The impact of physical and pore-structural characteristics on coral reef limestone strength: insights from interpretable ensemble learning models

• Proposed Stacking-XRB, an interpretable ensemble framework, for predicting coral reef limestone UCS with small, highly dispersed datasets. • SHapley Additive exPlanations (SHAP) and Individual Conditional Expectation (ICE) were adopted to quantify density-dependent nonlinear interaction between density and Anisotropy coefficient. • Identified a critical threshold (Anisotropy coefficient is −2.25) demarcating pore structure transition, defining the ±20 % error applicability domain. • Revealed that transitional pore structures (medium-strength CRL) dominate local prediction distortions. Coral reef limestone (CRL) serves as a critical foundation material for reef and island engineering, whose strength is controlled by physical properties and pore structure. However, the interaction influence of these characteristics on CRL strength remains insufficiently understood. Thus, confronting severe data limitations in current research, this study established a databset through original laboratory tests and computed tomography (CT) scans. This dataset includes dry density ( ρ d ), porosity, P-wave velocity, isotropic component of the pore tensor, anisotropy coefficient ( lo g 10 ( Γ ) ) and fractal dimension. Subsequently, an optimal ensemble learning model, Stacking-XRB combining extreme gradient boosting (XGBoost), random forest (RF) and Back propagation neural network (BPNN), was developed for predicting CRL strength, achieving high accuracy (R 2 = 0.93 on the test set). Furthermore, interpretable machine learning was employed to analyze the influence of physical properties and pore morphology on strength. SHapley Additive exPlanations (SHAP) and Individual Conditional Expectation (ICE) result reveal a nonlinear interaction between ρ d and lo g 10 ( Γ ) , which quantifies the competition between rock compaction and pore-structure in effecting UCS. The threshold ( lo g 10 ( Γ ) ≈ -2.25) of model applicability was determined. CT-based analysis confirms that this value demarcates a structural transition from nearly isotropic to oriented pore channels, which correlates with a decrease in strength. This work advances rock mechanics by quantitatively interpreting the interacting control mechanism of physical properties and pore structure, offering meso‑structural insights into the pore-scale mechanisms that controls the strength of CRL.