K-Means Cluster Analysis

Key Takeaways

• Seismic facies classification with k-means identifies spatial patterns in waveform character that geologists interpret as depositional environments — in a 3D seismic volume, the seismic waveform at each location (the amplitude pattern through a time window around a horizon of interest) reflects the acoustic character of the rock and fluids in that location; k-means clustering of these waveforms groups locations with similar acoustic character together into seismic facies, which often correspond to distinct depositional environments (channel sands, levee deposits, deep marine shales, carbonate reef structures); the resulting facies map shows the lateral distribution of similar waveform patterns across the field, which geologists interpret in the context of regional stratigraphy and known depositional systems; k-means seismic facies analysis has become a routine step in prospect evaluation and reservoir characterization for structurally and stratigraphically complex plays.
• Electrofacies analysis uses k-means to integrate multiple well logs into rock type classifications — individual well logs (gamma ray, density, neutron, resistivity, sonic, photoelectric) each reflect specific aspects of formation character, but no single log uniquely identifies rock type across diverse lithological environments; k-means clustering in the multi-dimensional log space groups depth intervals with similar combined log responses into electrofacies that can be correlated with core-described rock types (lithofacies, reservoir quality categories, diagenetic facies) at cored wells and propagated to uncored wells and inter-well volumes through 3D geostatistical modeling; this approach captures textural and compositional information from the full multi-log measurement suite that manual interpretation of individual logs would miss or inconsistently classify between different interpreters.
• Feature selection and normalization are critical to meaningful k-means results — k-means clusters based on Euclidean distance in the feature space, so variables with large numerical ranges dominate the clustering over variables with small ranges even if the latter are geologically more significant; z-score normalization (subtracting the mean and dividing by the standard deviation of each variable) scales all inputs to comparable range before clustering; feature selection (choosing which logs or attributes to include) requires geological judgment — including irrelevant variables adds noise that obscures meaningful groupings, while excluding important variables misses information that would improve cluster differentiation; exploratory principal component analysis (PCA) before k-means can help identify which combinations of original variables capture most of the variance and should drive the classification.
• K-means has a centroid-based limitation that makes it poorly suited for elongated or non-convex cluster shapes — k-means finds spherically shaped clusters in the feature space because it minimizes variance around centroids; real geological data often has complex cluster shapes (curved, elongated, or hierarchical distributions) that k-means cannot capture without choosing a large k that overpartitions the data; alternative clustering algorithms including DBSCAN (density-based clustering that finds arbitrarily shaped clusters and labels outliers as noise), Gaussian mixture models (GMM, probabilistic clustering that allows ellipsoidal cluster shapes and soft cluster assignments), and hierarchical clustering (which builds a dendrogram showing the nested structure of cluster relationships) may be more appropriate for datasets with complex cluster geometry, though each has its own limitations and parameter choices.
• Production clustering with k-means identifies well performance archetypes for targeted optimization — in unconventional plays with many wells, k-means clustering of production profiles (IP30, IP90, decline rate, GOR, water cut evolution) groups wells into distinct performance archetypes that may reflect underlying geological, completion, or operational differences; wells in low-performance clusters are candidates for investigation of what differentiates them from high-performing cluster neighbors, and the diagnosis may reveal actionable completion design improvements, landing zone adjustments, or production management changes; time-lapse clustering of the same wells with new production data can reveal which wells are improving or degrading relative to their archetype, providing an early warning system for developing production issues.