Integrating NMR and machine learning for pore-type driven rock classification in the heterogeneous Asmari carbonate reservoirs Authors: Maryam Veysi, Ali Kadkhodaie, Mehran Arian, Mohsen Aleali, Zahra Maleki, Pooria Kianoush Published: 2025 ABSTRACT The Asmari Formation’s complex heterogeneity presents fundamental challenges for reservoir characterization, where conventional lithology-based methods inadequately capture dynamic fluid behavior and pore-scale productivity controls. This study develops an integrated NMR-electrofacies framework combining Combinable Magnetic Resonance (CMR) logging with explainable machine learning to overcome such limitations. High-resolution CMR data were analyzed to quantify pore structure characteristics, fluid mobility parameters, and permeability indicators, with electrofacies classification performed using optimized machine learning algorithms incorporating SHAP interpretability. Three key innovations are presented: (1) a novel pore-connectivity index derived from T₂ distribution analysis, (2) dynamic permeability corrections accounting for clay-bound fluid effects, and (3) a lithology-independent rock typing system based on pore functionality. Results demonstrate that intervals with similar porosity (11% average) exhibit 3–5× variations in hydrocarbon transmissibility, confirming pore architecture’s dominant control. Notably, sandstone subunits representing just 15–20% of thickness contributed 60% of productive capacity through superior macroporosity (T₂ >100 ms). The machine learning framework achieved 92% prediction accuracy for flow units, with SHAP analysis identifying T₂ geometric mean (47% contribution) as the primary controlling parameter. These findings necessitate recharacterization of the Asmari Formation as a pore-connectivity governed system rather than lithology-dependent. Validation through comprehensive core-log integration (R²=0.89, RMSE = 0.32) and production data verification (84% PLT agreement) confirms method reliability. Practical applications include optimized well completion strategies and enhanced recovery in mature fields, while future adaptations encompass real-time implementation and CO₂ storage potential evaluation in analogous carbonate reservoirs. This work establishes a transformative approach for heterogeneous carbonate reservoir evaluation.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-32696-2. Introduction The pursuit of sustainable energy production requires a profound understanding of reservoir heterogeneity and rock fabric variability to maximize hydrocarbon recovery1,2. The Asmari Formation, among Iran’s most prolific carbonate reservoirs, exhibits complex diagenetically altered lithofacies, diverse depositional environments, and intricate porosity structures that challenge conventional characterization methods3–6. Traditional rock-typing approaches, while fundamental, often fail to capture dynamic pore geometry-fluid behavior interactions, particularly in carbonates with multi-scale heterogeneity7–9. The theoretical foundation of this research builds upon established pore system classifications10–12, while introducing critical adaptations for dynamic reservoir conditions. The methodology advances predictive modeling through NMR petrophysics and explainable machine learning integration, demonstrating significant improvements over conventional composition-based approaches13,14. These developments prove particularly valuable for mature fields where legacy data interpretations frequently overlook dynamic reservoir behavior and where precise flow unit identification is essential for recovery optimization. While conventional petrophysical logs provide valuable volumetric and textural insights, they often cannot resolve complete pore architectures governing flow capacity9,15, particularly in formations with complex wettability16. Nuclear Magnetic Resonance (NMR) logging offers superior microstructural characterization through direct quantification of pore throat distributions and fluid mobility via T₂ relaxation measurements. Recent advances in NMR-seismic integration have enhanced fluid typing capabilities, though carbonate-specific resolution remains limited17–20. NMR technology enables rock type differentiation based on pore system characteristics rather than bulk lithology, providing quantitative assessment of pore size distribution and permeability without lithologic assumptions21–23. This study develops a robust pore type-based framework for heterogeneous carbonate-clastic reservoirs, with specific application to the Asmari Formation. Primary objectives were systematically addressed through: Pore system quantification using NMR T₂ relaxation spectra to differentiate micro-, meso-, and macro-pore domains with core-calibrated accuracy (R² = 0.89 against helium porosity).Lithology-independent rock typing through NMR-electrofacies integration validated by machine learning classification (92% prediction accuracy for flow units).Framework validation via tri-modal calibration combining core data, NMR measurements, and SHAP-interpretable machine learning outputs (84% agreement with production logging). Pore system quantification using NMR T₂ relaxation spectra to differentiate micro-, meso-, and macro-pore domains with core-calibrated accuracy (R² = 0.89 against helium porosity). Pore system quantification using NMR T₂ relaxation spectra to differentiate micro-, meso-, and macro-pore domains with core-calibrated accuracy (R² = 0.89 against helium porosity). Lithology-independent rock typing through NMR-electrofacies integration validated by machine learning classification (92% prediction accuracy for flow units). Lithology-independent rock typing through NMR-electrofacies integration validated by machine learning classification (92% prediction accuracy for flow units). Framework validation via tri-modal calibration combining core data, NMR measurements, and SHAP-interpretable machine learning outputs (84% agreement with production logging). Framework validation via tri-modal calibration combining core data, NMR measurements, and SHAP-interpretable machine learning outputs (84% agreement with production logging). Key innovations establishing methodological advancement include: A dynamic pore-connectivity index (PCI) derived from T₂ distribution skewness analysis, providing a quantitative measure of flow potential that explains 73% of production variance in sandstone intervals despite comprising only 15–20% of reservoir thickness.Dynamic permeability corrections incorporating clay-bound fluid effects that reduce prediction errors by 37% in shaly zones (Vclay > 25%) compared to conventional static models.The first integration of explainable AI (SHAP) with NMR petrophysics for transparent electrofacies classification, with T₂ geometric mean identified as the dominant controlling parameter (47% contribution) through rigorous feature importance analysis. A dynamic pore-connectivity index (PCI) derived from T₂ distribution skewness analysis, providing a quantitative measure of flow potential that explains 73% of production variance in sandstone intervals despite comprising only 15–20% of reservoir thickness. A dynamic pore-connectivity index (PCI) derived from T₂ distribution skewness analysis, providing a quantitative measure of flow potential that explains 73% of production variance in sandstone intervals despite comprising only 15–20% of reservoir thickness. Dynamic permeability corrections incorporating clay-bound fluid effects that reduce prediction errors by 37% in shaly zones (Vclay > 25%) compared to conventional static models. Dynamic permeability corrections incorporating clay-bound fluid effects that reduce prediction errors by 37% in shaly zones (Vclay > 25%) compared to conventional static models. The first integration of explainable AI (SHAP) with NMR petrophysics for transparent electrofacies classification, with T₂ geometric mean identified as the dominant controlling parameter (47% contribution) through rigorous feature importance analysis. The first integration of explainable AI (SHAP) with NMR petrophysics for transparent electrofacies classification, with T₂ geometric mean identified as the dominant controlling parameter (47% contribution) through rigorous feature importance analysis. In mature hydrocarbon provinces such as the Dezful Embayment, where decades of production have altered pressure regimes and fluid distributions within reservoir systems, conventional petrophysical interpretations often depend on outdated static models that cannot account for evolving pore structures and wettability contrasts24–26. This limitation necessitates a paradigm shift toward dynamic rock typing approaches where reservoirs are evaluated not as homogeneous entities but as spatiotemporal continua of evolving petrophysical properties. The Asmari Formation, characterized by its multi-scale heterogeneity arising from cyclic shallow marine carbonate sedimentation interspersed with siliciclastic incursions and diagenetic overprints, serves as an ideal case study for implementing such innovative approaches. The fundamental challenge lies in quantifying subtle yet critically important variations in pore architecture across different facies - an endeavor for which NMR logging proves indispensable27,28. Unlike conventional porosity logs that provide only bulk averages, NMR technology enables discrimination of porosity into distinct size classes (micro-, meso-, and macro-pores), thereby revealing the true nature of fluid flow networks that govern both permeability and storage potential29,30. These capabilities become particularly valuable when evaluating mixed carbonate-clastic systems where traditional log interpretation methods struggle to distinguish between matrix effects and true reservoir quality indicators. By analyzing T₂ distribution signatures from NMR data, this study establishes a reservoir-aware classification of rock types that simultaneously captures depositional signatures, diagenetic imprints, and pore connectivity within a unified model. This model achieves geological realism through integration with stratigraphic stacking patterns, electrofacies analysis, and core-log calibrations - effectively transforming abstract log responses into geologically consistent, production-relevant insights. The workflow holds particular significance for reservoirs approaching secondary recovery phases, where accurate flow unit delineation and optimized stimulation strategies become increasingly critical for maintaining production efficiency. The geological and petrophysical complexity of the Asmari Formation has been extensively documented in regional studies. Gharechelou, et al31,32. demonstrated the superior hydrocarbon potential of sandstone-rich subunits within the predominantly carbonate sequence, while Honarmand and Amini33 established important relationships between flow units, facies distribution and diagenetic features34,35. Recent work by Saffari and Kianoush36 provided comprehensive petrophysical benchmarks for Zagros reservoirs, highlighting the need for improved pore-scale characterization methods. Recent technological advances in NMR applications37,38 and machine learning implementations39,40 have created new opportunities to overcome longstanding limitations in carbonate reservoir characterization. Notably, Varkouhi, et al41,42. investigated pore water typing in Japanese reservoirs using NMR and machine learning, demonstrating the critical role of fluid mobility in pore system classification. Their work highlighted the limitations of traditional lithology-based methods in capturing dynamic fluid behavior, strongly supporting the pore type-based approach developed in this study. Similarly, Eftekhari, et al43,44. applied electrofacies and rock typing in the Mansouri Oilfield (SW Iran), emphasizing the need for high-resolution pore-scale diagnostics in heterogeneous carbonate-clastic systems. Their findings revealed that conventional facies models often overlook sandstone-rich intervals with superior flow potential - a key gap addressed through NMR-derived pore architecture analysis in the current research. The conceptual innovation of this research lies in treating rock types as dynamic, pore-structured systems rather than static lithofacies labels. Methodologically, this represents the first integration of explainable AI (SHAP) with NMR petrophysics for transparent electrofacies classification, coupled with empirical demonstration of sandstone subunits as critical flow conduits—fundamentally challenging traditional carbonate-centric reservoir models. By integrating NMR-derived pore-scale insights with conventional logs, geological data, and core-calibrated benchmarks, a transformative framework is developed for identifying functional rock units directly tied to reservoir quality, zonation, and development strategy. Through application of digital log modeling, multi-scale data fusion, and sedimentological correlation, this study improves resolution of reservoir heterogeneity while enabling more predictive and tailored decision-making in field development. The practical importance of this research is underscored by its implications for reservoir simulation, enhanced recovery planning, and geostatistical modeling. Reliable prediction of production potential depends fundamentally on precise classification of rock types - particularly in aging fields where dynamic reservoir behavior may diverge from initial models. By applying this NMR-guided methodology to such a complex and economically significant field, the study contributes to both theoretical advancement and operational optimization in carbonate reservoir management. Geological setting Lithostratigraphic description The lithological framework presented in this study was established through integration of subsurface geological data, detailed petrographic analysis, and comprehensive well-log interpretation1,45,46. 1. Aghajari Formation (0–1554 m drilled depth; 1546 m thickness): This formation comprises a siliciclastic-dominated succession characterized by medium-grained, cherty sandstone and matrix-supported conglomerates47–49. Fig. 1a) Stratigraphy of Western (Lurestan), central and eastern Zagros from Triassic to Cambrian ages, b) Comparing Lower Triassic to Permian periods from Iran and Arab states in Dehram and Khuff Groups50–52. This figure was generated using AutoCAD 2022 (https://www.autodesk.com/products/autocad). a) Stratigraphy of Western (Lurestan), central and eastern Zagros from Triassic to Cambrian ages, b) Comparing Lower Triassic to Permian periods from Iran and Arab states in Dehram and Khuff Groups50–52. This figure was generated using AutoCAD 2022 (https://www.autodesk.com/products/autocad). 2. Gachsaran Formation (1554–3151 m; 1596 m thickness): Marked by the abrupt appearance of thick anhydrite intervals, the Gachsaran Formation displays strong evaporitic cyclicity interbedded with marl and halite. Stratigraphically, this unit is subdivided into six informal members: Units 3 & 6 (1554–1834 m; 321 m): Dominated by alternating red and gray marls and discrete anhydrite sheets, reflecting shallow marine to sabkha depositional settings with intermittent subaerial exposure. Unit 5 (1834–2134 m; 261 m): Exhibits increased gray marl and anhydrite content alongside a reduction in red marl, denoting a shift toward more restricted and saline environmental conditions. Unit 4 (2134–2824 m; 691 m): Initiated by the first occurrence of massive halite, this section presents rhythmic alternations of thick rock salt, anhydrite, and gray marl layers, consistent with lagoonal-evaporitic cycles. Unit 3 (2824–3134 m; 251 m): Characterized by persistent anhydrite-dominated cycles interspersed with gray marl and thin halite seams, implying sustained hypersaline conditions. Unit 2 (3134–3199 m; 65 m): Composed predominantly of massive halite with thin interbeds of gray marl and anhydrite, representing the culmination of evaporitic facies before transition to the overlying formations. Caprock Interval (3199–3151 m; 51 m): Serving as a seal to the underlying hydrocarbon-bearing zones, this interval consists of interlaminated anhydrite, gray and red marl, micritic limestone, and finely crystalline dolomite53–55. Units 3 & 6 (1554–1834 m; 321 m): Dominated by alternating red and gray marls and discrete anhydrite sheets, reflecting shallow marine to sabkha depositional settings with intermittent subaerial exposure. Units 3 & 6 (1554–1834 m; 321 m): Dominated by alternating red and gray marls and discrete anhydrite sheets, reflecting shallow marine to sabkha depositional settings with intermittent subaerial exposure. Unit 5 (1834–2134 m; 261 m): Exhibits increased gray marl and anhydrite content alongside a reduction in red marl, denoting a shift toward more restricted and saline environmental conditions. Unit 5 (1834–2134 m; 261 m): Exhibits increased gray marl and anhydrite content alongside a reduction in red marl, denoting a shift toward more restricted and saline environmental conditions. Unit 4 (2134–2824 m; 691 m): Initiated by the first occurrence of massive halite, this section presents rhythmic alternations of thick rock salt, anhydrite, and gray marl layers, consistent with lagoonal-evaporitic cycles. Unit 4 (2134–2824 m; 691 m): Initiated by the first occurrence of massive halite, this section presents rhythmic alternations of thick rock salt, anhydrite, and gray marl layers, consistent with lagoonal-evaporitic cycles. Unit 3 (2824–3134 m; 251 m): Characterized by persistent anhydrite-dominated cycles interspersed with gray marl and thin halite seams, implying sustained hypersaline conditions. Unit 3 (2824–3134 m; 251 m): Characterized by persistent anhydrite-dominated cycles interspersed with gray marl and thin halite seams, implying sustained hypersaline conditions. Unit 2 (3134–3199 m; 65 m): Composed predominantly of massive halite with thin interbeds of gray marl and anhydrite, representing the culmination of evaporitic facies before transition to the overlying formations. Unit 2 (3134–3199 m; 65 m): Composed predominantly of massive halite with thin interbeds of gray marl and anhydrite, representing the culmination of evaporitic facies before transition to the overlying formations. Caprock Interval (3199–3151 m; 51 m): Serving as a seal to the underlying hydrocarbon-bearing zones, this interval consists of interlaminated anhydrite, gray and red marl, micritic limestone, and finely crystalline dolomite53–55. Caprock Interval (3199–3151 m; 51 m): Serving as a seal to the underlying hydrocarbon-bearing zones, this interval consists of interlaminated anhydrite, gray and red marl, micritic limestone, and finely crystalline dolomite53–55. Asmari Formation (3151–3443 m; 293 m thickness): This key reservoir unit comprises a rhythmic sequence of limestone, dolomitic limestone, and crystalline dolomite intercalated with sporadic sandstone and shale beds3,54.Pabdeh Formation (3443–3523 m; 81 m thickness): The lowermost unit penetrated in this well consists predominantly of dark gray shale interbedded with marly limestone, forming the stratigraphic base and regional seal below the productive Asmari reservoir56,57. Asmari Formation (3151–3443 m; 293 m thickness): This key reservoir unit comprises a rhythmic sequence of limestone, dolomitic limestone, and crystalline dolomite intercalated with sporadic sandstone and shale beds3,54. Asmari Formation (3151–3443 m; 293 m thickness): This key reservoir unit comprises a rhythmic sequence of limestone, dolomitic limestone, and crystalline dolomite intercalated with sporadic sandstone and shale beds3,54. Pabdeh Formation (3443–3523 m; 81 m thickness): The lowermost unit penetrated in this well consists predominantly of dark gray shale interbedded with marly limestone, forming the stratigraphic base and regional seal below the productive Asmari reservoir56,57. Pabdeh Formation (3443–3523 m; 81 m thickness): The lowermost unit penetrated in this well consists predominantly of dark gray shale interbedded with marly limestone, forming the stratigraphic base and regional seal below the productive Asmari reservoir56,57. Reservoir formation of interest: the Asmari carbonate system The Asmari Formation is the youngest and most prolific carbonate reservoir in the Zagros Fold and Thrust Belt, serving as the primary hydrocarbon-bearing unit in numerous onshore oilfields of southwestern Iran37,58,59. Lithologically, the formation consists of platform carbonates (biomicrites, dolomitized limestones) with minor argillaceous layers, shale seams, and sandy carbonates55,60,61. Regionally, the Asmari exhibits considerable heterogeneity, including the Ahvaz Sandstone Member in the southwest and the Kalhur Evaporite Member in the Lurestan region62,63. A basal anhydrite unit further complicates the stratigraphy in fields like Masjed Soleyman, Haftgel, and Pars64,65. Chrono-stratigraphically, the Asmari spans the late Oligocene (Rupelian-Chattian) to early Miocene (Aquitanian-Burdigalian), representing a major transgressive-regressive cycle driven by regional tectonics and eustatic sea-level changes48,53,56. These complexities result in highly variable pore fabrics (interparticle, moldic, fracture) that are strongly influenced by diagenetic processes such as cementation, dolomitization, and pressure solution66–68. As a classic example of a multiscale, diagenetically modified carbonate reservoir32,33,69, the Asmari Formation is an ideal subject for pore-scale studies using NMR logging to link depositional features, diagenetic overprints, and dynamic fluid flow behavior. Study area The study area located approximately 52 km northeast of Dehluran in Ilam Province, occupies a strategic position within the central sector of Iran’s prolific Dezful Embayment. The field was discovered in 1967 following exploratory drilling initiatives that targeted the Asmari Formation8,31. To date, twelve wells have been drilled across the structure, with one additional well currently under drilling to further delineate the reservoir extent and productivity. The field demonstrates a production capacity exceeding 81,000 barrels of oil per day, with a current operational output of over 66,000 barrels per day. Commercial production commenced in 1977, and since then, crude oil extracted from the reservoir has been processed. Following separation and treatment, the crude is transported via a dedicated 153-kilometer pipeline to the Ahvaz-3 Production Unit for export or downstream utilization12,56. Ongoing expansion efforts, including the proposed drilling of eight additional wells by 2025, are aimed at sustaining or potentially increasing production to nominal capacity levels. Development strategies were informed by ongoing geomechanical studies, reservoir simulation, and enhanced data integration—of which the present study is a significant component. The geospatial positioning of the field and its neighboring reservoirs is illustrated in Fig. 1, while the structural framework and terrain attributes derived from satellite imagery are detailed in Fig. 2. The study area provides an ideal laboratory for advanced reservoir studies owing to its accessible data infrastructure, active production history, and well-preserved core and log records. These attributes render it especially well-suited for pore-scale reservoir characterization using NMR log data, as pursued in the current research50,70. Fig. 2Geospatial context of a carbonate giant: Location of the study area within Iran’s subsurface energy framework30,59. This figure was generated using CorelDRAW Graphics Suite 2021 (https://www.coreldraw.com/). Geospatial context of a carbonate giant: Location of the study area within Iran’s subsurface energy framework30,59. This figure was generated using CorelDRAW Graphics Suite 2021 (https://www.coreldraw.com/). Methodology Data acquisition and preprocessing The empirical foundation incorporated multiple data streams71–74: conventional wireline logs (gamma ray, density, neutron, sonic, and resistivity) provided baseline petrophysical parameters; NMR logs enabled T₂ relaxation analysis for pore-scale characterization; and core measurements from 32 plugs served as calibration benchmarks. Geological and stratigraphic context was established through well reports and regional studies50,75,76. All data processing and analysis were conducted using the Geolog software environment, beginning with essential preprocessing steps including depth alignment via stratigraphic markers, environmental corrections for borehole effects, and rigorous quality control excluding intervals with signal-to-noise ratios below 3:1 or caliper readings exceeding 12 inches. Feature selection identified 17 predictive parameters through mutual information scoring (> 0.3), including NMR-derived metrics (T₂ geometric mean, bound/free fluid volumes, pore connectivity index), conventional log measurements, and core-calibrated permeability values, all normalized using z-score standardization. The complete workflow, illustrated in Fig. 3, progresses systematically from raw data acquisition through to reservoir quality prediction. Particular attention was given to the NMR data processing chain, where multiple inversion algorithms were evaluated to optimize T₂ distribution reconstruction, ultimately selecting singular value decomposition (SVD) for its superior performance in heterogeneous carbonate intervals. The complete initial feature set considered for machine learning is provided in Supplementary Table 5. Fig. 3Structured workflow for multi-scale rock typing based on NMR-driven pore geometry integration in the Asmari reservoir. Structured workflow for multi-scale rock typing based on NMR-driven pore geometry integration in the Asmari reservoir. NMR data processing and pore type characterization Schlumberger’s Combinable Magnetic Resonance (CMR) tool was used for high-resolution pore system characterization77–79. Three inversion algorithms were evaluated for T₂ distribution calculation from CPMG echo decay data. The L1L2 regularization approach minimized the objective function as Eq. (1)80–82:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\min _f}\left\{ {{\mathbf{d}} - {\mathbf{K}}f_2^2 + {\lambda _1}f{_1} + {\lambda _2}f_2^2} \right\}$$\end{document} where d represents the vector of echo data, K is the kernel matrix, and f denotes the T₂ distribution vector, and λ1, λ2 are regularization parameters controlling sparsity and smoothness constraints, respectively. The Zero-Shift Deconvolution (ZSD) method preserved spectral peak fidelity through non-parametric inversion, while Singular Value Decomposition (SVD) with 15-value truncation provided optimal stability. Comparative validation established SVD’s superior performance with mean squared error (MSE) of 0.018 and strong correlation (R² = 0.92) against laboratory NMR measurements28,83. Key petrophysical parameters derived from NMR included total porosity (TCMR), bound fluid volume (BFV) using lithology-specific T₂ cutoffs (33 ms for sandstones, 100 ms for carbonates as initial defaults, subsequently calibrated to core data)37,84, and pore size classification: Microporosity (T₂ < 10 ms).Mesoporosity (10 ≤ T₂ ≤ 100 ms).Macroporosity (T₂ > 100 ms)34,84,85. Microporosity (T₂ < 10 ms). Microporosity (T₂ < 10 ms). Mesoporosity (10 ≤ T₂ ≤ 100 ms). Mesoporosity (10 ≤ T₂ ≤ 100 ms). Macroporosity (T₂ > 100 ms)34,84,85. Macroporosity (T₂ > 100 ms)34,84,85. The tool’s 6-inch vertical resolution and side-looking configuration minimized borehole fluid interference while maintaining sensitivity to thin-bed heterogeneities38,86. Calibration procedures incorporated deionized water standards and in situ mud filtrate T₂ measurements for enhanced environmental corrections61,87,88. The selected SVD methodology proved particularly effective for characterizing the Asmari Formation’s complex pore systems87,89, where rapid vertical variations in connectivity and pore geometry require robust inversion stability40,90. Final T₂ spectra were validated against independent petrophysical measurements to ensure geological consistency throughout the reservoir column42,70,91. (Detailed are provided in Supplementary Section D). The Total CMR Porosity (TCMR) log captured the complete pore size spectrum, from clay-bound water in microporous zones (< 10 ms T₂) to free fluids in macropores (> 100 ms T₂). TCMR-derived porosity values ranged from < 5% in tight zones to 25% in high-quality intervals, averaging 11% - consistent with regional analogues like the Bangestan and Dalan formations. The three volumetric formulations supporting porosity estimation, define porosity as Eq. (2) to (4)70,92:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\phi _t} = \frac{{{V_p}}}{{{V_b}}} (Total\, porosity\, definition)$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\phi _t} = 1 - \frac{{{V_g}}}{{{V_b}}}(Porosity\, from\, grain\, volume)$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\phi _t} = \frac{{{V_p}}}{{{V_p} + {V_g}}}(Porosity\, from\, pore\, and\, grain\, volumes)$$\end{document} where φt is total porosity, Vp is pore volume, Vb is bulk rock volume, and Vg is grain volume. This methodology enabled reliable discrimination between effective and bound porosity components, particularly valuable for identifying clay-bound water in shale interbeds versus connected macroporous networks81,93. Conventional log analysis for lithological context To provide a lithological framework for NMR pore-type integration, we employed neutron-density and M-N crossplot analyses94–96. Matrix lines for standard lithologies were calibrated and overlaid on environmentally corrected crossplots (Fig. 5). The neutron-density analysis revealed the formation’s inherent heterogeneity, with data clusters indicating carbonate-dominant intervals interspersed with sandstone-rich zones and localized anhydrite. The M–N crossplot technique was employed to address the mineralogical complexity of the Asmari Formation, integrating sonic (Δt), neutron (Φ N), and bulk density (ρ b) measurements into a ternary classification system. This approach normalizes fluid and porosity effects through dimensionless M and N parameters (Eqs. (5) and (6))40,97, where:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M=\frac{{\Delta {t_f} - \Delta t}}{{{\rho _b} - {\rho _f}}} \times 0.01$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N=\frac{{{\phi _{N,f}} - {\phi _N}}}{{{\rho _b} - {\rho _f}}}$$\end{document} where Δt is sonic travel time, Δf is fluid travel time, ρb is bulk density, ρf is fluid density, φN is neutron porosity and φN, f is neutron porosity of the fluid. For the Asmari reservoir, M–N coordinates were interpreted synergistically with NMR-derived T₂ distributions, advancing beyond static log interpretation by incorporating dynamic pore-scale fluid behavior. Diagnostic outliers and ambiguous trends were resolved by integrating spectral gamma ray and elemental capture spectroscopy logs39,52. This multi-log reconciliation proved particularly valuable in transitional intervals. The integration of these conventional discrimination methods with NMR pore architecture analysis provided a critical mineralogical foundation for the pore type-based rock typing framework. Detailed lithological discrimination methodology is provided in Supplementary Information, including Supplementary Table 1 for M-N matrix parameters. Machine learning methodology The integration of machine learning (ML) techniques - including Random Forest (RF), Support Vector Machines (SVM), and SHAP (SHapley Additive exPlanations) - provides a transformative approach for refining the pore type-based rock typing framework. These methods analyze NMR-derived pore system attributes, electrofacies classification, and petrophysical parameters, enabling data-driven validation of the hybrid NMR-electrofacies model7,76,93,98. The machine learning framework follows a structured workflow (Fig. 4) that integrates multi-scale data sources and maintains geological consistency throughout the classification process. This systematic approach ensures transparent model development from feature selection to interpretable prediction output. Fig. 4Integrated machine learning workflow for pore-type classification and electrofacies prediction. The framework demonstrates the sequential processing stages from multi-modal data input (NMR, conventional logs, core data) through feature engineering, model training with Random Forest and Support Vector Machines, to interpretable output generation using SHAP analysis. The workflow emphasizes the cyclic validation process integrating geological consistency checks and production data verification. Integrated machine learning workflow for pore-type classification and electrofacies prediction. The framework demonstrates the sequential processing stages from multi-modal data input (NMR, conventional logs, core data) through feature engineering, model training with Random Forest and Support Vector Machines, to interpretable output generation using SHAP analysis. The workflow emphasizes the cyclic validation process integrating geological consistency checks and production data verification. The machine learning workflow initially considered seventeen input features spanning NMR-derived parameters, conventional well logs, and mineralogy data (Table 1). Feature selection was performed using mutual information (MI) scoring to identify the most predictive features for pore type classification. Features with MI scores greater than 0.3 (n = 9) were retained for final modeling, ensuring that only the most informative parameters were used in subsequent ML algorithms. Random forest (RF) implementation The RF algorithm was implemented with 500 decision trees to classify pore types into three categories: macroporous (T2 > 100ms), mesoporous (10ms ≤ T2 ≤ 100ms), and microporous (T2 < 10ms). The Gini impurity criterion was used for node splitting, with maximum tree depth limited to 15 levels to prevent overfitting. The feature importance metric (FI) was calculated as Eq. (7)7,39:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{F}}{{\mathrm{I}}_i}=\frac{1}{{{N_{{\mathrm{trees}}}}}}\sum\limits_{T} {\sum\limits_{{n \in T}} \Delta } {\mathrm{Gin}}{{\mathrm{i}}_i}(n)$$\end{document} where FIi is the importance of feature i, Ntrees is the number of trees, T represents each tree, n is a node in the tree, and ΔGinii is the Gini impurity reduction at node n due to feature i. The model was trained on 70% of the dataset with 10-fold cross-validation. Support vector machine (SVM) configuration For electrofacies classification, a radial basis function (RBF) kernel SVM was employed. The decision function for n training samples (x i, y i) was formulated as Eq. (8)14:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f(x)={\mathrm{sign}}\left( {\sum\limits_{{i=1}}^{n} {{\alpha _i}} {y_i}K({x_i},x)+b} \right)$$\end{document} where α i are Lagrange multipliers, y i ∈ {−1,1} are class labels, K(xi, x) is the radial basis function (RBF) kernel, and b is the bias term. SHAP interpretation framework SHAP values were computed to quantify feature contributions to model predictions. For each instance x, the explanation model g(z’) was defined as Eq. (9)13,99:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g(z^{\prime})={\phi _0}+\sum\limits_{{j=1}}^{M} {{\phi _j}} {z^{\prime}_j}$$\end{document} where g is the explanation model, z’ ∈ {0,1} M is a simplified feature vector, M is the number of features, φ0 is the base value, and φj is the SHAP value for feature j. Pore type classification was performed using Random Forest (500 trees, max depth = 15) applied to the nine selected features from T₂ distributions and conventional logs, while electrofacies prediction employed Support Vector Machines with RBF kernel (γ = 0.1, C = 10) using the same optimized feature set. Feature importance was quantified through KernelSHAP analysis using 10,000 background samples. Table 1Initial feature set for machine learning model development.Feature TypeParameterMutual Information ScoreSelectedNMR-DerivedT₂ Geometric Mean (T₂LM)0.45 Yes Bound Fluid Volume (BFV)0.41 Yes Free Fluid Index (FFI)0.38 Yes T₂ Peak Amplitude0.35NoT₂ Variance0.32NoConventional LogsGamma Ray (GR)0.40 Yes Bulk Density (RHOB)0.39 Yes Neutron Porosity (NPHI)0.37 Yes Sonic Travel Time (DT)0.34 Yes Shallow Resistivity (RXO)0.31NoMineralogy (ECS)Calcite Volume (%)0.36 Yes Dolomite Volume (%)0.33 Yes Clay Volume (VCL)0.30 Yes Quartz Volume (%)0.29NoAnhydrite Volume (%)0.28NoCore-CalibratedHelium Porosity (PHIT)0.42No*Air Permeability (K)0.44No**Core data used for validation only, not as input features to prevent data leakage. Table 1Initial feature set for machine learning model development.Feature TypeParameterMutual Information ScoreSelectedNMR-DerivedT₂ Geometric Mean (T₂LM)0.45 Yes Bound Fluid Volume (BFV)0.41 Yes Free Fluid Index (FFI)0.38 Yes T₂ Peak Amplitude0.35NoT₂ Variance0.32NoConventional LogsGamma Ray (GR)0.40 Yes Bulk Density (RHOB)0.39 Yes Neutron Porosity (NPHI)0.37 Yes Sonic Travel Time (DT)0.34 Yes Shallow Resistivity (RXO)0.31NoMineralogy (ECS)Calcite Volume (%)0.36 Yes Dolomite Volume (%)0.33 Yes Clay Volume (VCL)0.30 Yes Quartz Volume (%)0.29NoAnhydrite Volume (%)0.28NoCore-CalibratedHelium Porosity (PHIT)0.42No*Air Permeability (K)0.44No**Core data used for validation only, not as input features to prevent data leakage. Initial feature set for machine learning model development. *Core data used for validation only, not as input features to prevent data leakage. Performance validation Model accuracy was assessed through three key metrics: classification accuracy (CA) for pore types achieved 0.91 ± 0.03, the coefficient of determination (R²) for permeability prediction reached 0.86 ± 0.04, and the area under the ROC curve (AUC) for electrofacies identification scored 0.93 ± 0.02, with all values derived from validation using 30% of the dataset. These results confirm robust performance across all predictive tasks. The methodology integrates NMR log processing, core data calibration, and machine learning to establish a comprehensive pore type-based rock typing framework. Results Lithological discrimination The generated crossplot in Fig. 5 reveals that the investigated reservoir interval comprises interbedded carbonates (limestone and dolomite), minor clastic sandstones, and frequent anhydrite occurrences. Notably, a concentration of points trending toward the lower-right quadrant of the plot—accompanied by elevated gamma-ray signals—strongly suggests clay-rich (shale) intervals, a conclusion reinforced by radiometric log overlays on the color-coded crossplot. These contours intersect lithological trends at varying angles, revealing maximum porosity values approaching 30% in dissolution-enhanced carbonate facies of the Asmari Formation—consistent with secondary porosity development. The neutron-density crossplot has limitations in gas-bearing zones and where borehole conditions are poor. In these scenarios, NMR-derived porosity and T₂ distributions provide essential verification. When integrated with NMR analysis and core validation, the neutron-density crossplot becomes a critical component of the multi-log rock typing framework for complex carbonate reservoirs. Fig. 5Neutron-density crossplot interpretation showing lithological trends and porosity distribution in the Asmari Reservoir. The plot displays the relationship between corrected density (RHOB_COR) and neutron porosity (NPHI_COR) across the entire logged interval. Matrix lines delineate lithological trends (limestone, dolomite, sandstone, and anhydrite), while gamma-ray-derived color shading enhances identification of shale-rich intervals and evaporite deposits. Data point dispersion reveals a heterogeneous lithological composition dominated by carbonates with intercalated siliciclastics and evaporites, characteristic of the Asmari Formation’s complex depositional and diagenetic history. Neutron-density crossplot interpretation showing lithological trends and porosity distribution in the Asmari Reservoir. The plot displays the relationship between corrected density (RHOB_COR) and neutron porosity (NPHI_COR) across the entire logged interval. Matrix lines delineate lithological trends (limestone, dolomite, sandstone, and anhydrite), while gamma-ray-derived color shading enhances identification of shale-rich intervals and evaporite deposits. Data point dispersion reveals a heterogeneous lithological composition dominated by carbonates with intercalated siliciclastics and evaporites, characteristic of the Asmari Formation’s complex depositional and diagenetic history. Lithological pattern recognition based on MLITH–NLITH crossplot analysis The MLITH–NLITH crossplot was employed to enhance lithological differentiation in the Asmari Formation, providing an integrated visualization of mineralogical variations derived from log-based lithology estimates. Figure 6 demonstrates a dispersed data distribution pattern, reflecting the formation’s heterogeneous composition of limestone (calcite), dolomite, sandstone (quartz-rich), and evaporites (anhydrite/gypsum). Gamma-ray intensity values, represented by a blue-to-red color gradient on the crossplot, facilitate identification of clay-rich intervals associated with illite and kaolinite in laminated or argillaceous zones. The MLITH–NLITH analysis advances upon basic crossplot techniques by: Resolving subtle mineralogical mixtures.Supporting precise electrofacies boundary definition.Enabling lithological subdivision by pore geometry and saturation state when combined with NMR T₂ data (Table 2). Resolving subtle mineralogical mixtures. Resolving subtle mineralogical mixtures. Supporting precise electrofacies boundary definition. Supporting precise electrofacies boundary definition. Enabling lithological subdivision by pore geometry and saturation state when combined with NMR T₂ data (Table 2). Enabling lithological subdivision by pore geometry and saturation state when combined with NMR T₂ data (Table 2). Key lithological distributions revealed by the analysis include: Carbonate dominance (65–70% of interval) with limestone-dolomite mixing trends.Sandstone subunits (15–20%) clustered along quartz matrix line.Evaporite/anhydrite nodules (8–12%) in Gachsaran contact zones.Shale-rich intervals (Vclay > 25%) showing gamma-ray correlation (R²=0.78). Carbonate dominance (65–70% of interval) with limestone-dolomite mixing trends. Carbonate dominance (65–70% of interval) with limestone-dolomite mixing trends. Sandstone subunits (15–20%) clustered along quartz matrix line. Sandstone subunits (15–20%) clustered along quartz matrix line. Evaporite/anhydrite nodules (8–12%) in Gachsaran contact zones. Evaporite/anhydrite nodules (8–12%) in Gachsaran contact zones. Shale-rich intervals (Vclay > 25%) showing gamma-ray correlation (R²=0.78). Shale-rich intervals (Vclay > 25%) showing gamma-ray correlation (R²=0.78). Table 2Lithology distribution from MLITH-NLITH analysis.LithologyVolume (%)Avg. Φ (%)Avg. K (mD)Limestone42.311.28.7Dolomite27.513.512.4Sandstone16.818.654.2Shale13.44.10.3 Table 2Lithology distribution from MLITH-NLITH analysis.LithologyVolume (%)Avg. Φ (%)Avg. K (mD)Limestone42.311.28.7Dolomite27.513.512.4Sandstone16.818.654.2Shale13.44.10.3 Lithology distribution from MLITH-NLITH analysis. Fig. 6MLITH/NLITH crossplot visualization showing mineralogical dispersion and shale identification in the Asmari Formation. Color intensity represents gamma-ray response (GR_COR), highlighting radioactive shale-rich zones. The data distribution demonstrates heterogeneous lithological composition critical for pore system characterization and electrofacies classification. MLITH/NLITH crossplot visualization showing mineralogical dispersion and shale identification in the Asmari Formation. Color intensity represents gamma-ray response (GR_COR), highlighting radioactive shale-rich zones. The data distribution demonstrates heterogeneous lithological composition critical for pore system characterization and electrofacies classification. This crossplot technique transforms static lithology predictions into dynamic classifications linked to pore structure behavior and reservoir quality evolution—a fundamental component of the pore type-based workflow. Pore system characterization from NMR analysis In this study, T₂ data were acquired throughout the reservoir interval of the studied oilfield. Rigorous quality control protocols were applied to the raw CMR data prior to interpretation. These steps ensured that the derived T₂ distributions accurately reflected formation-level porosity and fluid partitioning. The Bound Fluid Volume (BFV) and its volumetric spectrum—expressed as the Spectral Bound Fluid Volume (SBFV)—were derived from CMR T₂ distribution analysis across the depth profile. These parameters were computed using appropriate T₂ cutoff values calibrated to local core data and mineralogy. The results revealed average BFV and SBFV values of 3.9% and 2.2%, respectively, across the Asmari interval, highlighting zones of low-permeability facies, particularly in fine-grained dolomitic units and shale-dominated laminations. Figure 7 presents the processed CMR log data, illustrating multiple tracks of key reservoir parameters, including total and effective porosity, free and bound fluid volumes, and CMR-derived permeability estimates. The spatial variation of T₂ spectra along depth intervals provides insights into the vertical architecture of the pore system, revealing zones of micro-, meso-, and macro-porosity that directly inform the rock typing workflow. The processed CMR data delineated three distinct pore architecture regimes across the Asmari Formation (Fig. 8). Microporous zones (T₂ < 10 ms), comprising 18–23% of total porosity, were predominantly associated with dense micritic intervals, exhibiting bound fluid volumes (BFV) exceeding 85% of total porosity. Mesoporous zones (10 ≤ T₂ ≤ 100 ms) dominated the productive intervals, accounting for 62% of hydrocarbon-bearing zones and showing strong correlation with dolomitized units (R² = 0.79), with an average permeability of 8.7 mD (range: 1.2–24.3 mD). Macroporous zones (T₂ > 100 ms) were concentrated in sandstone subunits (Fig. 7), representing only 15% of total porosity but contributing 73% of total flow capacity, consistently demonstrating pore connectivity indices (PCI) greater than 1.8. Core-log integration confirmed these NMR-derived classifications, with vuggy macroporosity at 3564 m showing excellent agreement between NMR predictions and core photographs (Table 3; Supplementary Tables 2 and 3). The CMR logging data provided pore-scale resolution through T₂ distribution analysis: Bimodal distributions in sandstones (T₂ peaks at 12 ± 3ms and 142 ± 18ms).Unimodal distributions in carbonates (single peak at 8 ± 2ms).Microporosity (T₂<0ms) dominated shale-rich zones (BFV > 85%). Bimodal distributions in sandstones (T₂ peaks at 12 ± 3ms and 142 ± 18ms). Bimodal distributions in sandstones (T₂ peaks at 12 ± 3ms and 142 ± 18ms). Unimodal distributions in carbonates (single peak at 8 ± 2ms). Unimodal distributions in carbonates (single peak at 8 ± 2ms). Microporosity (T₂<0ms) dominated shale-rich zones (BFV > 85%). Microporosity (T₂<0ms) dominated shale-rich zones (BFV > 85%). The pore connectivity index (PCI) was derived from T₂ skewness using Eq. (10)11,100:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{PCI}}=\frac{{{\mu _3}}}{{{\sigma ^3}}}=\frac{{\frac{1}{N}\sum\limits_{{i=1}}^{N} {{{({T_{2,i}} - \overline {{{T_2}}} )}^3}} }}{{{{\left( {\sqrt {\frac{1}{N}\sum\limits_{{i=1}}^{N} {{{({T_{2,i}} - \overline {{{T_2}}} )}^2}} } } \right)}^3}}}$$\end{document} where µ3 is the third moment about the mean (skewness), σ is the standard deviation of the T₂ distribution, T2,i is the i-th T₂ value, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline {{{T_2}}}$$\end{document}is the mean T2, and N is the number of points in the distribution. Table 3Pore-type distribution in the Asmari Formation, NMR-derived pore-type classification. Macroporous zones (sandstone-rich) dominate flow capacity despite lower porosity.Pore TypeT₂ Range (ms)Porosity (%)Permeability (mD)BFV (%)Microporous< 1018–230.3–1.2> 85Mesoporous10–10035–421.2–24.315–40Macroporous> 10012–1524.3–54.6< 12 Table 3Pore-type distribution in the Asmari Formation, NMR-derived pore-type classification. Macroporous zones (sandstone-rich) dominate flow capacity despite lower porosity.Pore TypeT₂ Range (ms)Porosity (%)Permeability (mD)BFV (%)Microporous< 1018–230.3–1.2> 85Mesoporous10–10035–421.2–24.315–40Macroporous> 10012–1524.3–54.6< 12 Pore-type distribution in the Asmari Formation, NMR-derived pore-type classification. Macroporous zones (sandstone-rich) dominate flow capacity despite lower porosity. As demonstrated in Fig. 7, the graphic layout of CMR signal processing tracks clearly distinguishes between clean carbonate intervals and shaly sections. This distinction underscores the importance of preprocessing in ensuring that relaxation signatures reflect intrinsic pore structures. Fig. 7Integrated CMR log presentation showing nine tracks of advanced NMR-derived parameters across the Asmari Reservoir, including: (a) T₂ distribution heatmap, (b) bound vs. free fluid partitioning, (c) permeability profile, and (d) lithology-specific porosity calibration. Integrated CMR log presentation showing nine tracks of advanced NMR-derived parameters across the Asmari Reservoir, including: (a) T₂ distribution heatmap, (b) bound vs. free fluid partitioning, (c) permeability profile, and (d) lithology-specific porosity calibration. Fig. 8Depth-registered T₂ distribution analysis contrasting bimodal pore systems in sandstone intervals (3540–3560 m) with unimodal microporous signatures in tight carbonates (3730–3770 m), demonstrating the methodology’s resolution of depositional and diagenetic pore architecture. Depth-registered T₂ distribution analysis contrasting bimodal pore systems in sandstone intervals (3540–3560 m) with unimodal microporous signatures in tight carbonates (3730–3770 m), demonstrating the methodology’s resolution of depositional and diagenetic pore architecture. Inversion of CMR decay data and T₂ distribution reconstruction Once preconditioning was completed, the processed CPMG sequences were inverted to extract the T₂ relaxation time distribution—a direct proxy for pore size classification and fluid mobility analysis. This inversion transformed the decaying spin-echo signal into a histogram of relaxation times, effectively revealing the internal architecture of the pore system. To ensure robustness and validate interpretive reliability, three distinct inversion algorithms were evaluated: L1L2 regularized inversion, combining sparsity and smoothness constraints,Zeroth- and second-order deconvolution (ZSD), which prioritizes shape fidelity over smoothness,Singular Value Decomposition (SVD), a matrix-based method offering computational stability and noise resilience. L1L2 regularized inversion, combining sparsity and smoothness constraints, L1L2 regularized inversion, combining sparsity and smoothness constraints, Zeroth- and second-order deconvolution (ZSD), which prioritizes shape fidelity over smoothness, Zeroth- and second-order deconvolution (ZSD), which prioritizes shape fidelity over smoothness, Singular Value Decomposition (SVD), a matrix-based method offering computational stability and noise resilience. Singular Value Decomposition (SVD), a matrix-based method offering computational stability and noise resilience. Following a comparative assessment of inversion outputs across representative depth intervals, the SVD algorithm was selected for final implementation. This method yielded the most geologically consistent T₂ distributions, providing superior resolution of both fast- and slow-relaxing components—crucial for distinguishing micro-, meso-, and macro-pore systems. Figure 9 illustrates the output of the inversion step, where T₂ spectra are plotted alongside other reservoir indicators. The results highlight the vertical heterogeneity of the Asmari Formation and support the identification of intervals dominated by interparticle porosity, vuggy development, or cement-restricted flow pathways. The resulting T₂ data serve as the foundational layer for the pore type-based rock typing framework developed in this research. By anchoring reservoir classification in the physics of hydrogen relaxation behavior, rather than indirect log correlations or lithological assumptions, this methodology offers an innovative, data-rich pathway to capturing the functional diversity of carbonate systems. Fig. 9Vertical processing sequence of CMR data in the studied well: Signal behavior, noise evolution, and phase-amplitude tracking in the Asmari reservoir. This figure illustrates the stepwise transformation of CMR signal attributes across the 3500–3700 m interval. Multiple waveform panels capture the progression from raw echo sequencing to phase correction and amplitude normalization, showcasing the impact of formation composition and borehole conditions—especially the elevated noise response across shale-prone intervals. These insights anchor the inversion of T₂ relaxation spectra used for pore geometry analysis and advanced rock typing in the Asmari Formation. Vertical processing sequence of CMR data in the studied well: Signal behavior, noise evolution, and phase-amplitude tracking in the Asmari reservoir. This figure illustrates the stepwise transformation of CMR signal attributes across the 3500–3700 m interval. Multiple waveform panels capture the progression from raw echo sequencing to phase correction and amplitude normalization, showcasing the impact of formation composition and borehole conditions—especially the elevated noise response across shale-prone intervals. These insights anchor the inversion of T₂ relaxation spectra used for pore geometry analysis and advanced rock typing in the Asmari Formation. Figure 10 illustrates the resulting T₂ distributions across the reservoir interval. In mid-depth sections—characterized by relatively low shale content—the distributions exhibit clear bimodal patterns, reflecting the separation between bound-water-dominated micro-pores and larger, free-fluid-supporting macro-pores. These double peaks are distinctly resolvable, providing strong indicators of dual-porosity behavior and heightened reservoir quality. Conversely, in lower-depth zones where the gamma ray log indicates higher clay volume, the T₂ distributions become increasingly diffuse and disorganized. The absence of distinct peaks and the predominance of short relaxation times in these intervals suggest that the porosity is largely attributable to bound water in microporous clays—a form of porosity with negligible contribution to hydrocarbon flow. Such distinctions are critical in pore type-based rock typing, as they differentiate storage-dominated zones from flow-capable domains. This inversion process does not simply yield a mathematical function—it constructs a high-resolution fingerprint of the pore system at each depth. When combined with lithofacies interpretations and NMR-derived permeability models, the T₂ spectra lay the groundwork for classifying electrofacies and defining functional flow units within the Asmari Formation. Thus, inversion emerges not as a routine computation, but as an interpretive lens through which the dynamic interplay of rock, pore, and fluid can be quantified and modeled. Fig. 10High-Resolution T₂ Spectrum Mapping via SVD-Inverted CMR Logs: Pore System Differentiation Across the Asmari Reservoir in the studied well. This figure visualizes the vertical variation of T₂ relaxation-based porosity obtained through Singular Value Decomposition (SVD) inversion of CMR data across 3500–3700 m. The graphical track highlights zones of well-defined dual-porosity systems—reflecting both effective and bound fluid components—against stratigraphic layering. Variability in the porosity curve emphasizes lithological transitions, shale-prone intervals, and their influence on pore fluid behavior, offering crucial insight for pore-based rock classification in the Asmari Formation. High-Resolution T₂ Spectrum Mapping via SVD-Inverted CMR Logs: Pore System Differentiation Across the Asmari Reservoir in the studied well. This figure visualizes the vertical variation of T₂ relaxation-based porosity obtained through Singular Value Decomposition (SVD) inversion of CMR data across 3500–3700 m. The graphical track highlights zones of well-defined dual-porosity systems—reflecting both effective and bound fluid components—against stratigraphic layering. Variability in the porosity curve emphasizes lithological transitions, shale-prone intervals, and their influence on pore fluid behavior, offering crucial insight for pore-based rock classification in the Asmari Formation. Permeability Estimation through NMR-integrated petrophysics The CMR-based approach integrates porosity, fluid mobility, and pore size distribution into a coherent framework for dynamic flow assessment. Permeability (k), defined as the rock’s capacity to transmit fluids, is intimately related to the average size, distribution, and connectivity of its pore throats. In carbonate reservoirs—especially those with diagenetically altered textures like the Asmari Formation—these factors vary vertically and laterally, necessitating a resolution-rich, physics-based estimation method. This study employs two widely accepted NMR-derived permeability models: KTIM (Timur-Coates): A generic empirical model tailored to formations with variable pore structures. KSDR The Schlumberger–Doll Research model, expressed as Eq. (13)36,101. 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k_{{\mathrm{SDR}}}}=C \cdot \phi _{{{\mathrm{NMR}}}}^{4} \cdot T_{{2,{\mathrm{log}}}}^{2}$$\end{document} where kSDR is the NMR-derived permeability (mD), C is a lithology-sensitive calibration constant, φNMR is NMR porosity (fraction), and T2,log is the log-mean of the T₂ distribution (ms). In this analysis, permeability logs generated from both KTIM and KSDR equations were compared and cross-validated using available core measurements and hydrocarbon productivity indicators. The log-mean T₂ (T₂LM) was extracted directly from field-processed CMR datasets, ensuring sensitivity to pore size variability and effective fluid transmissivity. Importantly, model calibration was informed by mineralogical context, with special consideration for gas-bearing intervals where fluid effects may skew T₂LM values. The strength of this method lies in its dual capacity to detect subtle heterogeneities—such as vug-restricted flow zones or capillary-sealed microfacies—and to offer actionable inputs for reservoir simulation and well completion design. Moreover, by aligning permeability estimates with NMR-derived porosity and BFV trends, a multi-parameter view of reservoir quality emerges, enabling the classification of flow units that are compositionally distinct and hydraulically functional. Assumed parameter values for permeability modeling In this study, two distinct sets of empirical constants were applied to estimate permeability from NMR log data using models sensitive to lithological context. For detailed mineral-specific coefficients, refer to Tables 4 and 5. Representative values for major lithologies include: Sandstone formations: C1 = 5.16, b1 = 0.8, a1 = 1.13.Carbonate formations: C1 = 0.1–0.2, b1 = 1.5–2.5, a1 = 0.9–1.05. Sandstone formations: C1 = 5.16, b1 = 0.8, a1 = 1.13. Sandstone formations: C1 = 5.16, b1 = 0.8, a1 = 1.13. Carbonate formations: C1 = 0.1–0.2, b1 = 1.5–2.5, a1 = 0.9–1.05. Carbonate formations: C1 = 0.1–0.2, b1 = 1.5–2.5, a1 = 0.9–1.05. These values were incorporated into the generalized NMR-derived permeability (Eq. (12)): This approach allows the model to adapt to textural differences between clastic and carbonate units, particularly important in heterogeneous formations like the Asmari Reservoir.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k_{{\mathrm{NMR}}}}={C_1} \cdot {\phi ^{{a_1}}} \cdot T_{{2,{\mathrm{log}}}}^{{{b_1}}}$$\end{document} Where: KNMR is the estimated permeability (mD), ϕ is the NMR-measured porosity (fractional), T2log is the logarithmic mean of the T₂ distribution (ms), and C1,a1,b1 are empirical coefficients that vary with lithology. Permeability Estimation via the Ktim model To complement the KSDR permeability model, the Ktim model was applied, utilizing lithology-specific T₂ cutoff times to partition the T₂ relaxation spectrum. This approach distinguishes free and bound fluid components based on pore connectivity and capillary retention, independent of hydrocarbon presence94,102. In the Ktim method, fluid classification hinges on predefined T₂ cutoffs—set at 33 ms for sandstone and 100 ms for carbonate formations. Signals with T₂ values above these thresholds are attributed to Free Fluid Index (FFI), representing moveable fluids residing in macro- and mesopores, while those below correspond to the Bulk Volume Irreducible (BVI), indicative of capillary-bound fluids retained in micropores, often due to surface tension effects or mineral-fluid interactions. The model formulates permeability (K) as Eq. (13)103–105:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k_{{\mathrm{tim}}}}={C_2} \cdot {\left( {\frac{{{\phi _{{\mathrm{NMR}}}}}}{{100}}} \right)^{{a_2}}} \cdot {\left( {\frac{{{\mathrm{FFI}}}}{{{\mathrm{BVI}}}}} \right)^{{b_2}}}$$\end{document} where ktim is permeability (mD), C₂, b₂, and a₂ are empirical constants determined by lithofacies, φNMR is porosity (%), FFI is Free Fluid Index, and BVI is Bulk Volume Irreducible. In this study, default coefficients were initially selected from published values and then weighted by the relative mineralogical composition of each depth interval using volume-based lithology fractions. The mineral-specific coefficients used in the Ktim model are detailed in Table 5, with representative values including C2 = 1.35, b2 = 2.11, a2 = 0.005 for sandstones (WSAN) and C2 = 0.4–0.6, b2 = 5.79–6.51, a2 = 0.5 for carbonates (WCLC/WDOL). Additionally, for intervals with mixed mineralogy, lithology-specific coefficients were interpolated using the generalized mixing model as Eq. (14)106–108:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_{{\mathrm{eff}}}}=\sum\limits_{{i=1}}^{N} {{w_i}} \cdot {C_i}$$\end{document} where Ceff is the effective empirical constant for a mixed lithology interval, wi is the volumetric proportion of mineral i, and Ci is the mineral-specific constant. (Applies similarly to parameters a and b). Lithology-specific coefficient mixing model for permeability estimation in heterogeneous intervals. The model calculates effective empirical constants (C eff, a eff, b eff) by weighting mineral-specific values by their volumetric proportions (wi). Mineral phases include quartz (sandstone), calcite, dolomite, anhydrite, and clay, with corresponding empirical constants provided in Tables 4 and 5. This formulation allows the permeability model to adapt flexibly to interbedded or compositionally graded formations—such as those frequently observed in the Asmari interval. Given the absence of core-measured permeability data for direct calibration, cross-validation was conducted using synthetic Full Set log curves aligned with CMR-derived permeability estimates. This hybrid calibration strategy ensured consistency across lithological transitions and allowed for the identification of permeability zones with meaningful dynamic contribution to flow. Final results indicated that the average permeability obtained via the Ktim model was 5.1 mD, while the KSDR -derived estimate for the same interval was notably higher at 87 mD. The discrepancy reflects the models’ differing sensitivities: Ktim places greater emphasis on bound/free fluid segmentation, whereas KSDR integrates porosity and relaxation time dynamics more directly. Together, these models offer a multi-faceted view of reservoir transmissivity, providing valuable input for pore-type classification, flow unit definition, and reservoir simulation under a pore type-based paradigm. As the results, the hybrid KSDR-Ktim model showed significant improvement over conventional methods (Fig. 11; Supplementary Table 3): Table 4Coefficients corresponding to the KSDR equation.WANHWCLAWCLCWDOLWSAN C 1 0.050.050.10.25.16 b 1 442.51.50.8 a 1 0.10.10.91.051.13Note: Mineral volume abbreviations: WANH (anhydrite), WCLA (clay), WCLC (calcite), WDOL (dolomite), WSAN (sandstone). Coefficients remain consistent with previous permeability modeling approaches in the Asmari Formation. Table 4Coefficients corresponding to the KSDR equation.WANHWCLAWCLCWDOLWSAN C 1 0.050.050.10.25.16 b 1 442.51.50.8 a 1 0.10.10.91.051.13Note: Mineral volume abbreviations: WANH (anhydrite), WCLA (clay), WCLC (calcite), WDOL (dolomite), WSAN (sandstone). Coefficients remain consistent with previous permeability modeling approaches in the Asmari Formation. Coefficients corresponding to the KSDR equation. Note: Mineral volume abbreviations: WANH (anhydrite), WCLA (clay), WCLC (calcite), WDOL (dolomite), WSAN (sandstone). Coefficients remain consistent with previous permeability modeling approaches in the Asmari Formation. Table 5Coefficients corresponding to the Ktim equation.WANHWCLAWCLCWDOLWSAN C 2 0.010.010.40.61.35 b 2 5.145.956.515.792.11 a 2 0.50.120.50.50.005 Table 5Coefficients corresponding to the Ktim equation.WANHWCLAWCLCWDOLWSAN C 2 0.010.010.40.61.35 b 2 5.145.956.515.792.11 a 2 0.50.120.50.50.005 Coefficients corresponding to the Ktim equation. Fig. 11Core vs. predicted permeability crossplot. Cross-plot of core-measured versus predicted permeability. Dynamic model (orange) reduces scatter (RMSE = 0.21) compared to static KSDR (blue, RMSE = 0.31). Core vs. predicted permeability crossplot. Cross-plot of core-measured versus predicted permeability. Dynamic model (orange) reduces scatter (RMSE = 0.21) compared to static KSDR (blue, RMSE = 0.31). As the key findings of predicted permeability model: KSDR model: RMSE = 0.31 (carbonates), 0.28 (sandstones).Dynamic model: RMSE = 0.21 (carbonates), 0.19 (sandstones).Clay correction reduced overprediction by 37% in Vclay > 25% zones. KSDR model: RMSE = 0.31 (carbonates), 0.28 (sandstones). KSDR model: RMSE = 0.31 (carbonates), 0.28 (sandstones). Dynamic model: RMSE = 0.21 (carbonates), 0.19 (sandstones). Dynamic model: RMSE = 0.21 (carbonates), 0.19 (sandstones). Clay correction reduced overprediction by 37% in Vclay > 25% zones. Clay correction reduced overprediction by 37% in Vclay > 25% zones. Flow unit delineation and production validation Synthesis of NMR, electrofacies, and permeability data enabled identification of three principal flow units (FU1–FU3) with distinct production characteristics (Table 6). The high-quality flow unit (FU1), characterized by porosity > 16% and permeability > 25 mD, consisted predominantly (92%) of sandstone lithologies and accounted for 72% of tested production. Moderate-quality zones (FU2) with porosity 10–16% and permeability 5–25 mD were primarily dolomitized carbonates (68%) contributing 23% of production. Non-pay intervals (FU3) exhibited poor reservoir properties (porosity < 10%, permeability < 5 mD) due to high clay and anhydrite content. This flow unit classification showed 84% agreement with production logging tool (PLT) results, validating the pore type-based approach’s predictive capability. Table 6Flow unit properties versus production performance, ranked by dynamic reservoir performance, showing dominance of sandstone-rich FU1.Flow UnitPorosity (%)Permeability (mD)LithologyProduction Contribution (%)FU1> 16> 2592% Sandstone72FU210–165–2568% Dolomite23FU3< 10< 5High clay/anhydrite0 Table 6Flow unit properties versus production performance, ranked by dynamic reservoir performance, showing dominance of sandstone-rich FU1.Flow UnitPorosity (%)Permeability (mD)LithologyProduction Contribution (%)FU1> 16> 2592% Sandstone72FU210–165–2568% Dolomite23FU3< 10< 5High clay/anhydrite0 Flow unit properties versus production performance, ranked by dynamic reservoir performance, showing dominance of sandstone-rich FU1. Key findings. Pore architecture dominance: Sandstone subunits (15% thickness) contribute 73% of flow capacity due to macroporosity (PCI > 1.8).Dynamic permeability necessity: Clay corrections reduced errors by 37% in shaly zones (Vclay > 25%). Pore architecture dominance: Sandstone subunits (15% thickness) contribute 73% of flow capacity due to macroporosity (PCI > 1.8). Pore architecture dominance: Sandstone subunits (15% thickness) contribute 73% of flow capacity due to macroporosity (PCI > 1.8). Dynamic permeability necessity: Clay corrections reduced errors by 37% in shaly zones (Vclay > 25%). Dynamic permeability necessity: Clay corrections reduced errors by 37% in shaly zones (Vclay > 25%). Integrated petrophysical visualization of pore structure and fluid dynamics Figure 12 presents a comprehensive vertical log suite capturing the petrophysical behavior of the Asmari Formation in the studied well, spanning approximately 3424 to 3770 m. This multi-track synthesis integrates Combinable Magnetic Resonance (CMR), Elemental Capture Spectroscopy (ECS), and derived elastic and fluid partitioning metrics to resolve static and dynamic reservoir properties through a pore type-based lens. The initial tracks illustrate borehole stability and lithological zoning via tool-response overlays, while subsequent columns delineate variations in total porosity, free and bound fluid fractions, and T₂ relaxation trends. Track 5 merges CMR and ECS-derived signatures, marking transitions between matrix-dominated and fluid-dominated intervals with high vertical resolution. Track 6 displays T₂ distribution shifts, distinguishing zones of effective macroporosity from capillary-bound microstructures. Tracks 7 and 8 reveal fluid saturation and mobility regimes through T₂ log-mean (T₂LM) behaviors and Spectral Bound Fluid Volume (SBFV) trends. Elevated SBFV values in lower-depth intervals, coupled with subdued T₂LM signatures, indicate fluid retention in poorly connected pores, likely associated with clay-enriched microfacies or diagenetically altered dolomitic units. This integrated panel underscores the relationship between petrophysical signatures, mineralogical transitions, and pore architecture, offering a dynamic reinterpretation of reservoir quality beyond compositional markers. The integration of geochemical data provides a genetic context for the observed fluid dynamics. The boron concentration log (Fig. 12), a established paleo-salinity proxy, reveals a fundamental depositional control on the modern pore-fluid system. Intervals with elevated boron (> 80 ppm), indicative of restricted marine conditions, consistently correlate with high SBFV and suppressed T₂LM signatures—a direct link between depositional environment and the development of microporous, capillary-bound fluid networks. Conversely, the high-flow macroporous units (EF3) exhibit lower boron concentrations (30–50 ppm), supporting their origin in freshwater-influenced high-energy settings where superior initial permeability was established and preserved. This geochemical evidence solidifies that the present-day fluid partitioning and pore architecture are not random but are deeply rooted in the original depositional fabric of the Asmari Formation. Fig. 12Integrated interpretation panel of studied well: Pore system evolution, lithological variation, and fluid partitioning across the Asmari Formation. This panel presents an integrated CMR-driven petrophysical profile across the studied depth interval. The tracks highlight systematic transitions in porosity distribution, relaxation characteristics, and permeability—illustrating the pore-scale heterogeneity and mineralogical complexity inherent to the Asmari carbonate system. Elevated SBFV and suppressed T₂LM zones are mapped as indicators of microporosity and capillary trapping, while permeability overlays reveal hydraulically active intervals aligned with macro-porous facies. This visualization reinforces the effectiveness of a pore type-based approach in identifying functional flow units and refining rock typing in carbonate reservoirs. Integrated interpretation panel of studied well: Pore system evolution, lithological variation, and fluid partitioning across the Asmari Formation. This panel presents an integrated CMR-driven petrophysical profile across the studied depth interval. The tracks highlight systematic transitions in porosity distribution, relaxation characteristics, and permeability—illustrating the pore-scale heterogeneity and mineralogical complexity inherent to the Asmari carbonate system. Elevated SBFV and suppressed T₂LM zones are mapped as indicators of microporosity and capillary trapping, while permeability overlays reveal hydraulically active intervals aligned with macro-porous facies. This visualization reinforces the effectiveness of a pore type-based approach in identifying functional flow units and refining rock typing in carbonate reservoirs. Stratigraphic and petrophysical integration Figure 13 delivers a depth-aligned petrophysical mosaic for the studied well, incorporating elemental volumes, fluid saturations, and textural logs across the Asmari Formation. The log suite captures lithofacies variability through gamma-ray curves, mineral volumetrics, and lithology-specific corrections. Calcite- and dolomite-dominant zones interbed with evaporitic streaks and minor quartz-illitic clay incursions, reflecting depositional stacking patterns. The juxtaposition of irreducible (SWI) and effective water saturation (SWE) highlights pore accessibility, with elevated SWE-SWI divergence in illite-rich zones confirming capillary-bound fluids in microporous frameworks. Discrepancies in neutron-density porosity align with lithofacies mixing, necessitating lithology-agnostic indicators like Total CMR Porosity (TCMR). Advanced correction parameters (including bad-hole flags and borehole adaptors) maintain data fidelity across heterogeneous intervals. The integrated analysis correlates petrophysical anomalies with mineralogical endpoints, demonstrating that pore system architecture—rather than composition alone—governs reservoir functionality. Fig. 13Stratigraphic-petrophysical integration for the studied well: High-resolution evaluation of lithology, porosity, and pore connectivity in the Asmari Reservoir. This composite log displays depth-aligned petrophysical parameters, integrating gamma-ray response, porosity metrics, and fluid distribution to characterize depositional texture and flow potential. Results underpin electrofacies mapping and pore type-based stratigraphic modeling. Stratigraphic-petrophysical integration for the studied well: High-resolution evaluation of lithology, porosity, and pore connectivity in the Asmari Reservoir. This composite log displays depth-aligned petrophysical parameters, integrating gamma-ray response, porosity metrics, and fluid distribution to characterize depositional texture and flow potential. Results underpin electrofacies mapping and pore type-based stratigraphic modeling. Comparative multi-log analysis of reservoir properties Figure 14 presents a rigorous comparative analysis between Combinable Magnetic Resonance (CMR) measurements and conventional Full Set log responses across the 3500–3770-meter interval in the studied well of the study area. The dual-log evaluation methodology demonstrates remarkable consistency in porosity quantification between the two independent measurement systems, with both datasets identifying three principal high-porosity zones at 3540–3560 m, 3590–3709 m, and 3738–3748 m depths. This congruence validates the CMR-derived porosity curves while confirming the existence of reservoir-quality pore networks through multiple measurement techniques. The permeability characterization, incorporating both Ktim and KSDR model outputs, reveals enhanced fluid connectivity within these porous intervals, particularly in meso- to macro-porous domains where pore throat development appears most favorable. Track-based analysis of porosity-permeability cross-plots demonstrates systematic relationships between pore volume and flow capacity, with the highest permeability values corresponding to zones exhibiting well-connected pore architectures. However, detailed saturation profiling uncovers a critical reservoir complexity - significant portions of these high-quality pore systems contain water rather than hydrocarbons, as evidenced by divergent trends between hydrocarbon volume (VHC) and water saturation (Sw) logs. This fluid distribution anomaly manifests most prominently in the 3590–3709 m interval, where favorable porosity-permeability characteristics coincide with elevated water saturations. Multiple mechanisms may explain this phenomenon, including capillary entrapment in microporous networks, structural positioning within the hydrocarbon column, or the presence of bimodal pore systems where hydrocarbons preferentially occupy larger pore spaces while water remains bound in smaller throats. Track correlations between CMR-derived pore size distributions and conventional saturation logs provide compelling evidence for this complex fluid partitioning behavior. The study identifies two principal hydrocarbon-bearing intervals (3540–3560 m and 3590–3625 m) through integrated analysis of multiple petrophysical indicators. These zones exhibit the diagnostic convergence of high hydrocarbon volumes, low water saturations, and sustained porosity measurements across all log types. Mineralogical analysis reveals these productive intervals correlate strongly with sandstone-rich lithofacies, a finding that substantially modifies the traditional view of the Asmari Formation’s reservoir characteristics. The sandstone units demonstrate superior pore efficiency and fluid deliverability compared to adjacent carbonate-dominated sections, establishing their role as primary flow units despite occurring within a predominantly carbonate stratigraphic framework. This multi-log integration achieves comprehensive reservoir characterization by combining CMR-based pore-scale insights with conventional log responses. The methodology successfully discriminates between various reservoir components, identifying: (1) high-quality storage zones with optimal porosity development; (2) effective flow units exhibiting both storage capacity and permeability; and (3) capillary barriers where fluid mobility appears restricted despite favorable porosity indicators. The resultant reservoir model provides a robust framework for understanding fluid distribution patterns and predicting production behavior in this complex carbonate-clastic system. Fig. 14Comparative multi-log analysis of the Asmari Formation in the studied well: contrasting CMR-derived pore system attributes with full set petrophysical responses. This figure juxtaposes processed CMR data with traditional full set logs across the 3500–3770 m interval—offering stratified visualization of porosity, permeability, and fluid distribution characteristics. By aligning depth-resolved measurements—ranging from hydrogen-index-driven porosity to lithology-sensitive saturation profiles—it reveals both convergence and divergence in pore system interpretation. The integration supports pore type-based modeling by validating NMR-informed insights against conventional logs—enhancing confidence in electrofacies classification and reservoir flow modeling. Comparative multi-log analysis of the Asmari Formation in the studied well: contrasting CMR-derived pore system attributes with full set petrophysical responses. This figure juxtaposes processed CMR data with traditional full set logs across the 3500–3770 m interval—offering stratified visualization of porosity, permeability, and fluid distribution characteristics. By aligning depth-resolved measurements—ranging from hydrogen-index-driven porosity to lithology-sensitive saturation profiles—it reveals both convergence and divergence in pore system interpretation. The integration supports pore type-based modeling by validating NMR-informed insights against conventional logs—enhancing confidence in electrofacies classification and reservoir flow modeling. Machine learning integration for enhanced reservoir characterization Random forest (RF) for pore system classification Random Forest, an ensemble learning method, was utilized to classify pore systems based on NMR T₂ relaxation spectra and conventional log data. The model was trained on labeled datasets from the studied well, where pore types (micro-, meso-, and macro-pores) were identified using NMR-derived cutoffs. Feature importance analysis revealed that T₂ log-mean (T₂LM), bound fluid volume (BFV), and effective porosity (PHIE) were the most influential predictors. The RF model achieved an accuracy of 92% in pore-type classification, with a cross-validated correlation coefficient (R²) of 0.89 between predicted and observed pore distributions. This high accuracy confirms the reliability of NMR-derived pore typing and its alignment with geological facies. SVM and SHAP analysis for electrofacies prediction SVM was applied to predict electrofacies using a combination of NMR and conventional log inputs44,109. The kernel-based algorithm effectively handled the non-linear relationships between petrophysical parameters and facies boundaries. The SVM model achieved an overall accuracy of 88%, with distinct clusters corresponding to sandstone-rich (high permeability) and carbonate-dominated (variable porosity) intervals. Notably, the model identified subtle transitions in pore connectivity that were overlooked by traditional lithology-based methods, reinforcing the novelty of the pore type-based framework. Machine learning classification identified five electrofacies (EF1–EF5) with distinct petrophysical and geological characteristics. Tight carbonates (EF1) exhibited minimal porosity (4.2%) and permeability (0.3 mD), while dolomitized packstones (EF2) showed moderate reservoir quality (14.7% porosity, 9.1 mD permeability). The sandstone facies (EF3), though representing only 15% of total thickness, contained 58% of movable hydrocarbons due to superior macroporosity (18.2% porosity, 54.6 mD permeability). SHAP interpretability analysis revealed T₂ geometric mean as the primary electrofacies discriminator (42% contribution). Fig. 15Confusion matrix for electrofacies classification using the optimized Random Forest model. The matrix displays classification performance across the three primary electrofacies (EF1: microporous, EF2: mesoporous, EF3: macroporous), with EF3 (sandstone-rich) achieving 95% precision due to distinct pore architecture signatures. Overall balanced accuracy: 91.2% ± 2.3% (10-fold cross-validation). Confusion matrix for electrofacies classification using the optimized Random Forest model. The matrix displays classification performance across the three primary electrofacies (EF1: microporous, EF2: mesoporous, EF3: macroporous), with EF3 (sandstone-rich) achieving 95% precision due to distinct pore architecture signatures. Overall balanced accuracy: 91.2% ± 2.3% (10-fold cross-validation). The confusion matrix in Fig. 15 quantifies electrofacies classification performance, revealing particularly strong identification of EF3 (macroporous sandstone) intervals, which aligns with the study’s key finding regarding sandstone-dominated flow units. Misclassification primarily occurs between EF1 and EF2 in transitional carbonate intervals, reflecting the continuum of pore-size distributions in diagenetically altered zones. The model achieved 89% classification accuracy (κ = 0.87) as demonstrated in the confusion matrix (Table 7). SHAP values for feature ″i″ is computed as Eq. (15)50,76:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\phi _i}=\sum\limits_{{S \subseteq F \setminus \{ i\} }} {\frac{{|S|!(|F| - |S| - 1)!}}{{|F|!}}} (f(S \cup \{ i\} ) - f(S))$$\end{document} where φj is the SHAP value for feature j, F is the set of all features, S is a subset of features excluding j, |S| is the size of subset S, and f is the model prediction function. Table 7Machine learning-derived electrofacies characteristics. EF3 (sandstone) contributes disproportionately to flow.FaciesLithologyAvg. Φ (%)Avg. K (mD)Dominant Pore TypeEF1Tight carbonate4.20.3MicroEF2Dolomitized packstone14.79.1MesoEF3Sandstone18.254.6Macro Table 7Machine learning-derived electrofacies characteristics. EF3 (sandstone) contributes disproportionately to flow.FaciesLithologyAvg. Φ (%)Avg. K (mD)Dominant Pore TypeEF1Tight carbonate4.20.3MicroEF2Dolomitized packstone14.79.1MesoEF3Sandstone18.254.6Macro Machine learning-derived electrofacies characteristics. EF3 (sandstone) contributes disproportionately to flow. Furthermore, SHAP values were computed to elucidate the contribution of each input feature to the ML models’ predictions. Fig. 16SHAP summary plot illustrating feature contributions to pore-type classification. The plot displays the impact of the top 9 features on model predictions, with points colored by feature value (red: high, blue: low). T₂ geometric mean dominates feature importance (mean |SHAP| = 0.47), while negative SHAP values for clay volume confirm its inhibitory effect on flow capacity. Each point represents a single prediction from the test dataset (n = 1,250 samples). SHAP summary plot illustrating feature contributions to pore-type classification. The plot displays the impact of the top 9 features on model predictions, with points colored by feature value (red: high, blue: low). T₂ geometric mean dominates feature importance (mean |SHAP| = 0.47), while negative SHAP values for clay volume confirm its inhibitory effect on flow capacity. Each point represents a single prediction from the test dataset (n = 1,250 samples). SHAP analysis provides transparent interpretation of machine learning decisions, as visualized in Fig. 16. The horizontal dispersion of points reveals the magnitude and direction of each feature’s impact, with T₂ geometric mean demonstrating the strongest positive influence on macroporous classification, while elevated clay volumes consistently reduce predicted flow potential. Key findings include: T₂LM and BFV were the top contributors to pore-type classification, highlighting their role in differentiating flow units.Permeability (K) and free fluid volume (FFV) were critical for distinguishing high-quality reservoir zones.Clay volume (VCL) and shale content negatively impacted pore connectivity, aligning with the study’s observations of capillary-bound fluids in shale-rich intervals. T₂LM and BFV were the top contributors to pore-type classification, highlighting their role in differentiating flow units. T₂LM and BFV were the top contributors to pore-type classification, highlighting their role in differentiating flow units. Permeability (K) and free fluid volume (FFV) were critical for distinguishing high-quality reservoir zones. Permeability (K) and free fluid volume (FFV) were critical for distinguishing high-quality reservoir zones. Clay volume (VCL) and shale content negatively impacted pore connectivity, aligning with the study’s observations of capillary-bound fluids in shale-rich intervals. Clay volume (VCL) and shale content negatively impacted pore connectivity, aligning with the study’s observations of capillary-bound fluids in shale-rich intervals. The SHAP summary plot (Table 8; Fig. 17) quantifies these relationships, demonstrating how ML enhances the interpretability of NMR-derived pore systems. The high correlation (R² > 0.85) between SHAP-derived feature importance and NMR-based petrophysical trends validates the pore type-based model’s ability to capture dynamic reservoir behavior (Data presented in Supplementary Table 4). Table 8SHAP feature importance for pore system classification. SHAP values quantifying the contribution of NMR and conventional log features to pore system classification. High positive values indicate strong predictive power for flow units, while negative values (e.g., VCL) highlight inhibitory effects.FeatureSHAP Value (Importance)Correlation with NMR Metrics (R²)T₂LM0.450.91BFV0.380.87PHIE0.320.85Permeability (K)0.290.82Clay Volume (VCL)−0.250.79 Table 8SHAP feature importance for pore system classification. SHAP values quantifying the contribution of NMR and conventional log features to pore system classification. High positive values indicate strong predictive power for flow units, while negative values (e.g., VCL) highlight inhibitory effects.FeatureSHAP Value (Importance)Correlation with NMR Metrics (R²)T₂LM0.450.91BFV0.380.87PHIE0.320.85Permeability (K)0.290.82Clay Volume (VCL)−0.250.79 SHAP feature importance for pore system classification. SHAP values quantifying the contribution of NMR and conventional log features to pore system classification. High positive values indicate strong predictive power for flow units, while negative values (e.g., VCL) highlight inhibitory effects. In Table 8, SHAP values quantifying the contribution of NMR and conventional log features to pore system classification. High positive values indicate strong predictive power for flow units, while negative values (e.g., VCL) highlight inhibitory effects. Fig. 17SHAP analysis of machine learning model inputs. T₂ geometric mean dominates feature importance (47 ± 3%), validating NMR-derived parameters as primary controls on pore-type classification. SHAP analysis of machine learning model inputs. T₂ geometric mean dominates feature importance (47 ± 3%), validating NMR-derived parameters as primary controls on pore-type classification. Furthermore, the methodology was validated through core-log integration showing strong correlation between NMR and helium porosity (R² = 0.89) and between predicted versus measured permeability (RMSE = 0.32 log units). Production data verification demonstrated 83% agreement between machine learning-classified flow units and tested intervals. Monte Carlo simulation with 1000 iterations assessed parameter sensitivity and model uncertainty. Sensitivity analysis of results The ML results corroborate the study’s core findings, particularly the superior hydrocarbon potential of sandstone-rich intervals within the Asmari Formation. The RF and SVM models confirmed that traditional lithology-based classifications often misrepresent flow potential, whereas pore type-based metrics (e.g., T₂LM, BFV) provide a more accurate basis for reservoir zonation. The integration of SHAP further distinguishes this work by offering a transparent, physics-informed interpretation of ML outputs, bridging the gap between data-driven and petrophysical insights. To assess the robustness of the findings, a comprehensive sensitivity analysis was conducted on three critical components of the study: (1) pore connectivity index (PCI), (2) machine learning (ML) classification performance, and (3) dynamic permeability model accuracy. Pore connectivity index (PCI) The PCI was calculated from NMR T₂ distribution skewness, with the following sensitivity tests performed: Cutoff variation: The T₂ cutoff thresholds (33 ms for sandstone, 100 ms for carbonates) were perturbed by ± 20%. Results showed PCI values remained stable (ΔPCI < 5%) when cutoffs varied within ± 10%, but deviations beyond 15% led to misclassification of mesoporous zones.Noise robustness: Gaussian noise (σ = 5% of signal amplitude) was introduced to raw T₂ spectra. The PCI maintained a correlation coefficient (R²) > 0.85 with core-derived permeability, confirming resilience to moderate measurement noise.Depth resolution impact: Up-sampling and down-sampling NMR logs to 6-inch, 1-ft, and 2-ft resolutions revealed that PCI accuracy degraded by < 8% at 1-ft resolution but dropped by 18% at 2-ft resolution, justifying the use of high-resolution (6-inch) data. Cutoff variation: The T₂ cutoff thresholds (33 ms for sandstone, 100 ms for carbonates) were perturbed by ± 20%. Results showed PCI values remained stable (ΔPCI < 5%) when cutoffs varied within ± 10%, but deviations beyond 15% led to misclassification of mesoporous zones. Cutoff variation: The T₂ cutoff thresholds (33 ms for sandstone, 100 ms for carbonates) were perturbed by ± 20%. Results showed PCI values remained stable (ΔPCI < 5%) when cutoffs varied within ± 10%, but deviations beyond 15% led to misclassification of mesoporous zones. Noise robustness: Gaussian noise (σ = 5% of signal amplitude) was introduced to raw T₂ spectra. The PCI maintained a correlation coefficient (R²) > 0.85 with core-derived permeability, confirming resilience to moderate measurement noise. Noise robustness: Gaussian noise (σ = 5% of signal amplitude) was introduced to raw T₂ spectra. The PCI maintained a correlation coefficient (R²) > 0.85 with core-derived permeability, confirming resilience to moderate measurement noise. Depth resolution impact: Up-sampling and down-sampling NMR logs to 6-inch, 1-ft, and 2-ft resolutions revealed that PCI accuracy degraded by < 8% at 1-ft resolution but dropped by 18% at 2-ft resolution, justifying the use of high-resolution (6-inch) data. Depth resolution impact: Up-sampling and down-sampling NMR logs to 6-inch, 1-ft, and 2-ft resolutions revealed that PCI accuracy degraded by < 8% at 1-ft resolution but dropped by 18% at 2-ft resolution, justifying the use of high-resolution (6-inch) data. Machine learning (ML) classification The robustness of the RF and SVM models was evaluated under a range of conditions to test their stability and dependence on input features: Feature ablation: Features were sequentially removed in ascending order of their SHAP importance to assess their individual contribution to model performance.Redundancy analysis: Cross-correlation analysis between NMR and conventional log features was conducted.Class imbalance handling: Synthetic Minority Over-sampling Technique (SMOTE) was applied to address underrepresented microporous facies.Hyperparameter perturbation: The sensitivity to key parameters like RF tree depth and SVM kernel width was tested. Feature ablation: Features were sequentially removed in ascending order of their SHAP importance to assess their individual contribution to model performance. Feature ablation: Features were sequentially removed in ascending order of their SHAP importance to assess their individual contribution to model performance. Redundancy analysis: Cross-correlation analysis between NMR and conventional log features was conducted. Redundancy analysis: Cross-correlation analysis between NMR and conventional log features was conducted. Class imbalance handling: Synthetic Minority Over-sampling Technique (SMOTE) was applied to address underrepresented microporous facies. Class imbalance handling: Synthetic Minority Over-sampling Technique (SMOTE) was applied to address underrepresented microporous facies. Hyperparameter perturbation: The sensitivity to key parameters like RF tree depth and SVM kernel width was tested. Hyperparameter perturbation: The sensitivity to key parameters like RF tree depth and SVM kernel width was tested. The results from these tests confirmed the model’s robustness: The feature ablation analysis (Fig. 18) confirmed model stability to the removal of low-importance features (SHAP value < 0.05), which reduced accuracy by only 2–3%. However, it also highlighted the critical dependence on NMR-derived parameters. A sharp accuracy decline followed the removal of T₂ geometric mean and BFV, underscoring their fundamental role in pore-type discrimination.To specifically address potential overfitting, a comparative analysis between single-feature and multi-feature configurations was performed. Using only the top-contributing feature (T₂ geometric mean) achieved 79% accuracy, while the full 9-feature model reached 91% accuracy. This demonstrates that the additional features provide complementary predictive power without overfitting, while confirming the indispensable role of T₂ geometric mean (its exclusion caused a 12% drop in accuracy).Redundancy analysis revealed only moderate correlations (R² < 0.65) between features, validating the feature selection strategy by confirming minimal information redundancy.Applying SMOTE successfully improved recall for minority classes from 0.82 to 0.89 without compromising precision (Δ < 2%).The model performance was robust to hyperparameter perturbations, with AUC fluctuations remaining below 3%. The feature ablation analysis (Fig. 18) confirmed model stability to the removal of low-importance features (SHAP value < 0.05), which reduced accuracy by only 2–3%. However, it also highlighted the critical dependence on NMR-derived parameters. A sharp accuracy decline followed the removal of T₂ geometric mean and BFV, underscoring their fundamental role in pore-type discrimination. The feature ablation analysis (Fig. 18) confirmed model stability to the removal of low-importance features (SHAP value < 0.05), which reduced accuracy by only 2–3%. However, it also highlighted the critical dependence on NMR-derived parameters. A sharp accuracy decline followed the removal of T₂ geometric mean and BFV, underscoring their fundamental role in pore-type discrimination. To specifically address potential overfitting, a comparative analysis between single-feature and multi-feature configurations was performed. Using only the top-contributing feature (T₂ geometric mean) achieved 79% accuracy, while the full 9-feature model reached 91% accuracy. This demonstrates that the additional features provide complementary predictive power without overfitting, while confirming the indispensable role of T₂ geometric mean (its exclusion caused a 12% drop in accuracy). To specifically address potential overfitting, a comparative analysis between single-feature and multi-feature configurations was performed. Using only the top-contributing feature (T₂ geometric mean) achieved 79% accuracy, while the full 9-feature model reached 91% accuracy. This demonstrates that the additional features provide complementary predictive power without overfitting, while confirming the indispensable role of T₂ geometric mean (its exclusion caused a 12% drop in accuracy). Redundancy analysis revealed only moderate correlations (R² < 0.65) between features, validating the feature selection strategy by confirming minimal information redundancy. Redundancy analysis revealed only moderate correlations (R² < 0.65) between features, validating the feature selection strategy by confirming minimal information redundancy. Applying SMOTE successfully improved recall for minority classes from 0.82 to 0.89 without compromising precision (Δ < 2%). Applying SMOTE successfully improved recall for minority classes from 0.82 to 0.89 without compromising precision (Δ < 2%). The model performance was robust to hyperparameter perturbations, with AUC fluctuations remaining below 3%. The model performance was robust to hyperparameter perturbations, with AUC fluctuations remaining below 3%. Fig. 18Feature ablation analysis showing model performance degradation with sequential feature removal. The plot demonstrates classification accuracy as features are removed in ascending order of SHAP importance. Performance remains stable until removal of core NMR parameters (T₂ geometric mean, BFV), causing significant accuracy drops, validating the feature selection strategy and confirming the indispensability of NMR-derived metrics. Feature ablation analysis showing model performance degradation with sequential feature removal. The plot demonstrates classification accuracy as features are removed in ascending order of SHAP importance. Performance remains stable until removal of core NMR parameters (T₂ geometric mean, BFV), causing significant accuracy drops, validating the feature selection strategy and confirming the indispensability of NMR-derived metrics. Dynamic permeability model sensitivity The hybrid KSDR-Ktim permeability model was evaluated under three scenarios: Clay content variability: Increasing Vclay from 10% to 30% reduced permeability predictions by 22% in carbonates but only 9% in sandstones, highlighting lithology-dependent sensitivity.Fluid saturation effects: At Sw = 40%, the model overpredicted permeability by 8% compared to core data; this error diminished to < 2% when Sw ≤ 25%.Empirical constant calibration: Adjusting the carbonate constant (C₁ = 0.4) by ± 0.1 altered permeability estimates by ± 15%, emphasizing the need for core-based calibration.However, excluding T₂ geometric mean (SHAP = 0.47) caused a 12% drop in accuracy.Class imbalance: Synthetic minority oversampling (SMOTE) was applied to address underrepresented microporous facies (7% of data). This improved recall from 0.82 to 0.89 without compromising precision (Δ < 2%).Hyperparameter perturbation: Varying RF tree depth (5–20 levels) and SVM kernel width (γ = 0.01–1.0) showed AUC fluctuations < 3%, indicating robust default parameterization. Clay content variability: Increasing Vclay from 10% to 30% reduced permeability predictions by 22% in carbonates but only 9% in sandstones, highlighting lithology-dependent sensitivity. Clay content variability: Increasing Vclay from 10% to 30% reduced permeability predictions by 22% in carbonates but only 9% in sandstones, highlighting lithology-dependent sensitivity. Fluid saturation effects: At Sw = 40%, the model overpredicted permeability by 8% compared to core data; this error diminished to < 2% when Sw ≤ 25%. Fluid saturation effects: At Sw = 40%, the model overpredicted permeability by 8% compared to core data; this error diminished to < 2% when Sw ≤ 25%. Empirical constant calibration: Adjusting the carbonate constant (C₁ = 0.4) by ± 0.1 altered permeability estimates by ± 15%, emphasizing the need for core-based calibration. Empirical constant calibration: Adjusting the carbonate constant (C₁ = 0.4) by ± 0.1 altered permeability estimates by ± 15%, emphasizing the need for core-based calibration. However, excluding T₂ geometric mean (SHAP = 0.47) caused a 12% drop in accuracy. However, excluding T₂ geometric mean (SHAP = 0.47) caused a 12% drop in accuracy. Class imbalance: Synthetic minority oversampling (SMOTE) was applied to address underrepresented microporous facies (7% of data). This improved recall from 0.82 to 0.89 without compromising precision (Δ < 2%). Class imbalance: Synthetic minority oversampling (SMOTE) was applied to address underrepresented microporous facies (7% of data). This improved recall from 0.82 to 0.89 without compromising precision (Δ < 2%). Hyperparameter perturbation: Varying RF tree depth (5–20 levels) and SVM kernel width (γ = 0.01–1.0) showed AUC fluctuations < 3%, indicating robust default parameterization. Hyperparameter perturbation: Varying RF tree depth (5–20 levels) and SVM kernel width (γ = 0.01–1.0) showed AUC fluctuations < 3%, indicating robust default parameterization. Key observations from sensitivity analysis. The PCI is most sensitive to T₂ cutoff selection but remains reliable within ± 10% variation.ML models are robust to feature reduction but critically depend on T₂ geometric mean.Permeability predictions require clay content and Sw constraints, particularly in carbonate intervals. The PCI is most sensitive to T₂ cutoff selection but remains reliable within ± 10% variation. The PCI is most sensitive to T₂ cutoff selection but remains reliable within ± 10% variation. ML models are robust to feature reduction but critically depend on T₂ geometric mean. ML models are robust to feature reduction but critically depend on T₂ geometric mean. Permeability predictions require clay content and Sw constraints, particularly in carbonate intervals. Permeability predictions require clay content and Sw constraints, particularly in carbonate intervals. Discussion Reservoir characterization in heterogeneous formations like the Asmari requires moving beyond conventional lithology-based approaches94,106. The NMR-electrofacies integration addresses these challenges through direct pore-scale characterization, establishing a dynamic rock typing framework grounded in pore functionality rather than mineral composition. Pore type-based reservoir characterization The integrated petrophysical visualization (Fig. 12) redefines reservoir evaluation by prioritizing fluid response and pore accessibility over compositional labels. Electrofacies classification via T₂ relaxation dynamics departs from traditional gamma-ray or lithology-based schemes, categorizing intervals as flow-conductive, storage-dominated, or non-contributing. For instance, high SBFV and low T₂LM in deeper sections signify shifts to microporous domains, likely from compaction or dolomitization. This approach directly links pore geometry to permeability trends and saturation-height functions, enhancing predictive capacity for reservoir engineering. Figure 13 extends this paradigm by correlating mineralogy with fluid behavior. Calcite-rich layers with SWE-SWI convergence denote optimal connectivity, whereas clay-anhydrite associations signal capillary barriers. Implications for heterogeneous reservoir modeling The dual-log analysis in Fig. 14 resolves ambiguities in hydrocarbon distribution, revealing that high porosity does not universally equate to producibility. Sandstone-rich intervals emerge as critical flow units within the Asmari’s carbonate-clastic heterogeneity, underscoring the need for lithology-sensitive saturation models. The convergence of CMR and conventional logs validates NMR-derived pore-scale insights, enabling high-resolution electrofacies mapping. This methodology advances reservoir characterization by unifying static properties (porosity, mineralogy) with dynamic fluid behavior. It demonstrates that pore architecture, shaped by depositional and diagenetic processes, ultimately governs reservoir performance—a principle applicable to other complex carbonate systems. Pore type-based rock typing: a paradigm shift The hybrid NMR-electrofacies framework challenges conventional lithology-driven models by prioritizing pore connectivity over mineral composition. Intervals with comparable porosity (e.g., 3540–3560 m vs. 3738–3748 m) exhibited marked divergence in hydrocarbon occupancy, underscoring the limitations of lithofacies models. Sandstone subunits, though comprising only 15% of thickness, emerged as primary flow units due to macroporous networks (PCI > 1.8)—a finding later validated by production correlations (Fig. 19; Supplementary Table 6) and prior studies31,32,35. NMR-electrofacies integration: resolving heterogeneity The integration of NMR T₂ distributions with electrofacies classification resolved vertical heterogeneity more effectively than conventional logs. For instance, bimodal T₂ distributions in mid-depth intervals (e.g., 3590–3709 m) identified dual-porosity systems (micro- and macro-pores), while diffuse spectra in clay-rich zones confirmed capillary-bound fluids. This echoes Kadkhodaie, et al17., who linked T₂ signatures to dynamic flow units. Discrepancies between Ktim (5.1 mD) and KSDR(87 mD) permeability estimates highlight the need for model calibration to pore-throat geometry, as noted by Al-Sudani, et al94.. Stratigraphic controls and diagenetic impacts The preferential hydrocarbon accumulation in sandstone intervals (3540–3625 m) fundamentally challenges the traditional carbonate-dominated reservoir model of the Asmari Formation. Three key stratigraphic factors explain this phenomenon: Depositional architecture Sandstone bodies within the Asmari represent tidally influenced deltaic incursions during Oligo-Miocene sea-level lowstands32,33. Their sheet-like geometries (3–5 m thick, 200–500 m lateral continuity based on well correlations) create interconnected high-permeability pathways (avg. 54.6 mD vs. 9.1 mD in carbonates). Diagenetic modification Carbonates: Early calcite cementation reduced primary porosity by 40–60% in grain-supported facies (thin section quantification).Sandstones: Chlorite grain coatings (SEM-EDS confirmed) inhibited quartz overgrowths, preserving avg. 18.2% porosity vs. 14.7% in dolomites.Pore network contrasts: Micro-CT analysis reveals sandstone macroporosity (30–100 μm pore throats) provides 3× higher flow efficiency than moldic pores in carbonates (T₂ cutoff > 100 ms). Carbonates: Early calcite cementation reduced primary porosity by 40–60% in grain-supported facies (thin section quantification). Carbonates: Early calcite cementation reduced primary porosity by 40–60% in grain-supported facies (thin section quantification). Sandstones: Chlorite grain coatings (SEM-EDS confirmed) inhibited quartz overgrowths, preserving avg. 18.2% porosity vs. 14.7% in dolomites. Sandstones: Chlorite grain coatings (SEM-EDS confirmed) inhibited quartz overgrowths, preserving avg. 18.2% porosity vs. 14.7% in dolomites. Pore network contrasts: Micro-CT analysis reveals sandstone macroporosity (30–100 μm pore throats) provides 3× higher flow efficiency than moldic pores in carbonates (T₂ cutoff > 100 ms). Pore network contrasts: Micro-CT analysis reveals sandstone macroporosity (30–100 μm pore throats) provides 3× higher flow efficiency than moldic pores in carbonates (T₂ cutoff > 100 ms). The dolomitization signature (δ¹⁸O = − 2.5‰ to − 4.0‰ VPDB; fluid inclusion Th = 60–80 °C) indicates mesogenetic alteration that preferentially enhanced porosity in bioturbated wackestones but occluded fractures in packstones. This dual diagenetic effect explains the 37% permeability anisotropy observed in core plug measurements (n = 32). These findings align with global analogues (e.g., Arab-D Formation, Saudi Arabia) where sandstone interbeds contribute > 50% of flow in mixed systems. The critical implication is that reservoir performance correlates with preserved depositional porosity in sandstones and diagenetically enhanced connectivity in carbonates—not bulk lithology. This necessitates revising the Asmari play concept to incorporate: Stratigraphic trapping potential at carbonate-sandstone transitions.Diagenetic sweet spots where early dolomitization (10–20% replacement) enhanced permeability without complete pore occlusion. Stratigraphic trapping potential at carbonate-sandstone transitions. Stratigraphic trapping potential at carbonate-sandstone transitions. Diagenetic sweet spots where early dolomitization (10–20% replacement) enhanced permeability without complete pore occlusion. Diagenetic sweet spots where early dolomitization (10–20% replacement) enhanced permeability without complete pore occlusion. Production data validate this model: PLT logs show 72% of flow originates from sandstone-rich FU1 (Fig. 16), despite comprising only 15–20% of gross thickness. This matches NMR-derived PCI values > 1.8 in these intervals, confirming their conduit role. Synergy between ML and NMR-electrofacies analysis The synergy between ML and NMR-electrofacies analysis introduces a paradigm shift in reservoir characterization. By leveraging RF and SVM, this study transcends subjective lithological labels, instead classifying rocks based on measurable pore functionality. SHAP analysis further illuminates the geological relevance of ML predictions, ensuring that the model’s decisions align with petrophysical principles. This dual approach—combining advanced ML with NMR diagnostics—redefines reservoir evaluation as a dynamic, data-rich science, offering a scalable template for heterogeneous carbonate systems worldwide. The novelty lies in the methodological integration and also in its capacity to reconcile machine learning with geological realism, setting a new standard for predictive reservoir modeling. Implications for reservoir development The framework’s predictive capacity was demonstrated by correlating SBFV (2.2%) and T₂LM values with production data, enabling targeted stimulation of high-FFI (Free Fluid Index) zones. This approach mitigates risks in mature fields, where legacy models often misrepresent dynamic fluid behavior (Fig. 19). The methodology’s transferability to analogous carbonate-clastic systems (e.g., Ghawar Field) is supported by Wang, et al100. and Lin, et al73, who advocated NMR-based pore typing for Middle East reservoirs. Figure 19 illustrates the robust correlation between the pore connectivity index (PCI) and production rates, color-coded by lithology. Sandstone intervals (blue) exhibit a stronger linear trend (R² = 0.82) compared to carbonates (red), reinforcing their dominance as flow conduits despite their minority presence in the Asmari Formation. This visualization underscores PCI’s practical value as a predictive tool for reservoir performance, bridging pore-scale petrophysics with field-scale production behavior. Fig. 19PCI vs. Production Rate Correlation. Scatter plot of PCI against production rates, color-coded by lithology. Sandstone intervals (blue) exhibit stronger linear trends (R²=0.82) than carbonates (red). PCI vs. Production Rate Correlation. Scatter plot of PCI against production rates, color-coded by lithology. Sandstone intervals (blue) exhibit stronger linear trends (R²=0.82) than carbonates (red). Comparative results with prior works The novelty of this study is best contextualized by comparing its methodology and findings with prior research (Table 9). While Varkouhi, et al42. and Eftekhari, et al43. advanced fluid typing and electrofacies classification, respectively, their approaches relied heavily on static lithological descriptors. In contrast, this work redefines rock typing through dynamic pore connectivity metrics (e.g., T₂LM, BFV) and ML-aided interpretability, offering a more predictive and adaptable framework. For example, Zhang, et al74. achieved high accuracy in pore size prediction but did not link these results to reservoir-scale flow units. Here, the integration of SVM-based electrofacies and SHAP explanations bridges this gap, correlating pore geometry with production potential. Similarly, while Jamshidi Gohari, et al110. and Murtaza, et al111. used SHAP for facies interpretation, their study lacked NMR validation—a limitation overcome in this work through core-calibrated T₂ distributions (Table 9). This multi-parameter integration advances upon techniques like the joint inversion of spontaneous potential and resistivity for saturation assessment112, and frequency-domain electrical dispersion methods73,74,113. The NMR-ML framework directly quantifies pore-scale fluid mobility, thereby overcoming the limitations posed by complex wettability16 and static lithological models45,55. Key advancements include: Lithology-independent rock typing: Prior works (e.g43,87,114.,) treated lithology as a primary classifier, whereas this study uses pore-fluid dynamics.Explainable AI for reservoir decisions: SHAP analysis quantifies the impact of pore structure on flow, a step beyond black-box ML models in earlier studies.Economic implications: By identifying sandstone-rich zones as flow “sweet spots,” this work challenges regional paradigms (e.g31,115.,) that prioritized carbonate facies. Lithology-independent rock typing: Prior works (e.g43,87,114.,) treated lithology as a primary classifier, whereas this study uses pore-fluid dynamics. Lithology-independent rock typing: Prior works (e.g43,87,114.,) treated lithology as a primary classifier, whereas this study uses pore-fluid dynamics. Explainable AI for reservoir decisions: SHAP analysis quantifies the impact of pore structure on flow, a step beyond black-box ML models in earlier studies. Explainable AI for reservoir decisions: SHAP analysis quantifies the impact of pore structure on flow, a step beyond black-box ML models in earlier studies. Economic implications: By identifying sandstone-rich zones as flow “sweet spots,” this work challenges regional paradigms (e.g31,115.,) that prioritized carbonate facies. Economic implications: By identifying sandstone-rich zones as flow “sweet spots,” this work challenges regional paradigms (e.g31,115.,) that prioritized carbonate facies. Table 9Comparative analysis of this study with prior works.AspectPrior WorksThis StudyNovelty Rock Typing Basis Lithology-dominated10,43,104Pore type-based (NMR T₂ spectra + ML)Dynamic classification based on fluid mobility, not static composition. ML Integration Limited to prediction9,74 or facies clustering44,116RF/SVM for pore typing + SHAP for interpretabilityTransparent, physics-informed ML decisions. Fluid Dynamics Focused on pore water typing42,66,91Quantifies bound/free fluid interplay via BFV and FFVLinks pore-scale fluids to reservoir-scale flow. Validation Core/log correlations33,117,118Core-NMR-ML tri-validation (R² > 0.85)Higher fidelity in heterogeneous zones. Economic Impact Carbonate-centric31,32,41Identifies sandstone subunits as key contributorsRedefines development strategies in mixed systems. Table 9Comparative analysis of this study with prior works.AspectPrior WorksThis StudyNovelty Rock Typing Basis Lithology-dominated10,43,104Pore type-based (NMR T₂ spectra + ML)Dynamic classification based on fluid mobility, not static composition. ML Integration Limited to prediction9,74 or facies clustering44,116RF/SVM for pore typing + SHAP for interpretabilityTransparent, physics-informed ML decisions. Fluid Dynamics Focused on pore water typing42,66,91Quantifies bound/free fluid interplay via BFV and FFVLinks pore-scale fluids to reservoir-scale flow. Validation Core/log correlations33,117,118Core-NMR-ML tri-validation (R² > 0.85)Higher fidelity in heterogeneous zones. Economic Impact Carbonate-centric31,32,41Identifies sandstone subunits as key contributorsRedefines development strategies in mixed systems. Comparative analysis of this study with prior works. This study advances reservoir characterization by unifying NMR pore physics, machine learning, and explainable AI—a triad absent in prior literature. By addressing the gaps in dynamic pore-fluid modeling and scalable ML integration, it provides a replicable framework for heterogeneous reservoirs worldwide. The comparative analysis underscores its transformative potential, shifting the focus from “what rocks are” to “how rocks flow.” Future work could extend this approach to real-time drilling applications or CO₂ storage assessments, further solidifying its role in next-gen petrophysics. Limitations and future work While NMR data resolved microporosity, core calibration remains critical for validating clay-bound fluid volumes. Future studies could incorporate micro-CT imaging to refine pore-network models. Additionally, the framework’s reliance on Schlumberger’s CMR tool necessitates adaptation for other NMR tools in legacy datasets. Advantages Pore type-based approach: The integration of NMR-derived pore architecture with electrofacies classification provides a more accurate representation of fluid flow dynamics compared to traditional lithology-based methods. This enhances reservoir characterization by directly linking pore-scale properties to reservoir performance.Machine learning validation: The use of RF, SVM, and SHAP analysis ensures robust and interpretable classification of pore systems, improving predictive accuracy and reducing subjectivity in rock typing.Practical implications: The identification of sandstone-rich intervals as key flow units challenges conventional models and offers actionable insights for optimizing hydrocarbon recovery in carbonate-clastic reservoirs.Scalability: The proposed framework is transferable to other heterogeneous reservoirs, particularly in mature fields where dynamic pore behavior governs productivity. Pore type-based approach: The integration of NMR-derived pore architecture with electrofacies classification provides a more accurate representation of fluid flow dynamics compared to traditional lithology-based methods. This enhances reservoir characterization by directly linking pore-scale properties to reservoir performance. Pore type-based approach: The integration of NMR-derived pore architecture with electrofacies classification provides a more accurate representation of fluid flow dynamics compared to traditional lithology-based methods. This enhances reservoir characterization by directly linking pore-scale properties to reservoir performance. Machine learning validation: The use of RF, SVM, and SHAP analysis ensures robust and interpretable classification of pore systems, improving predictive accuracy and reducing subjectivity in rock typing. Machine learning validation: The use of RF, SVM, and SHAP analysis ensures robust and interpretable classification of pore systems, improving predictive accuracy and reducing subjectivity in rock typing. Practical implications: The identification of sandstone-rich intervals as key flow units challenges conventional models and offers actionable insights for optimizing hydrocarbon recovery in carbonate-clastic reservoirs. Practical implications: The identification of sandstone-rich intervals as key flow units challenges conventional models and offers actionable insights for optimizing hydrocarbon recovery in carbonate-clastic reservoirs. Scalability: The proposed framework is transferable to other heterogeneous reservoirs, particularly in mature fields where dynamic pore behavior governs productivity. Scalability: The proposed framework is transferable to other heterogeneous reservoirs, particularly in mature fields where dynamic pore behavior governs productivity. Limitations Data dependency: The accuracy of the model relies heavily on the quality and resolution of NMR and core data, which may not always be available in older or less-explored fields. Computational complexity: The integration of machine learning and SHAP analysis requires significant computational resources, which may limit real-time applications in some operational settings. Generalizability: While effective in the Asmari Formation, the framework may require calibration for reservoirs with markedly different diagenetic histories, fluid properties, or geological settings (e.g., siliciclastic-dominated or complex evaporite systems)3,4,69,119,120. This underscores a common challenge in subsurface characterization, where methodologies often need significant adaptation when applied to systems with different fundamental controls, such as the transition from hydrocarbon carbonate reservoirs to complex evaporitic playa aquifers26,121. Core-log integration: Discrepancies between core measurements and log-derived parameters, such as permeability, can introduce uncertainties in model validation (Table 10). Data dependency: The accuracy of the model relies heavily on the quality and resolution of NMR and core data, which may not always be available in older or less-explored fields. Data dependency: The accuracy of the model relies heavily on the quality and resolution of NMR and core data, which may not always be available in older or less-explored fields. Computational complexity: The integration of machine learning and SHAP analysis requires significant computational resources, which may limit real-time applications in some operational settings. Computational complexity: The integration of machine learning and SHAP analysis requires significant computational resources, which may limit real-time applications in some operational settings. Generalizability: While effective in the Asmari Formation, the framework may require calibration for reservoirs with markedly different diagenetic histories, fluid properties, or geological settings (e.g., siliciclastic-dominated or complex evaporite systems)3,4,69,119,120. This underscores a common challenge in subsurface characterization, where methodologies often need significant adaptation when applied to systems with different fundamental controls, such as the transition from hydrocarbon carbonate reservoirs to complex evaporitic playa aquifers26,121. Generalizability: While effective in the Asmari Formation, the framework may require calibration for reservoirs with markedly different diagenetic histories, fluid properties, or geological settings (e.g., siliciclastic-dominated or complex evaporite systems)3,4,69,119,120. This underscores a common challenge in subsurface characterization, where methodologies often need significant adaptation when applied to systems with different fundamental controls, such as the transition from hydrocarbon carbonate reservoirs to complex evaporitic playa aquifers26,121. Core-log integration: Discrepancies between core measurements and log-derived parameters, such as permeability, can introduce uncertainties in model validation (Table 10). Core-log integration: Discrepancies between core measurements and log-derived parameters, such as permeability, can introduce uncertainties in model validation (Table 10). Table 10Limitations vs. prior studies.AspectPrior StudiesThis StudyRock TypingLithology-based8,19Pore functionality (PCI + ML)PermeabilityStatic models17Dynamic clay/Sw correctionsValidationCore-only32Core-NMR-ML tri-validation Table 10Limitations vs. prior studies.AspectPrior StudiesThis StudyRock TypingLithology-based8,19Pore functionality (PCI + ML)PermeabilityStatic models17Dynamic clay/Sw correctionsValidationCore-only32Core-NMR-ML tri-validation Limitations vs. prior studies. Future directions should focus on Real-time applications: Developing streamlined workflows for real-time NMR and ML integration to support drilling and completion decisions. Expanded geological contexts: Testing the framework in diverse carbonate-clastic systems to refine its universal applicability. Advanced fluid typing: Incorporating multi-dimensional NMR and geochemical data to enhance fluid saturation and wettability predictions16,112. AI optimization: Exploring lightweight ML models and edge computing to improve computational efficiency for field-scale deployment. Real-time applications: Developing streamlined workflows for real-time NMR and ML integration to support drilling and completion decisions. Real-time applications: Developing streamlined workflows for real-time NMR and ML integration to support drilling and completion decisions. Expanded geological contexts: Testing the framework in diverse carbonate-clastic systems to refine its universal applicability. Expanded geological contexts: Testing the framework in diverse carbonate-clastic systems to refine its universal applicability. Advanced fluid typing: Incorporating multi-dimensional NMR and geochemical data to enhance fluid saturation and wettability predictions16,112. Advanced fluid typing: Incorporating multi-dimensional NMR and geochemical data to enhance fluid saturation and wettability predictions16,112. AI optimization: Exploring lightweight ML models and edge computing to improve computational efficiency for field-scale deployment. AI optimization: Exploring lightweight ML models and edge computing to improve computational efficiency for field-scale deployment. Conclusions and recommendations This study established that pore architecture, rather than lithology alone, governs fluid flow behavior in the heterogeneous Asmari Formation. The hybrid NMR-machine learning framework successfully quantified three distinct pore systems (micro-, meso-, and macroporous) and demonstrated that sandstone intervals, despite comprising only 15% of the reservoir thickness, contribute 58% of movable hydrocarbons and 73% of total flow capacity due to superior pore connectivity (PCI > 1.8). The explainable AI approach achieved 89% accuracy in electrofacies classification, with SHAP analysis confirming T₂ geometric mean (42% contribution) and bound fluid volume (−0.38 correlation) as dominant controls on permeability distribution. These findings fundamentally challenge traditional carbonate-centric reservoir models and provide quantitative evidence that pore-scale characterization should supersede lithological classification in heterogeneous systems. This study introduces three fundamental advances in carbonate reservoir characterization: (1) a NMR-derived pore connectivity index (PCI) that quantifies flow potential independent of lithology, (2) the first integration of SHAP-interpretable machine learning with NMR petrophysics for electrofacies classification, and (3) dynamic permeability corrections that reduce prediction errors by 37% in clay-rich zones. The workflow’s validation against core and production data demonstrates unprecedented accuracy, with 89% agreement in flow unit identification and < 0.32 log unit permeability errors. Most significantly, the results overturn conventional models by proving sandstone subunits (15% of thickness) dominate production (58% of hydrocarbons) due to superior pore connectivity - a finding with immediate implications for redevelopment strategies in mature fields. Key findings NMR T₂ skewness analysis revealed sandstone intervals have 3.2× higher PCI values than carbonates.Machine learning achieved 91.2% pore-type prediction accuracy using only 9 optimized features.The hybrid KSDR-Ktim permeability model improved predictions by 28% over conventional approaches.SHAP analysis proved T₂ geometric mean (47% contribution) outweighs lithology in flow control.84% of machine-classified flow units matched production logging results. NMR T₂ skewness analysis revealed sandstone intervals have 3.2× higher PCI values than carbonates. NMR T₂ skewness analysis revealed sandstone intervals have 3.2× higher PCI values than carbonates. Machine learning achieved 91.2% pore-type prediction accuracy using only 9 optimized features. Machine learning achieved 91.2% pore-type prediction accuracy using only 9 optimized features. The hybrid KSDR-Ktim permeability model improved predictions by 28% over conventional approaches. The hybrid KSDR-Ktim permeability model improved predictions by 28% over conventional approaches. SHAP analysis proved T₂ geometric mean (47% contribution) outweighs lithology in flow control. SHAP analysis proved T₂ geometric mean (47% contribution) outweighs lithology in flow control. 84% of machine-classified flow units matched production logging results. 84% of machine-classified flow units matched production logging results. Three key recommendations emerge for operational implementation: First, NMR logging should be prioritized over conventional porosity logs in development wells, as demonstrated by the 88% agreement between NMR-derived permeability and core measurements. Second, completion strategies should target sandstone-rich intervals (EF3 electrofacies) identified through the machine learning workflow, which showed 84% concordance with production logging results. Third, the dynamic permeability correction for clay-bound fluids should be incorporated into reservoir simulations, having reduced permeability overprediction in shaly zones by 37%. These measures collectively address the critical gap between static geological models and dynamic flow behavior in mature carbonate-clastic reservoirs. Recommendations for implementation Prioritize NMR logging in development wells for pore-scale characterization.Target completions in EF3 electrofacies (sandstone-rich zones).Implement dynamic clay corrections in reservoir simulations.Expand real-time ML applications using drilling data streams. Prioritize NMR logging in development wells for pore-scale characterization. Prioritize NMR logging in development wells for pore-scale characterization. Target completions in EF3 electrofacies (sandstone-rich zones). Target completions in EF3 electrofacies (sandstone-rich zones). Implement dynamic clay corrections in reservoir simulations. Implement dynamic clay corrections in reservoir simulations. Expand real-time ML applications using drilling data streams. Expand real-time ML applications using drilling data streams. The methodology establishes a new paradigm for reservoir evaluation, shifting focus from static lithology to dynamic pore functionality. Its successful application in the Asmari Formation provides a template for heterogeneous carbonate-clastic systems worldwide, particularly where traditional models underperform. Future work should validate the approach in other stratigraphic contexts while developing automated implementations for field-scale deployment. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1