Globally optimized machine-learning framework for CO2-hydrocarbon minimum miscibility pressure calculations

Abstract

Accurate determination of CO2-hydrocarbon minimum miscibility pressure (MMP) is critically important for CO2 geological storage and utilization in oil and gas reservoirs. Here we propose a machine-learning framework, the conditional generative adversarial network, together with the Bayesian optimization algorithm, to calculate the CO2-hydrocarbon MMPs. A total of 180-set MMP data are collected from the public resources to facilitate and validate the proxy model. Also, 21 MMP-influential factors covering fluid compositions and operating conditions are specifically evaluated to analyse their effects on the MMP. In comparison with the existing artificial neural network as well as support vector regression models based on radial basis function kernel and polynomial function kernel, the newly-proposed model does not only outperform with the lowest calculation error (MAPE of 6.81% and MSE of 3.2006), also vividly reflect the interactive relationships of each influential factor and the MMP