A Master Well, Multilayer Artificial Intelligence / Machine Learning (AI/ML) Framework for Scalable Porosity Permeability Prediction and Log Reconstruction in Carbonate Reservoirs

Abstract

Abstract Carbonate reservoirs exhibit complex petrophysical behavior arising from heterogeneity, uneven core coverage and variable log data quality. These challenges significantly limit the scalability and repeatability of conventional petrophysical workflows, particularly when extended to large well populations. To address these limitations, this study presents a Master Well centric artificial intelligence and machine learning (AI/ML) framework for scalable prediction of porosity and permeability and log reconstruction with a focus on improving model generalization, validation robustness and operational efficiency across the field. Conventional petrophysical machine learning is bound to the well in which it was calibrated. This study shifts the unit of training from the well to the field by constructing a Master Well, a single core calibrated reference population assembled by randomized sampling across nine wells. Two advances underpin the approach. Pore aspect ratio is estimated from the fractional lithology curves of the multimineral solution, so the elastic properties follow the local calcite and dolomite volumes and stay consistent with theoretical endmember values. The soft and stiff pore endmembers defined by Kumar and Han (2005) and Xu and Payne (2009) provide the envelope within which each sample is placed, and the aspect ratio contours between those limits separate the pore classes. Each pore system is normalized within its class to decompose total porosity into comparable fractions, which enter the models as engineered features alongside the pore classes carried through training and testing as categorical classifiers. The secondary porosity index and the pore aspect ratio emerge as the strongest permeability indicators of those tested, and both outrank porosity alone. Seventy two models were developed across multiple algorithm families, with the best realization of each retained for comparison. Held out testing returns R2 between 0.71 and 0.95, the wireline responses being reconstructed most accurately and the rock properties limited by genuine reservoir heterogeneity rather than by model capability. Because the reference population spans the full multilayer succession rather than a single interval, the models carry that layered behavior into every well to which they are applied. Deployment to a new well requires only environmental correction and normalization to the Master Well reference distributions, with no additional core. Accuracy is expected to hold within the calibration envelope of approximately 20 km lateral extent defined by the present dataset, beyond which predictions should be treated as indicative until confirmed against local core.