the range of RTO experiments in the dataset. When available data is sufficiently dense or the query point is within the range of RTO curves in the training set, then linear interpolation has comparable or better accuracy than the proposed method. The biggest advantage of the proposed method is that it is able to compute confidence intervals for experimentally measured or estimated O2 consumption curves. We believe that future methods will be able to use the efficiency and accuracy of interpolation-based methods with the statistical properties of the proposed machine learning approach to better characterize and predict heavy oil combustion.