مقايسه عملكرد الگوريتم‌هاي يادگيري ماشين SVM و Gradient Boosting در شناسايي رخساره‌هاي سنگي با استفاده از داده‌هاي چاه‌پيمايي

Lithofacies identification based on well-log data plays a key role in hydrocarbon reservoir modeling an‎d in reducing uncertainty in exploration an‎d production decision-making. This study eva‎luates an‎d compares two machine-learning models Support Vector Machine (SVM) an‎d Gradient Boosting (GB) for lithofacies classification. The dataset consists of well-log measurements from the Hugoton Field in Kansas an‎d includes nine facies classes labeled based on core interpretation. A combination of petrophysical features (natural gamma ray, neutron porosity, density, resistivity, an‎d photoelectric factor) an‎d geological variables (depositional environment type an‎d relative position within the depositional cycle) was used as model input. The results indicate that the GB model, with a macro-average F1-score of 0.58, significantly outperforms the SVM model, which achieved a macro-average F1-score of 0.40, particularly in distinguishing facies with similar petrophysical characteristics. Feature-importance analysis shows that the geological variable representing depositional environment type is the most influential factor in both models, while GB demonstrates a stronger ability to leverage lithological information, especially natural gamma ray. Visual eva‎luation of predictions in test wells also confirms that this model produces outputs with greater continuity an‎d stronger geological consistency. However, the performance of both models is affected by two categories of limitations: (1) inherent data challenges, including the fuzzy an‎d continuous nature of facies, feature overlap, an‎d physical limitations of log measurements; an‎d (2) methodological constraints, particularly the assumption of sample independence an‎d the neglect of the sequential structure of well-log data. Accordingly, future work is recommended to employ sequence-based methods such as LSTM an‎d Transformer models, deep convolutional networks with signal-image inputs, an‎d intelligent resampling approaches that preserve data order, in order to improve accuracy while enhancing geological continuity. Overall, this research shows that acceptable results can be achieved even with non-sequential algorithms using petrophysical an‎d geological features; nevertheless, architectures that account for the natural structure of the data offer a promising direction for more accurate an‎d geologically meaningful facies classification.

حمزه صادقي سرخني

نادر فتحيان پور , اميرعباس جهانگرد