A data-driven workflow using shallow learning for regional facies classification in carbonates

Regional facies interpretation supports basin-scale exploration by defining basin boundaries and identifying transitions in depositional settings that flag potential areas of interest. One conventional regional product for this purpose is the gross depositional environment (GDE) map, which synthesizes core descriptions, well logs, and seismic structure. In carbonate formations spanning tens of thousands of square kilometers, the GDE-based approach is constrained by three compounding factors: carbonate heterogeneity produces nonunique seismic responses, well control is too sparse to resolve the lateral length scales of facies variation, and manual interpretation is labor-intensive and interpreter-dependent. A data-driven, seismic data-only workflow was proposed to address these limitations through two stages: a genetic algorithm (GA) constructed laterally consistent horizons directly from the seismic data in a target-oriented and reproducible manner, and self-organizing maps (SOMs) clustered horizon-consistent seismic attributes, drawn from amplitude, waveform-character, and lateral-discontinuity families, into facies-style classes. The SOM result was benchmarked against the GDE map. On a large regional carbonate survey, GA-based horizon construction and SOM classification were completed in about 2 hours, automating the horizon-picking effort that would otherwise require days to weeks of manual work. Two SOM configurations demonstrated the contribution of attribute families: an interior-properties view (amplitude + waveform character) and an interior + boundary view (with the lateral-discontinuity family added). The resulting facies maps captured the major depositional trends present in the GDE and revealed regional patterns between control points, including a long-recognized out-inter shoal anomaly that was mapped in the GDE and independently resolved by the SOM facies classification, with well property contrasts inside the feature being consistent with both. The shallow learning methods used here provided an efficient, reproducible, and interpreter-consistent regional reconnaissance product, demonstrating practical effectiveness at field scale and complementing, rather than replacing, hard-data interpretation in regions of sparse control.