Optimizing Corrosion Inhibitor Package to Mitigate Metal Corrosion and Formation Damage: A Machine-Learning Advisory Model

Abstract Acid fracturing and matrix acidizing are widely used techniques to stimulate productivity of hydrocarbon-producing wells in carbonate and sandstone reservoirs. While increased acid concentration benefits stimulation designs, it also introduces several challenges, including corrosion of the surface and downhole equipment and tubulars. We use a data science approach to predict the corrosion rate (CR) and pitting index (PI) of a metal surface in acid-containing fluid under arbitrary conditions without conducting physical experiments. We explored a database containing more than 10,000 experiments over almost 30 years of laboratory testing. After primary data cleaning, we performed exploratory data analysis to select the most meaningful set of features. Two separate neural networks (NNs) were trained to predict the CR and PI. Validation was conducted on treatments where real-time modification of the corrosion inhibitor recipe was required due to dynamic pumping times associated with limited injectivity. The models were deployed on the cloud to integrate with stimulation design software. A classic set of regression-based machine learning (ML) models were used for training, including linear regression, random forest, and boosting. These models did not produce satisfactory results due to lack of interconnections between additives in one composition. With an alternate approach, we implemented an NN with a multilabel binarization encoding extension by representing each additive by its own concentration. We derived a tailored metric that enables us to minimize the number of incorrect predictions and to overcome singularities introduced by near-zero CRs. There are two incorrect prediction types of any classification model. They are false positive (FP) and false negative (FN). For our problem FP happens when a model predicts a high CR while it is low, and FN is when a model predicts a low CR while it is high. We were able to build a model with an ML approach, while the model cannot be built using first principles techniques. Our approach can be used in any similar domain where a response variable depends on explanatory variables of arbitrary length. The best model we trained had an FP of ~7% and MAE of 0.0344 lb/ft2. In light of the laboratory test outcomes, where the variability was 10%, this result is remarkable. The implemented model was validated on multiple use cases to (1) predict acid-metal interaction with the CR and PI of an arbitrary additives set and physical conditions (pressure, temperature, duration), (2) examine model behavior while sensitizing multiple inputs, and (3) modify fluid formulations on the wellsite when injectivity behavior changes dynamically. It enabled us to reduce the number of laboratory experiments with hazardous acids by 30% and substantially reduced health, safety, and environmental risks.