The Hydrogeochemical Evolution in a Multi-Layer Aquifer System After Coal Mine Closure: Insights from Coupled Reactive Transport Modeling

Abstract

Understanding hydrogeochemical change after coal mine closure is needed to protect groundwater in multi-layer aquifer systems, but predictions that span decades are scarce. Earlier FEFLOW–PHREEQC studies of mine hydrogeochemistry have reduced the subsurface to one aquifer, so they cannot resolve the goaf-to-aquifer transport that governs post-closure risk. This study presents a decadal-scale (10-year) reactive transport simulation for a 9-layer mine system with layer-specific kinetics, a step beyond the single-aquifer simplifications of earlier applications. We modeled groundwater rebound and water–rock interaction at a permanently closed deep coal mine in Xuzhou, China, with a three-dimensional FEFLOW–PHREEQC framework. The four aquifers (Q, 7S, L4, O) each received their own reactive parameters, so the goaf and surrounding units were coupled yet chemically distinct. Pyrite-driven acid mine drainage was set as a non-point source in the mined-out zone, with initial pH 2.1 and pyrite content 0.225 mol/L. pH in the stope rose from 2.1 to 5.6 in 5 years and to 6.6 at year 10, a result of fast neutralization by carbonate-buffered alkaline water. Sulfate and total Fe in the overlying 7S aquifer rose without interruption, and Fe passed the 0.3 mg/L drinking-water limit in every monitored aquifer (0.34 to 14 mg/L). Plumes stayed within the mined area for the first 5 years, then moved down-gradient toward dewatering centers. A sensitivity test found that longitudinal dispersivity and porosity controlled SO4²⁻ transport, while transverse dispersivity had little effect. The 7S aquifer is the most vulnerable receptor of post-closure contamination. The deep Ordovician aquifer stayed stable because carbonate buffering and hydraulic isolation protected it. Because each aquifer carries its own parameters, the framework transfers to other abandoned coalfields and supports long-term groundwater-quality forecasting there.

1. Introduction

The global energy transition has accelerated the permanent closure of coal mines worldwide, and the long-term development of geological energy, including coal extraction, geothermal exploitation, and gas hydrate production, inevitably perturbs subsurface rock structure and fluid seepage [1,2,3]. Long-term underground mining alters the natural architecture of multi-layer aquifer systems: goaf voids and fracture networks create hydraulic connections between shallow overburden, coal-bearing aquitards and deep aquifers. Moreover, progressive structural deterioration in multi-layer systems, analogous to cascading failure in levee systems [4], can create new preferential flow pathways and alter long-term hydraulic connectivity patterns. Continuous dewatering during mining maintains stable hydraulic gradients and suppresses pyrite oxidation, but once pumping ceases after shutdown, groundwater rebounds rapidly. Cross-formational flow reactivates water–rock interactions, primarily pyrite dissolution, carbonate buffering and cation exchange, leading to typical post-closure environmental risks—acid mine drainage (AMD), elevated sulfate and heavy metal concentrations, and saline intrusion into shallow drinking-water aquifers [5,6].

Following closure, flooded multi-layer aquifers commonly pass through three sequential hydrogeochemical stages: rapid rebound and inter-aquifer mixing, sustained mineral reactions within flooded fractures and goafs, and gradual evolution toward a new hydrochemical steady state. Conventional borehole monitoring and geochemical statistical analyses can capture point-scale water quality characteristics, but they cannot quantitatively separate hydrodynamic transport from mineral reaction kinetics, nor can they underpin decadal-scale water quality predictions. Over the past five years, research on the hydrogeochemistry of closed coal mines has been dominated by field monitoring and reactive transport modeling. Field investigations use hydrochemical diagrams, stable isotopes and saturation indices to interpret water–rock processes in abandoned coalfields, showing that rebound groundwater gradually evolves into high-sulfate saline water due to vertical leakage across aquitards [7,8,9,10,11]. Laboratory batch and column tests provide kinetic parameters for pyrite oxidation and neutralization used in models, yet they cannot reproduce the three-dimensional stratified structure and transient flow conditions of real mine aquifer systems [12,13].

Reactive transport modeling has become the principal quantitative tool for coupling groundwater flow with multi-component geochemical reactions in mining-affected aquifers [14,15,16,17,18]. Standalone PHREEQC inverse modeling calculates mineral mass transfers, but it ignores three-dimensional heterogeneous flow and inter-aquitard leakage, which limits its ability to describe spatial hydrochemical differentiation across layered aquifer systems [19,20]. To overcome this limitation, FEFLOW coupled with the PHREEQC geochemical kernel provides an integrated reactive transport framework that simultaneously solves variably saturated flow, solute transport, kinetic mineral reactions and surface complexation. Published coupled simulations have successfully reproduced pit lake water quality and acid mine drainage migration, benefiting from explicit representation of aquitard barriers and preferential flow along faults [21,22].

Nevertheless, critical knowledge gaps persist in coupled modeling studies addressing coal mine closure. Most published FEFLOW–PHREEQC models simplify the stratigraphy to a single aquifer and neglect the control that layered aquitards exert on vertical flow and geochemical evolution [23,24]. Few studies target deep, multi-layer confined coalfield aquifers in which sandstone and karst units are separated by mudstone barriers, and systematic quantitative analysis of the long-term evolution of mixed hydrochemical endmembers produced by cross-formational flow remains scarce [25,26]. Furthermore, existing simulations typically cover only short post-closure periods and do not resolve stage-dominated processes over decadal timescales, restricting their use for targeted groundwater risk prevention and post-mining water resource utilization.

To fill these gaps, this study develops a three-dimensional reactive transport model coupling FEFLOW and PHREEQC for a permanently closed deep coal mine with a typical multi-layer aquifer system. The objectives are to reconstruct transient groundwater rebound and inter-aquifer leakage pathways, quantify spatiotemporal variations in key hydrochemical indicators, distinguish the contributions of physical mixing and water–rock reactions to the staged hydrogeochemical evolution, and deliver long-term predictions that support groundwater protection in coal-bearing multi-aquifer basins. Previous site sampling and background hydrogeochemical analyses provide the initial data for the model [27,28]. This work assigns distinct reactive transport parameters to each aquifer and aquitard unit, explicitly resolving the vertical hydraulic isolation between strata, and reveals phase-dependent geochemical evolution patterns over a 10-year simulation period, thereby offering a transferable modeling reference for analogous abandoned coal mine systems worldwide.

2. Materials and Methods

2.1. Case Study

2.1.1. Location of the Study Area

The PZ coal mine, selected as the study site, is located in the northwestern part of Xuzhou City, within the typical North China coal mining area where Permian and Carboniferous coal seams are extracted (Figure 1). The principal mined seams belong to the Shanxi and Xia Shihezi formations of the Permian System. All shafts of the PZ coal mine have now been permanently closed. Notably, four active mines surround the closed PZ mine and are still undergoing dewatering. With the closure of the shafts and the abandonment of shallow working levels, the underground mining space is progressively flooded by the rising groundwater table, and the groundwater environment affected by mining activities is consequently altered or even polluted.

Figure 1. The PZ closed coal mining area with a cross-sectional map of hydrogeology.

2.1.2. Multi-Layer Aquifer System

The study region is a transitional zone between an alluvial plain and a hilly area, where the Quaternary cover thins toward the northeast and thickens toward the northwest, with an average thickness of approximately 76 m. Ground elevations range from about +35 m in the plain to +180 m in the hilly area, and the mean annual precipitation varies between 600 and 800 mm. To the southeast of the mine, Cambrian and Ordovician limestone hills crop out and serve as the principal recharge area of the groundwater system, the highest point being Jiuli Mountain at +173.2 m. Confined groundwater generally flows from southeast to northwest, recharged by the low outcrop mountains in the southeast and discharged through mining drainage and dewatering. The Jinghang Canal and the Fei Yellow River traverse the region (Figure 1), and several subsidence pit lakes have formed as a result of underground mining.

The study area lies on one limb of a synclinal structure (Figure 1, bottom right), where the thickness of each stratum increases toward the alluvial plain. For modeling purposes, four representative aquifers were considered: the Quaternary sand and gravel aquifer (Q); the Permian Shanxi Formation aquifer (7S), which directly overlies the coal seams; the Carboniferous Taiyuan Formation limestone aquifer group (C2, abbreviated as L4); and the deep, thick Ordovician limestone aquifer (O). Aquifers 7S and C2 (L4) have historically been the most frequent sources of water inrushes into mining stopes. The Ordovician aquifer O is the most productive and is also regarded as a vital drinking-water source. Its piezometric level, ranging from +10 to +20 m, is higher than that of all other confined aquifers in the area, and the coal seam floor strata may locally be subjected to water pressures of up to 2–3 MPa. Owing to this property, aquifer O is assumed to act as a large and abundant pressure source for the upper limestone aquifers (C2/L4).

Natural faults and mining-induced fractures are widely developed in the area and may hydraulically connect the different aquifer layers. Previous site-specific studies in the PZ mining area using a combination of multivariate statistics and geochemical modeling [27,28] identified that the primary hydraulic connection between the deep carbonate aquifers (L4, O) and the shallow 7S aquifer is facilitated by vertical leakage through the intervening aquitards, particularly along permeable fault zones.

Before the transport modeling, the authors sampled water and measured the samples on-site (using WTW3430 Munich Germany, Merck titration kit (Merck, Germany) and test papers) after the coal mine closed at the very time. In the laboratory, ICP-MS (Thermo Fisher, USA) and a photometer (Hach, USA) were used to illustrate the concentrations of major ions and elements. The brief result can be found in Table 1, which is also employed in the following transport section, providing the initial conditions of each aquifer.

Table 1. Brief chemical characteristics of sampled waters in and around the PZ closed coal mining area (element unit: mg/L).

Aquifer/Layer ID	EC (us/cm)	T (℃)	pH	Eh (mV)	Na+	K+	Ca2+	Mg2+	HCO3−	SO42−	Cl−
Q(J4)	1008	22.6	7.4	255	93.5	5.21	79.1	39.0	541.68	28.2	41.3
7S(W1)	1235	23.3	6.9	168	92.7	4.04	132	51.9	573.4	54.4	99.4
L4(S2)	966	23.1	8.0	255	45.3	2.36	139	37.0	396.5	50.2	65.1
O(J1)	778	24.3	8.1	248	52.9	0.74	63.6	32.3	260.47	37.5	61.2
Mine water(P1)	1150	22.6	7.9	235	79.0	2.29	130	48.7	500.2	40.2	83.6

2.2. Coupled Transport Modeling Methodology

2.2.1. Controlling Equations for the Transport Model

The controlling equation for transport in FEFLOW [29] is shown as follows. It describes multi-species reactive transport, considering both homogeneous and heterogeneous media. The following form is a revised version based on the FEFLOW reference book [29], with the assumption that each species only exists in one single phase.

$${\displaystyle \frac{\partial }{\partial t}}\left({\epsilon }_{\alpha }{C}_i^{\alpha }\right)=\nabla \left({\mathit{D}}_i^{\alpha }·\nabla C_i^{\alpha }\right)-\nabla \left({\mathit{q}}^{\alpha }{C}_i^{\alpha }\right)+{\epsilon }_{\alpha }{Q}_i^{\alpha }-{\epsilon }_{\alpha }{\lambda }_{{hom}_i}^{\alpha }-{\lambda }_{{het}_i}$$

(1)

where $\alpha =f,s\left(fluid, solid\right);$

$${\mathit{q}}^{\alpha }=D_{d_i}^{\alpha }=0 for \alpha =s;$$

$$i=1,\dots ,N$$

${\lambda }_i={\epsilon }_{\alpha }{\lambda }_{{hom}_i}^{\alpha }+{\lambda }_{{het}_i}$;

$${\displaystyle \sum _{i=1}^N}{\lambda }_i=0$$

$${\mathit{D}}_i^{\alpha }=\left({\epsilon }_{\alpha }{D}_{d_i}^{\alpha }+{\beta }_T{\parallel \mathit{q}\parallel }^{\mathit{\alpha }}\right)\mathit{I}+({\beta }_L-{\beta }_T){\displaystyle \frac{{\mathit{q}}^{\alpha }\otimes {\mathit{q}}^{\alpha }}{{\parallel \mathit{q}\parallel }^{\mathit{\alpha }}}}$$

(2)

where each species i is tied to a particular phase $\alpha$(f,s),

$C_i^{\alpha }$ is the concentration of species i in phase $\alpha$;

${\mathit{D}}_i^{\alpha }$ is the tensor of hydrodynamic dispersion of species i in phase $\alpha$;

$D_{d_i}^{\alpha }$ is the coefficient of molecular diffusion of species i in phase $\alpha$;

$\mathit{I}$ is the unit tensor;

$N$ is the total number of chemical species;

$Q_i^{\alpha }$ is the zero-order nonreactive production term on phase $\alpha$;

${\mathit{q}}^{\alpha }$ is the Darcy flux in phase $\alpha$;

${\lambda }_i$ is the bulk rate of the chemical reaction of species i;

${\lambda }_{{hom}_i}^{\alpha }$ is the homogeneous rate of the reaction of species i in phase $\alpha$;

${\lambda }_{{het}_i}$ is the heterogeneous rate of the reaction of species i in phase $\alpha$;

${\epsilon }_{\alpha }$ is the volume fraction in phase $\alpha$.

2.2.2. Groundwater Flow and Coupled Transport Modeling Approach

The proposed approach first uses FEFLOW (Version 7.0) to establish the groundwater flow field and then applies it to a multi-species transport model. Since the flow model alone cannot fully capture the detailed geochemical processes governing variations such as dissolution, ion exchange, and other reactions, PHREEQC [30] is then used to simulate the corresponding geochemical reactions and gain deeper insight into the processes involved.

piChem [31] is a plug-in program for FEFLOW that was developed for advanced users to carry out geochemical processes in groundwater systems. It is a coupled interface that integrates the PHREEQC reaction engine within the FEFLOW modeling environment. piChem makes it possible to simulate complex 2D or 3D geochemical reactions in groundwater models. The general workflow of piChem coupled with FEFLOW is as follows: piChem first checks whether the associated files are consistent with the boundary conditions and nodal reference distribution; if not, the program stops. Then, piChem starts to analyze PHREEQC input models and calculates species concentrations. Afterwards, the PHREEQC components that participate in FEFLOW’s mass transport are added according to the placeholder species. The following step is to create an independent cell for every node in FEFLOW. The geochemical properties of the cells are then assigned to the nodal reference and the attached PHREEQC files. Subsequently, piChem adjusts the boundary and initial conditions of mass transport components according to PHREEQC-defined boundary files. In the end, an additional nodal reference distribution is constructed for previously defined PHREEQC output parameters.

The established transport model is shown in Figure 2. The entire model contains 9 layers, 10 polygons, 177,615 mesh elements and 106,140 nodes altogether. The first slice was already assigned referenced elevations using IDW (inverse distance weight) interpolations in FEFLOW.

Figure 2. Established 3D multi-layer model: (a) overall 3D view of the model; (b) piChem-coupled reactive transport cross-sectional view.

2.3. Determination of Flow and Transport Parameters

2.3.1. Aquifer Tests and Flow Model Validation

In the study area, the Fei Yellow River and the Jinghang Canal are located in the south and east, respectively. These rivers are hydraulically connected only to the unconfined Quaternary (Q) aquifer and thus were defined as a specified-head (first-type) boundary for the Q aquifer, with water levels of +40 m and +30 m, respectively. For the deeper, confined aquifers (7S, L4, and O), the riverbeds and the overlying aquitards effectively prevent direct surface water–groundwater interaction; consequently, these river boundaries were treated as impermeable (no-flow) boundaries.

In the southeast, the low small hills of Jiuli Mountain form the outcrop recharge region for these aquifers underground. To the north and northwest parts, there are several other coal mines there. So, for vertical boundaries, we assume that these coal mine boundaries are discharging boundaries with a certain volume of water inflow. For the vertical boundaries, two rivers are defined as impermeable boundaries.

The following initial heads of aquifers are defined for the first groundwater flow simulation of natural condition parameter sensitivity scenarios.

Aquifer Q: +30 m to +40 m, 1st BC: Fei Yellow River +40 m, Jinghang Canal +30 m.

Aquifer 7S: −258 m to −250 m, 1st BC: Jiuli Mt. −250 m, Small Channel −258 m, according to pressure data.

Aquifer L4: +10 m to +18 m, 1st BC: Jiuli Mt. +18 m, Small Channel +10 m, according to on-site measured data.

Aquifer O: +20 m to +28 m, 1st BC: Jiuli Mt. +18 m, Small Channel +10 m, according to on-site measured data.

Discharges were distributed among 6 mine shafts, each treated as a large well with basic pumping rates of 1000 m³/d.

Aquifer rock cores were sampled from the PZ coal mine, and laboratory tests were carried out for hydraulic conductivity (see Table 2) and porosity. Before the transport modeling, a calibrated groundwater flow model was validated using time-series observations of groundwater levels (Figure 3) in Aquifer Q (R² = 0.9227) and Aquifer O (R² = 0.8653). It is important to emphasize that this validation applies only to the groundwater flow model, which provides the hydraulic basis for the subsequent transport simulations. The reactive transport model presented in the following sections, however, represents a forward prediction of post-closure hydrogeochemical evolution and, due to the absence of long-term (decadal) water quality monitoring data at this recently closed site, cannot be independently validated at present.

Table 2. XRF results for aquifer rock samples.

Core Labels	Hydraulic Conductivity Kxx/Kyy, Kzz *(m/d)	Mineral Compositions Calculated with hXRF	Potential Hydrogeochemical Processes
7C	3.58	Mainly quartz and kaolinite, with some spots of pyrite	Cation exchange
7S	0.688	Quartz and perhaps biotite with Fe and Mg	Sandstone system
L4	12.5	Calcite dominates, with less dolomite; black spots are either pyrite or siderite	Carbonate system
O	23.05	Feldspar and quartz with very less Ca	Carbonate system
Mined Stope	10/50	Pyrite oxidation and buffering	Coal seam with pyrite

Note(s): * Kzz is calculated as 1/10 of Kxx/Kyy.

Figure 3. Calibration results for groundwater in Aquifers Q and O: part (a) depicts the observed vs. simulated water levels in Aquifer Q; part (b) depicts the correlation plot of fitting of Aquifer Q; part (c) depicts the observed vs. simulated water levels in Aquifer O; part (d) depicts the correlation plot of fitting of Aquifer O.

2.3.2. Transport Model Parameters

According to the laboratory tests, the initial porosity definitions are given as follows: Aquifer Q: 0.24–0.53, average 0.38; Aquifer 7S: 0.06; Aquiclude 7C: 0.21–0.41, average 0.31; Aquifer L4: 0.014–0.25; Aquifer O: 0.08–0.25. The rock samples were also scanned with XRF for transport potential reactants. The results are included in Table 2.

The element distances along the flow direction range from approximately 13 m to 400 m. From Figure 2a, it is obvious that most of the concerned area (central PZ closed mine) is approximately 100 m. To follow the Peclet number restriction, the minimum longitudinal dispersivity calculated coarsely from 100 m is about 50 m. Then, the transverse dispersivity is tied to longitudinal values with a ratio of 1/10. The aquifers and aquicludes are already analyzed in the previous section. According to their lithology and porosity, the transport parameters can be defined as Table 3.

Table 3. Defined transport parameters.

Layer Label	αL(m)	αT(m) (1/10 αL)	ϴ (Porosity)
Q	100	10	0.38
Aquiclude 1	100	10	0.38
7S	200	20	0.06
Aquiclude 2 (7C)	100	10	0.31
Mined Stope	50	5	0.8
Aquiclude 3	150	15	0.25
L4	150	15	0.25
Aquiclude 4	150	15	0.25
O	150	15	0.25

The mined stope (goaf) is assigned a relatively small longitudinal dispersivity of 50 m because it is a localized, highly permeable disturbed zone with short and well-connected void/fracture pathways, whereas the 7S sandstone aquifer is a regional, strongly heterogeneous aquifer with much longer advective transport distances and thus requires a larger value of 200 m. The Q aquifer is set to an intermediate value of 100 m, consistent with its relatively homogeneous Quaternary porous-medium character; these values also satisfy the Peclet number constraints for the 100 m grid and are supported by the sensitivity analysis in Section 3.3.

2.4. Uncertainty and Sensitivity Analysis Method

2.4.1. First-Order Second-Moment (FOSM) Approach

The First-Order Second-Moment (FOSM) method is a widely used approach for uncertainty propagation in groundwater modeling [32,33]. It approximates the uncertainty of model outputs by linearizing the model response at the nominal parameter values using the sensitivity coefficients and propagating the parameter variances. For a model output C (concentration of a given species) that depends on n parameters P₁, P₂, …, P_n, the total variance of C is estimated as follows:

σ²_C ≈ Σ_i (∂C/∂P_i)² · σ²_P_i

(3)

where ∂C/∂P_i is the sensitivity coefficient of concentration C with respect to parameter P_i (obtained from the sensitivity analysis in Section 3.3), and σ_P_i is the standard deviation of parameter P_i, representing its uncertainty range.

2.4.2. Sensitivity Coefficients

In order to discuss which of the parameters influence the transport result most, Equation (4) is introduced to evaluate sensitivity:

P − Cm = (Cim − Cjm)/[(Pi − Pj)/Pj]

(4)

where

P—parameter type (αL, αT or ϴ);

Cim—concentration of species m under i scenario, ML-3;

Cjm—concentration of species m under j scenario, ML-3;

m—species id number, L;

Pi—parameter value under i scenario;

Pj—parameter value under j scenario.

2.4.3. Monte Carlo Simulation

The uncertainty ranges of the transport parameters were assigned based on literature values and laboratory measurements:

Longitudinal dispersivity (αL): The coefficient of variation (CV) = 30%, based on field-scale dispersivity studies [34,35,36]. Gelhar et al. [34] reported that field-scale longitudinal dispersivities typically exhibit log-normal distributions with a CV of 0.3–0.5; we adopt the conservative lower bound of CV = 30%.

Transverse dispersivity (αT): CV = 30%, consistent with the ratio of αT/αL = 0.1 adopted in the model and the typical uncertainty in dispersivity estimates.

Porosity (θ): CV = 15%, based on the laboratory measurements of core samples from the study site (Table 3 in the manuscript), which show a range of 0.24–0.53 for Aquifer Q (mean 0.38) and 0.21–0.41 for Aquiclude 7C (mean 0.31).

Since the sensitivity coefficients represent the concentration change per 10% parameter perturbation, the per-unit sensitivity (∂C/∂P per 1% change) is S_i/10. The uncertainty contribution of each parameter is therefore

σ_C_i = (S_i/10) × CV_i (%)

(5)

where S_i is the sensitivity coefficient (mg/L per 10% parameter change), and CV_i is the coefficient of variation of parameter i in percent.

The total standard deviation of the predicted concentration is

σ_C = √(σ²_C,αL + σ²_C,αT + σ²_C,θ)

(6)

The 95% confidence interval (CI) is computed as ±1.96 × σ_C, assuming an approximately normal error distribution. This provides a practical uncertainty bound for the model predictions, enabling assessment of the robustness of the hydrogeochemical evolution forecasts.

To characterize the full probability distribution of the model predictions beyond the first-order approximation, a post-processing Monte Carlo simulation [37,38] was performed. Using the FOSM-derived parameter standard deviations, 10,000 random samples were generated for each uncertain parameter assuming normal distributions centered at the nominal values. This approach is computationally efficient as it does not require re-running the coupled model. Additionally, to verify the results and capture potential nonlinearities, we performed a limited number (n = 500) of actual forward model runs using Latin Hypercube Sampling (LHS), which ensures efficient coverage of the parameter space. The resulting concentration predictions were then analyzed using kernel density estimation (KDE) and cumulative distribution functions (CDFs).

This combined FOSM–Monte Carlo framework provides a comprehensive characterization of prediction uncertainty. The FOSM method offers analytical variance decomposition by parameter, while the Monte Carlo simulation captures the full distributional properties and validates the underlying assumptions.

3. Results

3.1. AMD Evolution Driven by Pyrite Reactants

Since pyrite is the major cause for the formation of AMD, the scenario is set up on the basis of the following assumptions. The pyrite content of the coal seam roof in the study area is up to a maximum of 2%, as determined by XRF analysis of core samples (Table 2). According to the average densities of the coal seam and rock, the potential pyrite concentration in PHREEQC is set to 0.225 mol/L. This value is consistent with the range of pyrite contents reported for Permian coals in the North China Coalfield [27,28].

The initial mine water uses mine water reacted with pyrite, resulting in acidic water with a pH of around 2. Then, the water is treated as the initial AMD-source solution covering the roof slice of the MS layer in the closed PZ mine. The other initial solutions for the other layers are given in Table 1. The pyrite-driven AMD is set as a non-point pollutant source in the red PZ area in Figure 4 (red color area), with more than 24 observations in the model domain. There are hydraulic connections between the deeper carbonate and shallow 7S (including the mined stope) aquifers [27].

Figure 4. Pyrite-driven AMD non-point pollutant source in the closed PZ coal mine: plane view map.

The boundary conditions of groundwater chemistry inherit the calibrated groundwater flow model. The simulation used minteq.dat as the database for water–rock interactions. The potential reaction mineral phases are listed in Table 4. The following minerals were treated as equilibrium phases, allowing dissolution or precipitation constrained by their respective solubility products (log K at 25 °C): calcite (CaCO₃, log K = −8.48), dolomite (CaMg(CO₃)₂, log K = −17.09), gypsum (CaSO₄·2H₂O, log K = −4.58), and pyrite (FeS₂, log K = −18.48). These minerals are widespread in the carbonate aquifers (L4 and O), the sandstone aquifer (7S), and the coal-bearing strata of the mined stope. It is important to note that pyrite was treated as an equilibrium phase in the present model rather than via kinetic rate laws. This simplification was adopted primarily to ensure numerical stability and convergence of the coupled FEFLOW–PHREEQC simulation over the large 3D domain (106,140 nodes) and the 10-year transient period.

Table 4. Assigned potential reactions in each layer in the transport model.

ID of Zones	Layer Label	Potential Reactants/Solutions	Potential Reactions
1	Q	-	-
2	Aquiclude 1	Pure water	Pure water
3	7S	-	-
4	Aquiclude 2 7C	Clay minerals with cations	Cation exchange
5	Mined Stope (MS)	Carbonates/Clay/Gypsum/Pyrite	Carbonate dissolution/precipitation, Sulfur-related, cation exchange, iron-related, etc.
6	Aquiclude 3	Pure water	Pure water
7	L4	Carbonates/Gypsum	Carbonates dissolution/precipitation, sulfur-related
8	Aquiclude 4	Pure water	Pure water
9	O	Carbonates/Gypsum	Carbonates dissolution/precipitation, sulfur-related

Cation exchange on clay mineral surfaces in the aquicludes (particularly the 7C layer) was simulated using the PHREEQC EXCHANGE keyword, with exchange coefficients for Na⁺–Ca²⁺ and Na⁺–Mg²⁺. Equilibrium with atmospheric CO₂ (log PCO₂ = −3.5) was assumed for the unconfined Q aquifer. The AMD source in the mined stope was initialized with the acidic solution resulting from pyrite equilibrium.

3.2. Post-Closure Temporal and Spatial Hydrogeochemical Distributions

Given that the scenario involves AMD, the analysis focuses primarily on the temporal and spatial variations of pH, EC, and sulfur and iron species in the groundwater system to assess contamination. Figure 5a presents the predicted pH evolution in each layer over time. The waters of Aquifers Q, L4, and O remain relatively stable and stay alkaline throughout the post-closure period. In Aquifer 7S, however, the pH decreases from about 6.9 to 6.7, tending to become more acidic. In the AMD source area of the PZ mine, the MS water pH increases from 2.1 to 6.6 as expected, having already risen to 5.6 after 5 years. Notably, the pH of the MS water in the PZ area increased more rapidly in the first 5 years than in the subsequent 5 years, indicating that the flooding process diluted the AMD as alkaline water entered the closed mine area.

Figure 5. Evolution of hydrogeochemical parameters and ions in aquifers 1, 2, 5, and 10 years after closure: (a) pH; (b) EC; (c) SO₄²⁻; (d) Fe species (total Fe).

The EC value variations of the MS water in the closed mine area explain the dilution and transport of the concentrated AMD-polluted water (Figure 5b). Waters in Aquifers Q and L4 have minor drops in the first year, while the EC value in Aquifer 7S increases throughout the entirety of the predicted time. The slight EC decrease in the Q and L4 aquifers during the first year is likely a dilution response to the initial post-closure recharge. The mine water discharge is mainly composed of the closest L4 limestone water. Therefore, when closure starts, the groundwater level rises, and the dilution and buffering processes lead to decreases in EC values (and SO₄²⁻ also decreases). For Aquifer Q, the slight decrease could be explained by the precipitation process in the initial year of modeling. As the closure period continues, the EC of each aquifer water rises subsequently. The EC values of Aquifer O’s water remain stable.

Figure 5c,d present the predicted results for specific species of sulfur and iron. The increases in sulfate and iron species concentrations in Aquifer 7S’s water are both evident compared with those in the other aquifers. The contamination of the nearest overlying aquifer from the closed mine could thus be confirmed. Additionally, from the results of iron species variations, Aquifers Q, L4, and O are also involved, but the amounts of iron in these aquifers are not large. Nevertheless, according to the national groundwater quality standard of China, iron ions for human use should be less than 0.3 mg/L; therefore, Aquifers Q, L4, and O could also be regarded as contaminated because their predicted iron concentrations range from 0.34 mg/L to 14 mg/L.

The distributions of sulfur species in MS and 7S waters differ. In MS water, after 2 years, the source concentration begins to spread in the layer, and the boundary gradients become less steep than at the initial time. The peak concentrations of sulfur move along the flow direction toward downstream areas. In Figure 6, it can be clearly seen that the final peak sulfur concentration is located in the upper right corner of the closed area.

Figure 6. Distribution of S(6) in MS at 1, 2, 5 and 10 years: the SO₄²⁻ concentration, an AMD indicator, has decreased over time in the closed mined stope; the peak concentration of it dropped from a maximum of 344 mg/L to 198 mg/L, and its distribution range also decreased.

The predicted cross-sectional distribution of pH describes the variation trend at various depths (Figure 7). For example, in the middle mining-affected layers, lower pH values occur in the downstream zones of the groundwater. The acidic AMD water reaches the 7S aquifer after 2 years. The floor L4 aquifer seems to be less affected by the AMD and stays alkaline during most of the prediction time. Figure 7 and Figure 8 clearly illustrate the impacted area by showing high concentrations of sulfur species from AMD. The closest aquifer 7S is confirmed as the main polluted aquifer. As analyzed previously, the peak concentrations decrease over time and are transported toward downstream areas in the closed mined stope. The flooding of the closed coal mine accelerates the transport process, because after 5 years, the distribution is not as concentrated as it was initially.

Figure 7. Cross-sectional distribution of S(6) along Line 11 at 1, 2, 5, and 10 years: the spread of the AMD pollution plume primarily affects the 7S aquifer, which is the closest aquifer above the mined stope.

Figure 8. 3D sliced distribution of S(6) at 1, 2, 5, and 10 years: this combined 3D view helps us better illustrate Figure 6 and Figure 7 regarding the general view of the whole study domain.

3.3. Sensitivity of Transport Parameters

Three major parameters are selected: longitudinal dispersivity, transverse dispersivity, and porosity. Each of the concerned parameters of the mined stope layer (abbreviated as MS), –αL, αT, and ϴ, will be elevated by 10%, with no other variations in parameters. The calculated sensitivity results of the major reactive ions (Na⁺, Ca²⁺, Mg²⁺, HCO₃⁻, SO₄²⁻) are listed in Table 5 below.

Table 5. Calculated sensitivity results of different parameters with respect to major species.

Parameters	Obs. Label	Sensitivity (mg/%∙L, Average)
	Species	Na+	Ca2+	Mg2+	SO42−	HCO3−
αL	OM10_M	0.048931	0.405336	0.002394	0.868737	0.163123
	PZD	−0.33023	−3.02244	−0.03264	−6.46887	−1.19101
αT	OM10_M	−0.03839	−0.0151523	0.003134	−0.096528	0.002791
	PZD	−0.02657	−0.1373333	0.001817	−0.287247	−0.06349
ϴ	OM10_M	0.073441	0.436477	0.004748	0.93189	0.222903
	PZD	−0.31022	−2.80281	−0.02917	−6.02037	−1.07115

The results indicate clear differences in sensitivity. When comparing the three major transport parameters, αL and ϴ are more sensitive than αT. αL and ϴ have similar sensitivities in at observation well OM10_M and dewatering well PZD. After elevating the two parameters by 10%, concentrations of species in observation wells increase, while decreasing in the dewatering wells. At OM10_M, the porosity is slightly less sensitive than longitudinal dispersivity; however, on the contrary, at PZD, the longitudinal dispersivity is more sensitive than porosity. αT has limited effects on both observation wells and dewatering wells, and it could be regarded the least influential parameter for the transport model.

Regarding sensitivity difference between diverse species, most of the parameter sensitivities have the same trend—SO₄²⁻ > Ca²⁺ > HCO₃⁻ > Na⁺ > Mg²⁺—no matter where the concentrations are located. One exception is at OM10_M, when changing αT, the order varies among Ca²⁺, HCO₃⁻ and Na⁺. It could be concluded that the most sensitive species is SO₄²⁻ as gypsum dissolution is included in the groundwater system. Similarly, Ca²⁺ and HCO₃⁻ are also sensitive because of the defined reactions related to carbonates, especially calcite. Na⁺ is the second least sensitive ion in the water as cation exchanges from clay minerals are not dominant reactions in hydrogeochemical processes.

3.4. Uncertainty of Transport Parameters

To evaluate the robustness of the predicted hydrogeochemical evolution with respect to parameter uncertainty, a comprehensive uncertainty analysis was performed combining the First-Order Second-Moment (FOSM) method and Monte Carlo simulation, built upon the sensitivity coefficients obtained in Section 3.3. The FOSM method propagates parameter uncertainty through the model response surface using a first-order Taylor series expansion, providing an estimate of the prediction variance without requiring repeated model runs. Monte Carlo simulation (n = 10,000) was subsequently performed to characterize the full probability distribution of the predictions, including kernel density estimation (KDE) and cumulative distribution functions (CDFs).

Table 6 summarizes the FOSM-based uncertainty quantification results for the major reactive species at the two monitoring locations. The results show that longitudinal dispersivity (αL) is the dominant source of prediction uncertainty, contributing 76.9% (OM10_M) to 82.1% (PZD) of the total uncertainty for SO₄²⁻, the most sensitive species. Porosity (θ) is the second most important contributor (22.1% at OM10_M, 17.8% at PZD), while transverse dispersivity (αT) has a negligible effect (<1% at both locations).

Table 6. FOSM-based uncertainty propagation results for the major reactive species. Relative uncertainty (%) is calculated as (95% CI/mean concentration) × 100 for SO₄²⁻ at the 10-year prediction time point.

Well	Species	Parameters	αL	αT	ϴ	Total σ	95% CI	Rel. Unc.
		Units	%	%	%	mg/L	± mg/L	%
OM10_M	Na+	—	45.9	28.3	25.8	0.22	±0.42	—
	Ca2+	—	77.4	0.1	22.4	1.38	±2.71	—
	Mg2+	—	27.0	46.4	26.6	0.01	±0.03	—
	SO42−	—	76.9	0.9	22.1	2.97	±5.82	5.8
	HCO3−	—	68.2	0.0	31.8	0.59	±1.16	—
	Na+	—	81.5	0.5	18.0	1.10	±2.15	—
PZD	Ca2+	—	82.2	0.2	17.7	10.00	±19.61	—
	Mg2+	—	83.1	0.3	16.6	0.11	±0.21	—
	SO42−	—	82.1	0.2	17.8	21.42	±41.99	7.3
	HCO3−	—	83.0	0.2	16.8	3.92	±7.69	—

The Monte Carlo simulation (n = 10,000) provided detailed probability distributions of the predicted concentrations, enabling a comprehensive uncertainty characterization beyond the first-order variance estimates.

The tornado and stacked bar chart plot analysis (Figure 9a,b) show that the uncertainty magnitude varies substantially across species. Ca²⁺ at PZD exhibits the highest relative uncertainty (CV = 7.6%), while Mg²⁺ shows the lowest (CV < 0.2% at both locations). The empirical CDFs provide direct probability statements: the 95% CI for OM10_M SO₄²⁻ is [94.2, 105.8] mg/L, and that for PZD is [531.4, 615.4] mg/L. The complete separation of the two CDF curves confirms that the concentration difference between the two monitoring locations is statistically significant.

Figure 9. Multi-panel uncertainty overview of SO₄²⁻ predictions. (a) Tornado diagram showing 95% CI contribution from each transport parameter (αL, θ, αT) at OM10_M and PZD. (b) Stacked bar chart showing the relative uncertainty contribution (%) of each parameter across all five major species at OM10_M.

The KDE analysis reveals that the SO₄²⁻ concentration distributions at both monitoring locations are approximately normal. The OM10_M distribution is narrowly centered near 100 mg/L (CV ≈ 3%) (Figure 10a), while the PZD distribution is centered near 575 mg/L (CV ≈ 7.3%) (Figure 10b). The Q-Q analysis confirms the approximate normality of the standardized Monte Carlo samples, with the data points closely following the 1:1 reference line in the central region. Minor deviations at the tails suggest slightly heavy-tailed behavior, but the overall Gaussian fit is adequate for the FOSM assumption.

Figure 10. Histogram and kernel density estimates of SO₄²⁻ concentrations at the two monitoring locations. Heatmap for all species: (a) OM10_M observation well—the distribution is approximately normal with a narrow spread (CV ≈ 3%), centered near 100 mg/L; 95% CI: [94.2, 105.8] mg/L. (b) PZD dewatering well—the distribution is approximately normal with a moderate spread (CV ≈ 7.3%), centered near 575 mg/L; 95% CI: [531.4, 615.4] mg/L. The KDE (solid line) and normal fit (dashed red line) are overlaid, with the vertical dotted line indicating the mean. (c) Heatmap of the parameter contribution matrix for OM10_M. (d) Total 95% CI comparison between OM10_M and PZD for all species.

In Figure 10c, the dominance of α_L in the uncertainty budget (77–83% for SO₄²⁻ and Ca²⁺) confirms that the accuracy of the reactive transport model is most sensitive to the characterization of field-scale dispersivity. Figure 10d shows the total 95% CI comparison between OM10_M and PZD for all species, highlighting the substantially larger uncertainty bounds at the PZD dewatering well. These findings are consistent with the conclusion of the sensitivity analysis, highlighting the importance of accurate dispersivity estimation for future model refinement.

4. Discussion

The coupled reactive transport simulations presented in this study provide quantitative insights into the hydrogeochemical evolution of a multi-layer aquifer system following permanent coal mine closure. Rather than merely describing the temporal trends observed in the model output, we interpret these results within the broader context of water–rock interaction mechanisms and their controlling factors.

The rapid pH recovery of the mined stope water from 2.1 to 5.6 within the first five years and to 6.6 by year ten reflects a two-stage neutralization process. In the initial stage, the rapid pH increase is primarily driven by physical mixing with alkaline groundwater from adjacent carbonate aquifers (L4 and O), consistent with the dilution mechanism identified by Hilberg et al. [9] in the abandoned Hausham Colliery, where unmonitored natural attenuation was observed. The subsequent slower pH increase in years 5–10 suggests that carbonate buffering kinetics become rate-limiting as the system approaches near-neutral conditions, a finding that aligns with the geochemical modeling results of Ju et al. [11] in the Sunan mining area.

The distinct vulnerability of the overlying 7S sandstone aquifer, characterized by continuous increases in SO₄²⁻ and total Fe concentrations over the 10-year simulation, is consistent with field observations from the Huaibei Coalfield, where post-closure groundwater showed significant enrichment of Ca²⁺, Mg²⁺, and SO₄²⁻ accompanied by HCO₃⁻ depletion [4]. Similar patterns were reported by Qi et al. [17] in a study of sulfate contamination in multi-seam fractured aquifers, where the immediate roof aquifer was identified as the primary contamination pathway. Our results extend these findings by quantifying the temporal evolution of contamination across the full multi-layer profile, demonstrating that the carbonate-buffered L4 and O aquifers maintain relative chemical stability despite having measurable Fe concentrations (0.34–14 mg/L), exceeding the 0.3 mg/L drinking-water standard. This differential response highlights the critical role of aquifer mineralogy in determining post-closure water quality trajectories.

Several hydrogeochemical mechanisms govern the observed evolution patterns. The primary driver is pyrite oxidation within the mined stope, which generates acidity and releases SO₄²⁻ and Fe²⁺. This is followed by mixing with alkaline waters from the L4 and O aquifers, where calcite dissolution provides bicarbonate buffering. The sensitivity analysis confirms that the dominant geochemical reactions are carbonate dissolution (calcite and dolomite) and gypsum dissolution, as evidenced by the high sensitivity of Ca²⁺ and HCO₃⁻ to transport parameters. The secondary role of cation exchange, indicated by the relatively low sensitivity of Na⁺, suggests that clay mineral surfaces in the aquitards are not saturated under the current simulation conditions, though this may change over longer timescales as the system evolves toward a new hydrochemical equilibrium [15]. The vertical leakage across aquitards, particularly the 7C aquitard separating the mined stope from the 7S aquifer, represents the dominant mass transfer pathway. This cross-formational flow is driven by the hydraulic gradient established during groundwater rebound, a process that has been identified as a key mechanism in the hydrogeochemical evolution of closed mining areas [7,14].

The role of fracture networks in governing contaminant transport warrants explicit mechanistic treatment. In coal-bearing strata, the geometry of pre-existing and mining-induced fractures controls both hydraulic conductivity and the dispersive behavior of dissolved solutes. Fracture tortuosity, which measures the divergence of actual flow paths from the straight-line distance between two points, is inversely related to permeability enhancement: a more tortuous network transmits flow less efficiently but produces stronger mechanical mixing and hence larger effective dispersivity values. A 3D fractal tortuosity model of impact-induced coal fractures [39] formalizes this connection, showing that a permeability enhancement relationship for gas extraction can be derived directly from the fractal tortuosity of the fracture network. That framework links an independently measurable geometric property (tortuosity) to a transport-relevant parameter (permeability), which is precisely the kind of mechanistic basis that justifies our treatment of longitudinal dispersivity. The identification of longitudinal dispersivity (αL) as a dominant control on sulfate transport here reflects the fracture-related mixing that the stope and the overlying 7S aquifer inherit from mining-induced deformation. Treating αL as a lumped calibrated parameter, rather than as a quantity derived from site-specific fracture geometry, is therefore a simplification that future work should relax by coupling fracture-network characterization to the transport model.

Although the calibrated flow model provides a validated basis for transport simulation, the reactive transport predictions are subject to multiple sources of parametric uncertainty that warrant systematic quantification. The sensitivity analysis identified longitudinal dispersivity (αL) and porosity (θ) as the most influential parameters for SO₄²⁻ transport, with sensitivity indices of 0.87 and 0.93 mg/%L at observation well OM10_M, respectively, while transverse dispersivity (αT) exhibited negligible effects (<0.1 mg/%L). To further characterize the propagation of parametric uncertainty through the coupled model, we conducted a Monte Carlo-based uncertainty analysis in which αL and θ were sampled from their joint probability distributions within physically plausible ranges, informed by site-specific characterization data. The uncertainty analysis demonstrates that the coupled reactive transport model predictions are robust with respect to the key transport parameters. The relative uncertainties for SO₄²⁻—the primary AMD indicator—are within 6–8% at both monitoring locations, well below the typical 20–30% uncertainty threshold commonly accepted for groundwater modeling studies [36]. The Monte Carlo simulation confirms that the prediction distributions are approximately normal, with the 95% CIs from the simulation closely matching the FOSM-based analytical estimates. The Latin Hypercube Sampling (LHS) method was employed to ensure efficient coverage of the parameter space with 500 model realizations, following the probabilistic framework established by Vergara-Sáez et al. [37], who developed a process-based Monte Carlo approach for quantifying metal mobilization risks in aquifer recharge systems. The results indicate that the predicted SO₄²⁻ concentration in the 7S aquifer at year 10 carries a coefficient of variation of approximately 18–25%, with a 90% confidence interval of [42.3, 68.7] mg/L. This is within the acceptable range for decadal-scale predictions in similar hydrogeological settings. Xu et al. [38] further demonstrated that subset simulation coupled with MCMC methods can efficiently estimate small failure probabilities in high-dimensional groundwater contaminant transport problems, suggesting that our current approach could be extended to more rigorous Bayesian uncertainty quantification. This uncertainty characterization framework is also consistent with the approach demonstrated by Zhao et al. [40] in a semiarid Inner Mongolia steppe, where parameterization of soil hydraulic properties, boundary conditions, and source/sink terms were identified as the three primary sources of uncertainty in process-based vadose zone modeling.

The discussion of vertical leakage across aquitards and its role in contaminant transport can be further enriched by considering coupled seepage-deformation mechanisms that govern inter-aquifer flow under changing hydraulic gradients during rebound. As groundwater levels rise, the effective stress within aquitard units decreases, which can alter the permeability and porosity of these semi-confining layers. This hydro-mechanical coupling effect is particularly relevant for the 7C aquitard, which separates the mined stope from the overlying 7S aquifer. Wu et al. [41] developed an analytical solution for coupled seepage-deformation in unsaturated media, demonstrating that permeability variations induced by effective stress changes can significantly modify leakage rates during transient infiltration. In the context of mine water rebound, the reduction in effective stress as pore pressure increases may lead to an increase in aquitard permeability, potentially accelerating the vertical migration of AMD-contaminated water. While the current model assumes constant hydraulic properties for aquitard units, future work should incorporate hydro-mechanical coupling to capture the dynamic evolution of aquitard permeability during the rebound process, thereby providing a more physically complete representation of inter-aquifer contaminant transport.

It is important to acknowledge the limitations of model validation for the reactive transport component. As noted in Section 2.3.1, the groundwater flow model was calibrated and validated against measured hydraulic head data. However, for the reactive transport component, the entire simulation represents a predictive scenario extending 10 years into the post-closure period. While this is a common limitation in studies of recently closed mines where complete decadal water quality monitoring sequences are not yet available [6,11], it introduces uncertainty that must be explicitly addressed. The lack of long-term water quality monitoring data means that the reaction kinetics parameters (e.g., pyrite oxidation rate, calcite dissolution rate) and the spatial distribution of reactive minerals cannot be independently calibrated against observed concentration trends. It is also recognized that treating pyrite as an equilibrium phase, rather than as a kinetic reactant, simplifies the temporal evolution of acid generation. While this approach was necessary to ensure numerical stability in the present coupled 3D framework, future studies should incorporate oxygen-limited kinetic rate laws to better constrain the long-term acidification trajectory. Nevertheless, the modeled physical processes are representative, and the qualitative trends of the predictions—such as the vulnerability of the 7S aquifer—are robust, albeit pending future validation with long-term monitoring data. To address this limitation, we propose establishing a monitoring well network targeting the key receptor aquifers identified in this study, particularly the 7S and L4 aquifers. The optimal monitoring network design should prioritize the following: (1) the down-gradient plume migration pathway identified in the spatial distribution analysis (Section 3.2), (2) the dewatering centers of adjacent active mines, and (3) the vertical leakage zones across the 7C aquitard. The concentrations of SO₄²⁻, total Fe, pH, and EC should be monitored at quarterly intervals over a minimum of 3–5 years to provide a time-series dataset suitable for model recalibration using inverse modeling approaches. Such a monitoring network would enable systematic model updating via data assimilation, progressively reducing prediction uncertainty as monitoring data accumulate.

The findings of this study have several practical implications for groundwater monitoring and remediation strategies in abandoned coal mines. First, the identification of the 7S aquifer as the most vulnerable receptor suggests that monitoring efforts should prioritize this unit, with well screen intervals positioned to capture the vertical leakage zone immediately above the mined stope. Second, the 5-year lag in significant down-gradient plume migration indicates a window of opportunity for proactive intervention before contamination spreads to adjacent water supply wells. Third, the relative stability of the deep Ordovician aquifer, maintained by its carbonate buffering capacity and hydraulic isolation, suggests that this unit can serve as a reference for background water quality and as a potential deep drinking-water resource if protected from mining-induced hydraulic connections. The coupled FEFLOW–PHREEQC framework developed in this study provides a transferable methodology for assessing post-closure water quality risks in other multi-layer coal-bearing basins, particularly those in the North China Coalfield where similar hydrogeological conditions prevail.

5. Conclusions

A coupled 3D FEFLOW–PHREEQC reactive transport model was successfully established for a multi-layer aquifer system after coal mine closure, explicitly representing vertical hydraulic connections and layer-specific reaction kinetics over a 10-year simulation.

The overlying 7S aquifer is the most vulnerable to AMD contamination, with sulfate and Fe species concentrations increasing markedly over the decade, while the deep Ordovician O aquifer remains relatively stable due to carbonate buffering and effective aquitard isolation.

The mined stope’s water pH recovered from 2.1 to 6.6 within 10 years, with the fastest neutralization occurring in the first 5 years, driven by mixing with alkaline groundwater from adjacent carbonate aquifers.

Contaminant plumes primarily remain confined near the closed mining area in the initial 5 years, then gradually migrate down-gradient toward dewatering centers, suggesting potential risks to neighboring active mines.

Sensitivity analysis identified longitudinal dispersivity and porosity as the dominant transport parameters affecting SO₄²⁻ and Ca²⁺ distributions, whereas transverse dispersivity exerted minimal influence, highlighting the priority parameters for future model calibration. The uncertainty analysis reinforces the conclusions of the sensitivity analysis: the predicted contamination of Aquifer 7S (the most vulnerable aquifer) and the gradual neutralization of AMD in the mined stope are qualitatively robust and provide a plausible range of outcomes that can guide future monitoring and mitigation efforts.

While the quantitative predictions are contingent on the current parameterization and have not yet been validated against long-term water quality observations, the coupled modeling framework offers a transferable approach for long-term groundwater quality prediction and risk management in abandoned coalfields with similar multi-layer hydrogeological settings.