Abstract
Efficiently designed gel treatments play a vital role in extending the lifespan of brownfields through rejuvenating oil production. Recently, a three-mode mathematical methodology named the VCR approach has been proposed for designing effective treatments. To optimize this approach, it is crucial to determine the appropriate design mode systematically rather than relying solely on the intuitive judgments of field operators. This study introduces an advanced methodology for predicting the optimal design type of gel treatments using 12 reservoir and production variables. The methodology integrates ensemble machine-learning (EML) models with historical data from 65 field projects across 11 countries (1985-2020). The Random Forest (RF) algorithm was utilized due to its resistance to overfitting, thereby mitigating the issue of poor model generalization caused by data scarcity. RF models were also built using synthetic data intelligently generated by the MOSTLY AI platform to enhance generalization further. Data preprocessing examined 22 variables for outliers and missing values that were imputed using the MICE algorithm. A hybrid approach was utilized to select input variables, considering design considerations, data availability, statistical significance, and contributions to RF models. Synthetic data was generated and verified using univariate and bivariate quality measures. Four RF models were built using real and synthetic data, and their performance was evaluated on holdout datasets using global performance metrics. The results demonstrate that RF models accurately predict the design type of gel treatments, achieving validation-sample accuracies of 92% and 100% for the original and synthetic datasets, respectively. The consistent accuracy observed between training and validation samples underscores the robust generalization of developed models. Significant enhancements in stability, predictivity, and generalizability were achieved through synthetic data up sampling. The prediction of design type is primarily influenced by average permeability, recovery factor, and formation temperature. An Excel dashboard, OptiGel-RF, was created to facilitate the usability of models and made publicly available. This novel methodology optimizes gel treatment design, improves success rates, and provides actionable insights for addressing design errors. It paves the way for future applications of machine learning and generative AI in reservoir engineering.
Recommended Citation
M. Aldhaheri et al., "Predictive Modeling for Optimal Gel Treatment Design in Brownfields using Ensemble Machine Learning and Data Upsampling via Generative AI," Society of Petroleum Engineers SPE Conference at Oman Petroleum and Energy Show Opes 2025, Society of Petroleum Engineers, Jan 2025.
The definitive version is available at https://doi.org/10.2118/225111-MS
Department(s)
Geosciences and Geological and Petroleum Engineering
Second Department
Chemical and Biochemical Engineering
Publication Status
Available Access
International Standard Book Number (ISBN)
978-195902574-0
Document Type
Article - Conference proceedings
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2025 Society of Petroleum Engineers, All rights reserved.
Publication Date
01 Jan 2025
Included in
Biochemical and Biomolecular Engineering Commons, Geology Commons, Petroleum Engineering Commons
