Evaluating the Effects of Acid Fracture Etching Patterns on ConductivityEstimation Using Machine Learning Techniques

Author ORCID Identifier

Maya Mincheva:https://orcid.org/0000-0001-5415-9613

Publication Title

Society of Petroleum Engineers - SPE Europec Featured at 82nd EAGE Conference and Exhibition



Document Type

Conference Proceeding


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 conductivitydistribution for the given reservoir. The literature includes several models developed to predict theconductivity of acid fractured rock. The most popular is empirical and based on measuring the conductivityof 25 acid fracture experiments. The present research provides empirical models utilizing machine learningtechniques 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 whileconsidering field conditions. We used several data preprocessing techniques to clean the data, fill in missingvalues, exclude outliers and failed experiments, and standardize the dataset. Regularization was employedto eliminate features that didn't contribute to accurate prediction. Feature engineering was used to constructnew features from our dataset. We began by measuring the correlations between features to better understandthe 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 significanterrors in conductivity estimation because different rock types result in different etching patterns that cannotbe explained by a single correlation. For instance, the channeling etching pattern is mostly observedin limestone formations, while a roughness pattern is seen in dolomite and chalk rock. Moreover, theconductivities 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 comparisonwith other models generated through ridge regression. We evaluated the performances of our models usingwell-known metrics such as accuracy, precision, recall, mean squared error, and correlation coefficients. Wealso employed cross-validation to avoid over-fitting, finding that certain features were the most importantin predicting acid fracture conductivity.

Publication Date





Department of Computer Science