Geological Concept-Driven Intervention in Machine Learning for Foraminiferal Reservoir Property Modeling

Abstract Machine learning has increasingly been applied in geomodelling to extract spatial relationships from well and seismic data and to support reservoir characterization. This study evaluates a geology-guided machine-learning approach for porosity modeling in the DA Field, a foraminiferal carbonate gas reservoir in the Madura Strait PSC. The objective is to test whether seismic-derived attributes, well-log porosity, and rock-type constraints can be integrated in a transferable workflow that preserves geological realism while improving modeling turnaround time. The DA Field reservoir was characterized using rock types defined from facies distribution and hydrocarbon flow capacity. Seismic attributes, including acoustic impedance, amplitude, RMS amplitude, and instantaneous phase, were resampled into the structural grid and used as predictor variables. The modeling framework combines supervised learning with spatial prior information so that property estimation is not controlled by seismic attributes alone, but is also evaluated against geological concepts and blind-well validation. When the model relied mainly on seismic attribute inputs, the resulting porosity distribution generated 12–13% lower resource estimates than the existing conventional model. The discrepancy was traced to the acoustic impedance volume, which had been generated before the development drilling campaign and therefore did not fully represent the updated rock-type understanding. A subsequent scenario incorporated the rock-type model as a geological constraint rather than as a deterministic output template. This revised approach reduced the volumetric difference to 0.008% relative to the conventional model, shortened the modeling turnaround substantially, and retained a blind-well correlation of approximately 62–63%. The implementation of machine learning with a domain-guided geological intervention capability enables the integration of depositional constraints, represented here by the rock type model. This approach demonstrates that newly available data can be directly used as input for model updates without repeating full variogram and geostatistical modeling routines, significantly reducing computational turnaround time. This workflow can be adapted to other reservoirs where seismic attributes, well control, and conceptual geological models need to be reconciled.