Application of Machine Learning for Reservoir Identification and Characterization in an Onshore Oilfield, Niger Delta, Nigeria

Reservoir characterization in mature Niger Delta fields is challenged by heterogeneous lithology and limited labeled data, where conventional supervised or unsupervised machine learning (ML) methods often underperform. This study addresses this gap by proposing a hybrid semi-supervised learning (SSL) framework tailored for datasets with mixed labeled and unlabeled well-log outcomes. We integrate cluster analysis, random forest (RF), and SSL to analyze five subsurface intervals using porosity, resistivity, gamma ray, and hydrocarbon saturation logs. Cluster analysis objectively delineates reservoir units, RF prioritizes predictive features, and SSL (self-training, neural networks, and KSVM classifiers) bridges data scarcity. Cluster analysis consolidated five intervals into three geologically viable reservoirs ranked by log-response robustness (Reservoir 1 > 2 > 3). The RF model has achieved 85.7% prediction accuracy, identifying resistivity (>20 Ωm) as the most critical hydrocarbon indicator. SSL improved reservoir classification in data-limited zones with self-training (71.0% accuracy) and neural networks (72.03%), outperforming Kernel Support Vector Machine (KSVM) (58.86%). This study demonstrates that SSL, combined with unsupervised clustering and supervised RF, overcomes the limitations of standalone ML approaches. By prioritizing resistivity-driven insights and scalable SSL workflows, the framework offers a cost-effective solution for bypassed hydrocarbon recovery in mature fields. Our hybrid methodology sets a precedent for applying semi-supervised techniques to complex reservoir systems globally, enhancing accuracy while reducing reliance on fully labeled datasets.