Supervised Machine Learning and Multiple Regression Approach to Predict Successfulness of Matrix Acidizing in Hydraulic Fractured Sandstone Formation

Abstract

The success rate of matrix acidizing in hydraulic fractured sandstone formation is less than 55%, much lower compared to the more than 91% success rate in carbonate formation. The need for alternative approaches to help the success ratio in matrix acidizing is crucial. This paper demonstrates a modeling technique to improve the success ratio of matrix acidizing in a hydraulic fractured sandstone formation. Supervised machine learning with 4 models of a neural network, logistic regression, tree, and random forest was selected to predict the successfulness of matrix acidizing in hydraulic fracturing. In parallel, multivariate analysis of principal component regression and partial least square regression approach were utilized to predict the oil gain of the job. For qualitative prediction, the results showed that the random forest was the best model to predict the successfulness of the job with the area under the curve (AUC) of 0.68 and precision of 0.73 in the training model with 70% of the data. Subsequently, the validation test with the rest of the data (30% data) gave 0.51 AUC and 61% precision. For quantitative prediction, the net oil gain was evaluated by using principal component regression (PCR) and partial least square regression (PLS-R). The PCR and PLS-R model gave a coefficient of determination (Rsquare) of 0.22 and 0.35, respectively. The p-value of PLS-R was 0.047 (95% confidence interval) which indicates that the model is significant. The results of this work demonstrate the potential application of supervised machine learning, principal component regression, and partial least square regression to improve candidate selection of oil wells for matrix acidizing especially in hydraulic fractured wells with limited design data.

Optimizing acidizing design and effectiveness assessment with machine learning for predicting post-acidizing permeability

“…XGBoost enhances performance and reduces overfitting, whereas random forest combines decision trees for robust predictions. Overall, these models were chosen due to their capabilities in handling the complexities of acidizing and their track record of accurate predictions 17 , 24 , 25 , 28 , 30 , 34 , 35 , 44 .…”

“…Gumrah et al describe a computer model that uses a genetic algorithm to optimize Damkohler and acid capacity numbers for predicting the permeability alteration of an acidization process 31 – 33 . Alkathim et al investigated the impact of rock, acid, and reaction properties on pore volume to breakthrough during calcite matrix acidizing, finding optimal injection rates 34 , while Kurniawan proposed a machine learning and regression analysis model to enhance success rates and net oil gain in hydraulic fractured sandstone formations, improving candidate selection 35 . Additionally, Abdollah Hatamizadeh and Behnam Sedaee optimized acidizing processes in carbonate reservoirs using neural networks, meta-learning algorithms, and genetic algorithms, achieving high simulation accuracy and minimizing acid consumption while enhancing permeability improvement 17 .…”