Three-dimensional numerical simulation of progressive rock caving using a tension-driven iterative element-removal method Authors: Jingwen Liu, Jiangyong Pu, Qinglei Yu, Xin Wang Published: 2026 ABSTRACT Caving mining is widely used for deep ore extraction because of its high productivity and cost efficiency. However, accurately predicting the three-dimensional (3D) caving morphology and collapse height remains challenging for safe mine design. Classical arch theories are limited to 2D plane-strain conditions, while conventional numerical methods are either computationally expensive or unable to capture the progressive nature of 3D rock caving. In this study, a 3D continuum-based numerical framework is developed to simulate rock caving by defining tension-induced detachment of plasticized rock mass as the primary failure mechanism. Progressive collapse is represented through an iterative element-removal scheme implemented in FLAC3D. The proposed method is first verified against classical caving arch theories for different undercut spans, showing good agreement with the self-stable arch theory. It is then applied to the 37–2# orebody of the Anhui Taiping Mine to simulate the evolution of the caving zone and quantify its height, volume, and spatial morphology during staged mining. The results show that the caving zone expands upward progressively and develops an arch-like geometry in cross-section. Sensitivity analyses indicate that the internal friction angle has the strongest influence on caving height and volume, as it controls shear resistance, stress arching, and tensile stress redistribution at the caving boundary. The proposed approach provides an efficient and mechanically interpretable tool for evaluating caving behavior and supporting risk assessment and backfill design in underground mining. Introduction Caving mining is a high-efficiency underground extraction method that uses blasting, in situ stress, and gravity to promote progressive rock failure and fragmentation1,2. Once caving initiates, the rock mass progressively fails when induced stresses reach to the rock mass strength and fractured rock masses unravel under gravity into the undercut and draw points3. Because of its high productivity and low cost, it has become a major approach for extracting deep ore bodies. As mining depth increases, the hanging wall collapses gradually into the goaf, forming a complex caved zone above the extracted orebody4,5. A clear understanding of rock caving mechanisms is therefore essential for mine safety. This is especially important in block caving, which depends on continuous ore draw and stable cave propagation, and in sublevel caving, where uncontrolled voids may threaten mine infrastructure6,7. To describe the geometry of caving zones, many analytical models have been proposed. Protodyakonov’s equilibrium arch theory (PEAT) is a classical method for estimating rock pressure around underground openings. It represents the caving arch as a parabolic profile and assumes that the ground behaves as a loosened medium8. The natural equilibrium arch theory (NEAT) assumes that no shear stress or bending moment acts on the caving-zone boundary, and it predicts an elliptical arch profile9,10. NEAT has also been extended to consider gradients in the horizontal component of in situ stress11. Based on NEAT, the self-stable arch theory (SSAT) assumes that rock in a limit-equilibrium state collapses from tensile regions toward the boundary where tensile stress vanishes12. The transition boundary between tensile and compressive stress is approximately elliptical13. The main assumptions and governing equations of these three theories are summarized in Table 1. However, these methods are limited to two-dimensional plane-strain conditions and are not well suited to complex three-dimensional problems. Recent experimental and numerical studies have further revealed that stress arching in caving mines creates a stress depletion zone associated with a dilation front, which governs cavity propagation toward the surface14. Table 1Summary of three main caving arch theories.TheoryDiagramEquation of caving zoneHeight of caving zone(a)8 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{y=}\frac{{\mathrm{x}}^{\mathrm{2}}}{\mathrm{fa}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{h=}\frac{\mathrm{a}}{\mathrm{f}}$$\end{document} (b)9 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{\mathrm{y}}^{\mathrm{2}}}{{\mathrm{b}}_{\mathrm{2}}^{\mathrm{2}}}\mathrm{+}\frac{{\mathrm{x}}^{\mathrm{2}}}{\lambda{\mathrm{b}}_{\mathrm{2}}^{\mathrm{2}}}\mathrm{=}\mathrm{1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{\mathrm{a}}_{\mathrm{2}}}{{\mathrm{b}}_{\mathrm{2}}}\mathrm{=}\sqrt{\lambda}$$\end{document} (c)11 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{R}}_{\mathrm{y}}\mathrm{-}\frac{\gamma\mathrm{H}{\mathrm{x}}^{\mathrm{2}}}{\mathrm{2}}\mathrm{-(}\frac{\lambda\gamma\mathrm{H}{\mathrm{y}}^{\mathrm{2}}}{\mathrm{2}}\mathrm{+}\frac{\lambda{\mathrm{y}}^{\mathrm{2}}}{\mathrm{6}}\mathrm{)=}\mathrm{0}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{l=}\sqrt{\lambda{\mathrm{h}}^{\mathrm{2}}\mathrm{+}\frac{\mathrm{2}\lambda{\mathrm{h}}^{\mathrm{3}}}{\mathrm{3}\mathrm{H}}}$$\end{document} (d)12 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{\mathrm{x}}^{\mathrm{2}}}{{\left(\frac{{\mathrm{W}}_{\mathrm{0}}}{\mathrm{2}}\right)}^{\mathrm{2}}}\mathrm{+}\frac{{\mathrm{y}}^{\mathrm{2}}}{\left(\frac{\mathrm{3}{\mathrm{P}}_{\mathrm{0}}{\mathrm{W}}_{\mathrm{0}}^{\mathrm{2}}}{\mathrm{A}{\mathrm{P}}_{\mathrm{0}}\mathrm{+}\mathrm{12}{{\sigma}}_{\mathrm{i}}}\right)}\mathrm{=}\mathrm{1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{h=W}\sqrt{\frac{\mathrm{3}{\mathrm{P}}_{\mathrm{0}}}{\mathrm{4}{\mathrm{P}}_{\mathrm{0}}\mathrm{+}\mathrm{12}{{\sigma}}_{\mathrm{t}}}}$$\end{document} Table 1Summary of three main caving arch theories.TheoryDiagramEquation of caving zoneHeight of caving zone(a)8 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{y=}\frac{{\mathrm{x}}^{\mathrm{2}}}{\mathrm{fa}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{h=}\frac{\mathrm{a}}{\mathrm{f}}$$\end{document} (b)9 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{\mathrm{y}}^{\mathrm{2}}}{{\mathrm{b}}_{\mathrm{2}}^{\mathrm{2}}}\mathrm{+}\frac{{\mathrm{x}}^{\mathrm{2}}}{\lambda{\mathrm{b}}_{\mathrm{2}}^{\mathrm{2}}}\mathrm{=}\mathrm{1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{\mathrm{a}}_{\mathrm{2}}}{{\mathrm{b}}_{\mathrm{2}}}\mathrm{=}\sqrt{\lambda}$$\end{document} (c)11 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{R}}_{\mathrm{y}}\mathrm{-}\frac{\gamma\mathrm{H}{\mathrm{x}}^{\mathrm{2}}}{\mathrm{2}}\mathrm{-(}\frac{\lambda\gamma\mathrm{H}{\mathrm{y}}^{\mathrm{2}}}{\mathrm{2}}\mathrm{+}\frac{\lambda{\mathrm{y}}^{\mathrm{2}}}{\mathrm{6}}\mathrm{)=}\mathrm{0}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{l=}\sqrt{\lambda{\mathrm{h}}^{\mathrm{2}}\mathrm{+}\frac{\mathrm{2}\lambda{\mathrm{h}}^{\mathrm{3}}}{\mathrm{3}\mathrm{H}}}$$\end{document} (d)12 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{\mathrm{x}}^{\mathrm{2}}}{{\left(\frac{{\mathrm{W}}_{\mathrm{0}}}{\mathrm{2}}\right)}^{\mathrm{2}}}\mathrm{+}\frac{{\mathrm{y}}^{\mathrm{2}}}{\left(\frac{\mathrm{3}{\mathrm{P}}_{\mathrm{0}}{\mathrm{W}}_{\mathrm{0}}^{\mathrm{2}}}{\mathrm{A}{\mathrm{P}}_{\mathrm{0}}\mathrm{+}\mathrm{12}{{\sigma}}_{\mathrm{i}}}\right)}\mathrm{=}\mathrm{1}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{h=W}\sqrt{\frac{\mathrm{3}{\mathrm{P}}_{\mathrm{0}}}{\mathrm{4}{\mathrm{P}}_{\mathrm{0}}\mathrm{+}\mathrm{12}{{\sigma}}_{\mathrm{t}}}}$$\end{document} Summary of three main caving arch theories. With advances in computational mechanics, numerical methods have become essential for solving engineering problems that cannot be handled by closed-form analytical solutions15–17. Rock masses are heterogeneous, discontinuous, and strongly nonlinear, so robust numerical treatment is required. Rock caving involves a transition from continuum deformation to a discontinuum of interacting fragments. This process is naturally related to the Discrete Element Method (DEM), which can explicitly simulate fragmentation and particle interaction18–22. However, DEM results are often affected by the initial discretization. Fractures usually develop along predefined element boundaries. In 3DEC and PFC3D, for example, fragments are commonly represented by polyhedra or spheres23–25. In these models, contact behavior must be defined in advance, and fracture paths follow preset interfaces. In addition, contact parameters are not always measurable in the laboratory and often require trial-and-error calibration, the calibration step frequently represents a major difficulty in large-scale DEM simulations26. At the engineering scale, DEM also becomes very expensive computationally because a large number of elements is needed to represent realistic geometry. The Finite Element Method (FEM)27–29 and the Finite Difference Method (FDM)30,31 are well-established continuum-based numerical frameworks. They are more efficient for large-scale mining problems under current computing conditions. However, they still struggle to capture progressive fracturing and rock detachment. As a result, static plastic zones are often used as simplified proxies for caving regions30. In reality, rock in a caving zone progressively fragments and detaches before falling. The failure first occurs in certain superimposed strata slightly far from the goaf and subsequently propagates both upward and downward as excavation proceeds, forming a large-scale fracture plane that connects with the goaf32. During cyclic caving, new free surfaces cause stress redistribution, which drives further yielding33. This feedback process means that the actual caving morphology can differ greatly from the static plastic zone predicted by conventional continuum analyses34. In this study, a three-dimensional numerical framework with an iterative element-removal scheme was developed in FLAC3D to investigate the progressive failure and caving of hanging wall rock masses. First, two-dimensional simulations were performed for different undercut spans, and the results were compared with classical arch theory to assess the reliability of the model. The framework was then applied to the 37–2# orebody in the Anhui Taiping Mine. This case study was used to quantify caving-zone evolution, volume change, and spatial morphology under site-specific geological and stress conditions. The proposed method provides a practical tool for predicting cave propagation in large-scale mining operations. Numerical model for rock caving process analysis Model description The caving process proceeds through undercutting, gravity-driven progressive rupture of the orebody, and rockfall until a self-supporting arch forms. Accordingly, caving mining is characterized by repeated cycles of undercutting-induced stress redistribution and subsequent structural collapse. This process is governed by the interaction between pre-existing structural discontinuities and newly formed tensile fractures. Although shear failure remains common within the stope boundaries, especially during induced caving, its macroscopic expression is often reduced by the temporary confinement provided by accumulated caved debris in the drawbell. Based on the above analysis, the proposed numerical framework uses a composite constitutive law. The maximum tensile stress criterion is adopted to identify the caving susceptibility of the ore and host rock, while the Mohr–Coulomb criterion is used to describe shear yielding. To capture the progressive transition from a continuous rock mass to fragmented debris, several physically based assumptions are introduced. In particular, elements in the incipient plastic zone at the arch crown are assumed to detach once the tensile failure criterion is satisfied. These elements are then removed from the computational domain through an iterative element-deletion procedure to simulate fragmentation and cavity growth. The method is implemented in FLAC3D, using its continuum-based finite-difference scheme to track the evolution of the caving front. In the FLAC3D model, rock mass failure is governed by a composite yield surface consisting of a Mohr–Coulomb shear envelope and a tension-cutoff limit (see Fig. 1). The envelope is divided into two regimes. The shear segment from Point A to Point B is described by the Mohr–Coulomb criterion in Eq. (1), whereas the tensile segment from Point B to Point C is defined by the tension-cutoff criterion in Eq. (2).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{f}\text{}}^{\mathrm{s}}\mathrm{=}{\sigma}_{\mathrm{1}}\mathrm{-}{\sigma}_{\mathrm{3}}\frac{\mathrm{1+}\mathrm{sin\:(}\\phi\mathrm{)}}{\mathrm{1-}\mathrm{sin\:(}\\phi\mathrm{)}}\mathrm{-2}\mathrm{c}\sqrt{\frac{\mathrm{1+}\mathrm{sin\:(}\\phi\mathrm{)}}{\mathrm{1-}\mathrm{sin\:(}\\phi\mathrm{)}}}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{f}\text{}}^{\mathrm{t}}\mathrm{=}{\sigma}_{\mathrm{1}}\mathrm{-}{\sigma}_{\mathrm{t}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\phi$$\end{document} is the friction angle, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{c}$$\end{document} is the cohesion, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma}_{\mathrm{t}}$$\end{document} is the tensile strength. Fig. 1Constitutive law description of the numerical model. Constitutive law description of the numerical model. In the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma}_{\mathrm{1}}-{\sigma}_{\mathrm{3}}$$\end{document} stress plane (Fig. 1), a diagonal line connecting the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{f}\text{}}^{\mathrm{s}}\mathrm{=0}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{f\:}}^{\mathrm{t}}\mathrm{=0}$$\end{document} representations divides the region where the elastic trial stress violates the composite yield surface into two subregions, hereafter referred to as Domain 1 and Domain 2. If the elastic trial stress state lies within Domain 1, a shear return mapping is applied. If it falls within Domain 2, a tension cutoff correction is performed. Elements in either domain that satisfy the tensile failure condition (σ1 > 0) are identified as the active caving zone. In the proposed numerical scheme, these elements represent rock mass that is ready for detachment, thereby simulating progressive cavity development through iterative removal of failed elements. The implementation of the proposed numerical procedure is illustrated in the flowchart shown in Fig. 2. In this study, the caving front is defined as the outermost boundary of elements that simultaneously satisfy the plastic yielding condition and the tensile failure criterion. During each iteration, only elements that (i) meet the tensile failure condition and (ii) are adjacent to previously removed (null) elements are considered eligible for removal. This definition ensures that the caving process propagates progressively from the excavation boundary toward the surrounding rock mass, rather than occurring as isolated or non-physical element deletions. The process begins by identifying the plastic zones induced in the surrounding rock mass by undercutting. The stress state of elements within these zones is then evaluated. Elements that satisfy the tensile failure criterion and are located along the outer boundary of the caving arch (the caving front) are assumed to detach under gravity. These elements are subsequently assigned a null material model in FLAC3D to represent fully detached and non-load-bearing material. Once converted to null elements, they no longer contribute to stiffness or stress transfer, effectively simulating fragmentation and void formation within the caving zone. This approach avoids the need for explicit contact detection and improves computational efficiency, although it may induce localized stress redistribution during the iterative process. This iterative procedure continues until no additional elements exceed the tension-cutoff limit, at which point the caving front reaches a temporarily stable configuration. By using tensile failure as an indicator of detachment, the method links the efficiency of continuum finite-difference modeling with an explicit representation of progressive cave development. While the proposed numerical model uses tensile failure as the primary criterion for element detachment, shear yielding is still considered via the Mohr–Coulomb criterion to govern plastic zone development. The tensile-driven assumption is most applicable to weak-to-moderately competent rock masses, strongly altered surrounding rocks, and conditions where excavation creates free surfaces and induces tensile stress concentration near the arch crown. For highly jointed or blocky rock masses, where shear along discontinuities dominates, the model may underestimate shear-controlled failure; in such cases, extension with discrete element or hybrid continuum–discrete methods may be necessary. Fig. 2Flow chart of the numerical model. Flow chart of the numerical model. Verification of the numerical model To evaluate the robustness and accuracy of the proposed numerical procedure, a discretized model was developed to simulate caving over a range of undercut spans. This validation stage primarily compares the numerically predicted caving height with theoretical values derived from classical caving arch theory, while also assessing the algorithm’s ability to reproduce progressive failure mechanisms. The computational domain, with dimensions of 400 m in length and 350 m in height, was discretized into a fine mesh comprising 102,811 zones to ensure numerical convergence. For the boundary conditions, horizontal displacements were constrained along the lateral boundaries, and vertical displacement was fixed at the base, corresponding to roller conditions. To represent a 110 m overburden, a uniform vertical stress of 3.96 MPa was applied to the top surface. Undercutting was implemented incrementally along the model length in steps of 5 m in height. To examine span-dependent behavior, five undercut spans of 20 m, 40 m, 60 m, 80 m, and 100 m were systematically simulated and analyzed. For theoretical validation, the rock mass in this section is idealized as a homogeneous and isotropic continuum. This simplification is consistent with the assumptions underlying the analytical solutions and enables a controlled and unbiased comparison between numerical results and theoretical predictions. However, natural rock masses are inherently heterogeneous and often anisotropic due to the presence of joints, bedding planes, and other discontinuities. As such, this assumption may limit the general applicability of the validation results, particularly in structurally complex geological conditions. This simplification is consistent with the assumptions underlying the analytical solutions and enables an unbiased assessment of the numerical algorithm’s performance. The mechanical properties assigned to the idealized rock mass are summarized in Table 2. These values were selected based on typical ranges reported for moderately jointed hard rock masses in standard rock mechanics references, including the engineering rock mass classification guideline GB/T 50,218–2014. They also fall within the range of recent experimental data for skarn and metal ore rocks35,36. The resulting parameter set represents a moderately competent rock mass and provides conservative, idealized conditions favorable for the formation of a clear arch structure. The extension of the method to heterogeneous and site-specific geological conditions is addressed in the subsequent engineering case study. Table 2Mechanical properties of the idealized rock mass used in the verification model.LithologyElastic modulus (GPa)Poisson’s ratioCohesion (MPa)Friction angle (°)Tensile strength (MPa)Density (kg/m3)Ore10.50.252.40360.403600 Table 2Mechanical properties of the idealized rock mass used in the verification model.LithologyElastic modulus (GPa)Poisson’s ratioCohesion (MPa)Friction angle (°)Tensile strength (MPa)Density (kg/m3)Ore10.50.252.40360.403600 Mechanical properties of the idealized rock mass used in the verification model. Figure 3 illustrates the caving process at a 100 m undercut span, showing the evolution of the maximum principal stress and the development of the plastic zones. After completion of the undercutting stage, stress redistribution occurs in the surrounding rock mass (Fig. 3a), resulting in the formation of the initial plastic zones (Fig. 3b). Following the proposed criterion, elements within the plastic zone whose tensile stress exceeded the tension cutoff were removed (Iteration step 1). Element removal altered the geometric and load distribution, so the stress field was recomputed and re-evaluated (Iteration step 2). These removal-recompute iterations were repeated five times. When all elements in the arch crown were in compression, the computation terminated and a stable caved arch was formed (Iteration step 5). The final caved zone (Iteration step 5) differs significantly from the initial plastic zone identified immediately after undercutting (Undercut). Figure 4a shows the final caved zones for different undercut spans. Each exhibits an approximately arch-shaped profile. The caving height increases progressively with increasing undercut span. Figure 4b compares caving heights from the numerical model against predictions from classical arch theories, using the parameters of the verification model. The theoretical curves were obtained from the arch theories introduced above and computed with the mechanical parameters listed for the verification model. The numerical predictions are in good agreement with the SSAT results and are consistently lower than the predictions of the other three classical arch theories. This close agreement arises because the stability criterion used in the numerical scheme defines a caved arch as stable when no tensile stress acts along the arch crown, which is consistent with the stability condition adopted in SSAT. To further interpret these differences, the discrepancies between the numerical results and other classical arch theories (e.g., PEAT and NEAT) can be attributed to differences in their underlying mechanical assumptions. PEAT is based on the assumption of a loosened granular medium, whereas NEAT neglects the contribution of shear stress along the arch boundary. In contrast, the present model incorporates both tensile and shear stress states and determines stability based on a tensile-failure criterion, which is consistent with the assumptions of SSAT. These differences in failure mechanisms and stress boundary conditions lead to variations in the predicted caving geometry. Most classical arch theories were originally developed for relatively small excavation spans, and their predictive accuracy may therefore deteriorate for the larger spans considered in this study. Given the differing applicability ranges of these theoretical models, the deviations between the numerical and analytical results are both expected and reasonable. Despite the limitations of classical analytical solutions, the numerical method demonstrates robust capability in simulating progressive rock caving. Moreover, because classical arch formulations are inherently two-dimensional, they cannot be directly extended to three-dimensional caving problems. The proposed numerical model overcomes this limitation and enables full three-dimensional caving analysis. It should be noted that this verification is based on a homogeneous material idealization. The ability of the proposed method to accommodate spatially variable rock mass properties is demonstrated in the three-dimensional mine-scale application presented below. Fig. 3The caving process at the undercut span of 100 m: (a) maximum principal stress; (b) plastic zone. The caving process at the undercut span of 100 m: (a) maximum principal stress; (b) plastic zone. Fig. 4(a) Caving arch with different undercut spans; (b) caving height comparison between numerical and theoretical results with different undercut spans. (a) Caving arch with different undercut spans; (b) caving height comparison between numerical and theoretical results with different undercut spans. Application of the method in anhui taiping mine Engineering background The Anhui Taiping Mine is located in Huaibei City, Anhui Province, China (Fig. 5). The 37–2# orebody extends from the − 261 m level to the − 305 m level. It is a stratiform orebody dipping to the northeast (Fig. 6a). The orebody exhibits pronounced thickness variations, and both the roof and floor rock masses are highly altered, with extensive skarn alteration, serpentinization, and chloritization37. Owing to the complex geometry of the orebody and the inherent instability of the surrounding rock mass, extraction of the 37–2# deposit presents significant technical challenges. The original mining method for the 37–2# orebody was underhand cut-and-fill mining, with a stope height of approximately 6 m. To date, the 1#, 2#, and 3# sub-orebodies of the 37–2# deposit have been excavated and backfilled with cemented fill. However, the low stability of both the ore and the surrounding rock has resulted in occasional roof and floor collapses, posing substantial risks to ongoing mining operations. Roof failure during late-stage extraction and backfilling has also caused ore–backfill mixing and significant ore loss. In addition, the presence of highly unstable vein zones has further deteriorated mining conditions and at times forced operations to stop, leading to considerable resource waste. To enhance production efficiency and satisfy the mine’s operational targets for thin orebodies: “two highs” (high efficiency, high economic benefit) and “three lows” (low ore loss, low dilution, low accident risk), a revised mining strategy was implemented. Based on the geological and technical conditions of the 37–2# thin orebody, a combined approach of sublevel caving followed by backfilling was adopted. Mining was scheduled at four levels: −277 m, − 285 m, − 290 m, and − 300 m (Fig. 6b). Extraction at the uppermost level follows the mine’s original retreat mining method, with natural or artificially induced caving of the overlying rock after ore recovery. For the lower levels, sublevel caving is employed, whereby ore is extracted beneath the caved overburden. The overburden is critical to sublevel caving and backfilling, and it exerts a strong influence on orebody stability during excavation. To restrict the volume of caved surrounding rock and ensure operational safety and design compliance, the overburden thickness was specified to be at least twice the mining layer height. A roadway network is located above the orebody at approximately − 220 m. Additionally, the Taiping area is underlain by a widespread high-yield Quaternary aquifer and contains numerous surface structures. As a result, excavation carries a risk of large-scale roof and surrounding rock collapse, which could trigger severe safety incidents. Accurate investigation of caving height and extent is therefore essential to guarantee safe and efficient resource recovery. Fig. 5Location of the Anhui Taiping Mine. The map was created using a commercial software: SuperMap GIS software (Version SuperMap iDesktop 9D (2019); https://www.supermap.com/cn/). Location of the Anhui Taiping Mine. The map was created using a commercial software: SuperMap GIS software (Version SuperMap iDesktop 9D (2019); https://www.supermap.com/cn/). Fig. 6The 37 − 2# orebody: (a) orebody morphology (b) stope overview at four levels. The 37 − 2# orebody: (a) orebody morphology (b) stope overview at four levels. Numerical model construction A three-dimensional numerical model was constructed to replicate the geological conditions of the 37–2# orebody (Fig. 7a). To minimize boundary effects, the model domain measures 160 m (length) × 100 m (width) × 170 m (height). The mesh comprises 2,600,166 elements with a minimum element size of 1 m. The upper boundary of the model is set at − 150 m, where the gravity−induced stress is about 4.8 MPa. A vertical gravitational acceleration of 9.8 m/s2 was applied downward. Zero normal displacement (roller) boundary conditions were imposed on the left, right, and bottom boundaries. The surrounding rock consists of highly unstable skarn. To calibrate the model, the mechanical parameters of the skarn rock mass were back‑calculated via an inversion method based on numerical simulations, and the results were validated against in‑situ displacement monitoring and measurements of the surrounding rock failure zone38. The mechanical properties of the 37–2# orebody were assigned based on the mine’s internal geotechnical investigation report39 and are consistent with the typical range reported for skarn-hosted magnetite deposits. The physical and mechanical properties of the orebody and surrounding rock are summarized in Table 3. Numerical simulations were performed according to the designed mining sequence for the 37–2# orebody. The orebody was divided into four mining levels at elevations of − 277 m, − 285 m, − 290 m, and − 300 m, which were extracted sequentially from top to bottom to replicate the underhand mining sequence. Following the extraction of each level, the ore-roof caving process was simulated using the numerical method developed in this study. Mining proceeded to the next lower level once a stable caved arch formed (per the stability criterion outlined above). This procedure was repeated until the entire orebody was excavated. Table 3Mechanical properties of rock mass.LithologyElastic modulus (GPa)Poisson’s ratioCohesion (MPa)Friction angle (°)Tensile strength (MPa)Density (kg/m3)Ore9.260.232.13380.353574Skarn2.10.31.5330.52629 Table 3Mechanical properties of rock mass.LithologyElastic modulus (GPa)Poisson’s ratioCohesion (MPa)Friction angle (°)Tensile strength (MPa)Density (kg/m3)Ore9.260.232.13380.353574Skarn2.10.31.5330.52629 Mechanical properties of rock mass. Results To facilitate a clear presentation of the numerical outcomes, the middle cross-section at the − 277 m level was selected to illustrate the progressive caving behavior of the surrounding rock mass following excavation. The mechanical evolution of the roof strata is visualized through a series of cross-sectional plots (see Fig. 7b). Using the tension-driven iterative element-removal framework established in this study, the redistribution of principal stresses after the extraction of the − 277 m level is depicted in Fig. 8, while the corresponding development of the caving zone across successive computational iterations is shown in Fig. 9. According to the prescribed caving criterion—wherein plasticized rock fails under tensile stress—the numerical model achieved a stable configuration after three removal–recalculation cycles. The final caving geometry exhibited a pronounced conical morphology. Fig. 73D numerical model. (a) 3D numerical model; (b) cut section locations. 3D numerical model. (a) 3D numerical model; (b) cut section locations. Fig. 8Contours of the maximum principal stress of each step after the − 277 m layer is mined out. Contours of the maximum principal stress of each step after the − 277 m layer is mined out. Fig. 9Caving zone of each iteration step after − 277 m layer is mined. Caving zone of each iteration step after − 277 m layer is mined. A distinct stress arch developed within the roof strata during this process: material within the arch crown experienced tensile failure and was subsequently removed from the computational domain, whereas the surrounding load-bearing arch redirected compressive stresses to limit further large-scale collapse. Consequently, a well-defined arch-shaped caving profile emerged in the vertical cross-section. Following the extraction of the uppermost − 277 m level, the weakened roof rock collapsed to a maximum height of 32 m above the excavation horizon, thereby establishing an initial overburden layer that provides cover for subsequent lower-level mining operations. The spatial and volumetric evolution of the caving zone across all mining levels is summarized in Fig. 10. In the central cross-section, the maximum caving heights recorded after mining the − 277 m, − 285 m, − 290 m, and − 300 m levels were 32 m, 37 m, 40 m, and 43 m, respectively. Throughout the downward progression of extraction, the caving zone propagated upward while maintaining an arch-like profile in vertical sections. The ultimate caving crest reached an elevation of approximately − 234 m, corresponding to a total vertical extension of 43 m. The cumulative caving volumes following the extraction of the − 285 m, − 290 m, and − 300 m levels were calculated as 17,243.5 m3, 21,220.6 m3, and 23,545.4 m3, respectively. These volumetric estimates provide a quantitative basis for the design of backfilling operations and the procurement of fill materials in the field. Fig. 10Caving morphology of surrounding rock after ore mining in different sections. Caving morphology of surrounding rock after ore mining in different sections. Following the complete extraction of the 37–2# orebody, displacement contours at the − 240 m and − 220 m levels (Fig. 11) were analyzed. At the − 240 m horizon, the simulation indicates the presence of a caving-affected zone, whereas at the − 220 m level, a larger deformation zone is observed. These results are consistent with the upward propagation of the caving boundary predicted by the model and indicate that the influence of excavation extends to the upper levels through stress redistribution and progressive roof deterioration. Fig. 11The displacement contours after the mining of the orebody. (a) −240 m level; (b) −220 m level. The displacement contours after the mining of the orebody. (a) −240 m level; (b) −220 m level. Discussion The host rock of the 37–2# orebody consists predominantly of skarn, a material known for its heterogeneous diagenetic alteration and consequent spatial variability in geomechanical behavior. In the present study, the caving process is primarily interpreted as a tensile-driven failure phenomenon. This assumption is particularly applicable to rock masses under relatively low confinement and with moderate strength, where stress redistribution after undercutting induces tensile stress concentrations, especially near the arch crown. Under such conditions, tensile cracking governs the detachment of rock mass and drives the progressive upward propagation of the caving zone. However, in highly jointed, blocky, or strongly anisotropic rock masses, failure mechanisms may be dominated by shear sliding along discontinuities or structurally controlled collapse. In these cases, the present model may underestimate shear-driven contributions, and its applicability should be evaluated with caution. To systematically quantify the influence of rock mass strength parameters on post-excavation caving dynamics, a parametric sensitivity analysis was conducted by independently varying cohesion (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{c}$$\end{document}), internal friction angle (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\phi$$\end{document}), and tensile strength (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma}_{\mathrm{t}}$$\end{document}). Using the baseline parameters from the numerical model construction, the caving evolution was simulated for parameter increments and decrements of 20% and 40%. As illustrated in Fig. 12, reductions in these strength components correlate with substantial increases in both caving height and cumulative volume. Among the parameters examined, the internal friction angle emerged as the dominant control on the extent of failure. Within the tested parameter range, the maximum predicted caving height reached 43.4 m, with an associated volume of 31.4 × 103 m3. These findings highlight the necessity of establishing safety buffer zones for critical mine infrastructure located beyond the predicted caving perimeter. Quantitatively, a 40% reduction in the friction angle (to 19.8 yielded a 34.8% increase in caving height and a pronounced 145.3% (an absolute percentage of the baseline volume) expansion in caving volume. In contrast, proportional variations in tensile strength produced relatively minor changes in the final caving geometry, indicating that the ultimate extent of the caving zone is governed primarily by the shear strength components of the rock mass. The three-dimensional morphological progression of the caving zone under different parameter regimes is presented in Fig. 13, which shows that as mechanical properties degrade, the failure envelope expands and increasingly approximates an idealized conical geometry. The continuum-based framework developed in this study addresses a critical methodological gap between conventional continuum analyses and the Discrete Element Method (DEM). While DEM is capable of explicitly modeling fragmentation and inter-particle dynamics, its application at the field scale is often computationally prohibitive. Standard continuum approaches, in contrast, typically approximate caving zones as static plastic regions and fail to capture the physical process of detachment. By implementing a tensile-stress-driven element removal strategy, the proposed method enables the simulation of the continuum-to-discontinuum transition with a favorable balance between computational efficiency and geometric fidelity. Nevertheless, the applicability of the model is subject to two principal limitations. First, it assumes that caving is driven by tensile instability, a premise that is well-suited to weak-to-moderate rock masses but may be less representative of highly jointed or blocky ground conditions where structural discontinuities dominate the failure process. Second, the simulation exhibits marked sensitivity to the internal friction angle, a parameter that directly governs the stability of the stress arch. Future research should therefore focus on the development of hybrid continuum-discrete frameworks and the integration of real-time field monitoring data, including microseismic and displacement measurements, to enhance the predictive reliability of the model in increasingly complex geomechanical environments. Fig. 12The height and volume of the caving zone under different parameters. (a) variation of the cohesion; (b) variation of the friction angle; (c) variation of the tensile strength. The height and volume of the caving zone under different parameters. (a) variation of the cohesion; (b) variation of the friction angle; (c) variation of the tensile strength. Fig. 133D morphology of caving zone under different properties. (a) variation of the cohesion; (b) variation of the friction angle; (c) variation of the tensile strength. 3D morphology of caving zone under different properties. (a) variation of the cohesion; (b) variation of the friction angle; (c) variation of the tensile strength. The predicted caving-crest elevation of approximately − 234 m indicates that the influence of excavation extends upward beyond the mined levels and that the overlying roadway system may still be affected by excavation-induced stress redistribution. Accordingly, the displacement contours at the − 240 m and − 220 m levels provide useful guidance for field verification of the caving boundary. A monitoring drift excavated from the existing roadway network could be used for the installation of instrumentation to confirm the actual extent of failure, while targeted boreholes drilled from the existing access may help define the true caving boundary. In addition, localized backfilling with granular materials at the − 220 m level may improve lateral confinement and reduce further roof deformation. These engineering implications are derived from the simulated caving geometry and deformation pattern and should be regarded as practical guidance for field validation and risk mitigation rather than as prescriptive design values40,41. In summary, the proposed model links stress redistribution, tensile instability, and the geometric evolution of the caving zone within a unified three-dimensional framework. By combining theoretical consistency with numerical implementation, it offers a practical tool for analyzing caving processes and supporting risk assessment in underground mining. Conclusion A three-dimensional continuum-based numerical framework incorporating an iterative element-removal scheme is developed to simulate progressive rock caving. By adopting tensile-stress-induced detachment as the primary failure criterion, the method successfully captures the transition from continuum deformation to discontinuous caving within a computationally efficient framework. Validation against classical arch theories demonstrated good agreement, particularly with the self-stable arch theory, confirming the reliability of the proposed numerical approach across various undercut spans. The application to the 37 − 2# orebody of the Anhui Taiping Mine reveals that the caving zone expands upward progressively while maintaining an arch-like profile in vertical sections, with a final caving crest reaching approximately − 234 m and a cumulative volume of 23,545.4 m3 after extraction of the − 300 m level. Sensitivity analysis identified the internal friction angle as the dominant parameter controlling caving extent, with a 40% reduction in friction angle leading to a 145.3% increase in caving volume. The framework has certain limitations that assumes tensile-dominated failure and may be less applicable to highly jointed ground where structural discontinuities govern collapse. Future work should focus on hybrid continuum-discrete approaches and the integration of real-time microseismic monitoring to enhance predictive capabilities under complex geomechanical conditions.