Combining Prior Knowledge and Self-Organising Feature Maps for Logging Data Stratification

ABSTRACT

Well log stratification is crucial for geological interpretation, yet it faces challenges in balancing automation efficiency with geological interpretability. This study proposes a novel unsupervised method, PK-SOM (Prior Knowledge - Self Organising Maps), which integrates geological prior knowledge into Self-Organising Maps to overcome the “black box” limitations of data-driven models. The method employs Support Vector Data Description (SVDD) for anomaly detection, followed by a constrained SOM clustering process. Applied to borehole data from the Songliao Basin, PK-SOM achieved high stratigraphic consistency with expert interpretations, reducing boundary errors by ~5.6% compared to standard SOM. Furthermore, using PK-SOM stratified data as preprocessing significantly enhanced downstream lithology identification, improving model accuracy by over 11%. This demonstrates the method’s effectiveness in achieving automated, interpretable, and geologically reliable stratification without manual labelling.