AI-Driven Petrophysical Predictions: Optimizing Performance in Reduced Dataset Environments

Abstract The objective of the study is to enhance the prediction of permeability by leveraging of advanced artificial intelligence technologies in reservoirs with limited datasets. Initially, the paper will focus on employing transfer learning techniques to establish analogies between the current reservoir and geologically similar reservoirs with larger datasets. In cases where no suitable analogy could be identified, ensemble methods were utilized to enrich the dataset by constructing robust predictive features. To address imbalanced datasets, the study focused on using data augmentation to expand the current dataset. These approaches aim to optimize performance, reduce uncertainty and deliver robust predictions in data constraint reservoirs. The study presented in this paper uses a multi-faceted approach to address the challenges of petrophysical attributes prediction in reservoirs with limited datasets. First, several transfer-learning tools, including Procrustes disparity, Kullback-Leibler divergence, Silhouette scores analyses are applied robustly identify analogies between the target reservoir and external datasets. For the reservoirs where no analogies exist, ensemble methods are used to generate robust synthetic features to enrich the dataset. In datasets that may be biased or imbalanced, techniques such as Synthetic Minority Oversampling Technique (SMOTE) are employed to create representative samples. This workflow integrates these techniques into a streamline process, starting with data processing, analysis, followed by model training and validation using statistical metrics to ensure reliability and robustness of the predictions. The deployment of these AI methodologies in permeability prediction has yielded promising results. The application of transfer learning techniques has clearly identified analogies between reservoirs, significantly improving prediction accuracy. In reservoirs lacking direct analogies, ensemble and data augmentation methods have effectively reduced bias and enhanced model robustness. A study conducted on number of fields have shown that models trained using this methodology have greatly improved applicability and performance compared to traditional methods. This study introduces a novel framework that moves beyond traditional approaches to identify analogies in reservoirs, which often relied on statistical comparisons or subjective geological expertise. By integrating these AI methods, the study provides a quantifiable and objective method to establish a deeper understanding of relationships between reservoirs, allowing for improved predictions. The study also addresses the critical point where data scarcity affects model reliability and how data augmentation and ensemble methods can be employed to improve model performance. The ability to integrate these advanced tools represents a significant leap forward in decision-making processes in field development and management.