The successful design of an acid fracture job requires accurate prediction of fractured well productivity. Productivity estimation demands knowledge of both the acid penetration length and conductivity distribution for the given reservoir. The literature includes several models developed to predict the conductivity of acid fractured rock. The most popular is empirical and based on measuring the conductivity of 25 acid fracture experiments. The present research provides empirical models utilizing machine learning techniques and incorporating 97 experiments and 563 datapoints.
We conducted an extensive literature review to collect the published data on acid fracture experiments. The objective of such experiments is to measure conductivity at different formation closure stresses while considering field conditions. We used several data preprocessing techniques to clean the data, fill in missing values, exclude outliers and failed experiments, and standardize the dataset. Regularization was employed to eliminate features that didn't contribute to accurate prediction. Feature engineering was used to construct new features from our dataset. We began by measuring the correlations between features to better understand the data. We then built various machine learning models to predict acid fracture conductivity.
It has been observed that developing one universal empirical correlation often results in significant errors in conductivity estimation because different rock types result in different etching patterns that cannot be explained by a single correlation. For instance, the channeling etching pattern is mostly observed in limestone formations, while a roughness pattern is seen in dolomite and chalk rock. Moreover, the conductivities of etching patterns formed in chalk, dolomite, and limestone formations behave differently. We built machine learning classification techniques to determine the most likely etching patterns (e.g., channeling, roughness). A linear regression-based model was then developed as a baseline for comparison with other models generated through ridge regression. We evaluated the performances of our models using well-known metrics such as accuracy, precision, recall, mean squared error, and correlation coefficients. We also employed cross-validation to avoid over-fitting, finding that certain features were the most important in predicting acid fracture conductivity.
Detailed empirical conductivity correlations and models were developed in this work for three carbonate rock types. Previously, a single empirical model has often been employed to estimate acid fracture conductivity or, at best, a model has been developed for a particular rock type. Most models have not considered the impact of etching patterns on conductivity, which was found to be significant in limestone.