Simulation of carbonate reservoirs acidizing using machine and meta-learning methods and its optimization by the genetic algorithm

Abstract

Optimizing the process of acidizing in carbonate reservoirs is essential to increase the effectiveness of permeability improvement, reduce the pressure drop of near-well bore, and minimize acid consumption. Experimental studies of carbonate reservoirs acidizing and its optimization are time-consuming and expensive. Moreover, numerical simulation of this process, which has been introduced as an alternative to experimental studies, is not practical at industrial scales due to the considerable computation time and the increasing complexities of acidizing itself. Machine learning (ML), neural networks, and meta-learning algorithms can act as a facilitator in simulating and optimizing this complicated process. Previous studies have not used machine learning and meta-learning algorithm to simulate multi-phase carbonate reservoirs acidizing and genetic algorithms for optimizing their operational parameters. In this study, multiphase carbonate reservoirs acidizing is simulated at core scale and validated using published experimental data Because of employing actual data and the nature of artificial intelligence approaches, it is feasible to model the complexities of acidizing in carbonate reservoirs and have acceptable results. Porosity and permeability and their distribution, core dimensions, fluids saturation, oil viscosity, oil density, oil compressibility, injection rate, temperature, pressure, acid concentration, acid molecular diffusion, and reaction rate are the effective parameters in this procedure whose objective function is the acid pore volume to breakthrough (PVBT). After gathering available valid data from the literature, various machine learning methods were employed to model this process. However, due to the problem's complexities, typical machine learning and neural networks methods would not provide the desired accuracy, and more recent approaches such as meta-learning are required to attain reasonable precision. Having modeled the acidizing of carbonate reservoirs, we utilized the genetic algorithm to optimize the operational parameters of this process, i.e., temperature, acid concentration, injection rate, and saturation. Using the stacking method and combining different ML algorithms such as DT, KNN, SVM, and MLP with the MLP algorithm of meta-learning, the prediction accuracy of PVBT improves and reaches 83%. However, 90% accuracy can be attained through combining various artificial neural networks and deep learning models with a neural network algorithm as the meta-learning model. The runtime of the metamodel is 99.9% less than that of numerical simulation; hence, it is used with the genetic algorithm to optimize the operational parameters of injection rate, acid concentration, temperature, and oil saturation.