The study area is located in the central Xiong'an New Area, Hebei Province, covering ∼ 536 km². It is bounded by the Nanjuma River (northeast), Baigouyin River Diversion Canal (east), Baiyangdian Lake (south), Dingxing County (northwest), and a shallow groundwater depression cone (west) (Fig. 1). The region lies at the center of the NCP, characterized by flat alluvial plains (average elevation 10–15 m) and Quaternary deposits ranging from 10 to 500 m in thickness. Groundwater depth ranges from 5 to 30 m. Groundwater in the Xiong'an New Area occurs in Quaternary alluvial–pluvial unconsolidated deposits forming four aquifer groups (I–IV). We focus on the shallow system (Groups I–II) and the unsaturated zone. Land use is dominated by arable land, with construction land as a secondary use, and some grassland in certain regions (Fig. 1). The study area has a warm temperate humid monsoon climate, with mean annual precipitation of 400–501 mm, more than 75 % of which occurs between June and September.

Groundwater recharge sources mainly include infiltration from atmospheric precipitation, lateral recharge, river infiltration and irrigation return, while pumping is the dominant discharge. Long-term over-extraction has formed a distinct groundwater depression funnel, typical of the NCP. Since 2018, artificial recharge has been implemented through diversion of South-to-North Water Transfer Project inflows into local rivers, aiming to restore groundwater levels and secure long-term water supply for Xiong'an.

Groundwater and river water samples were previously collected to characterize water quality and chemistry and measure chemical parameters (Chen, 2022). Ca²⁺ in groundwater is between 9.075 × 10⁻⁴ and 2.7 × 10⁻³ mol L⁻¹, primarily from the dissolution of calcite and gypsum. HCO3- concentration of groundwater was from 3.1 × 10⁻³ to 8.92 × 10⁻³ mol L⁻¹ and SO42- concentration of groundwater was from 3.13 × 10⁻⁴ to 2.41 × 10⁻³ mol L⁻¹. At some sampling points, nitrate concentrations reached 1.94 × 10⁻³ mol L⁻¹, exceeding the U.S. Environmental Protection Agency (EPA) drinking water NO3- standard of 8.87 × 10⁻⁴ mol L⁻¹.

Eight simulation scenarios (Table 3) were designed to evaluate the impact of geological structural conditions, artificial recharge measures, and denitrification processes on nitrate concentration dynamics in groundwater. In the groundwater flow simulation, a homogeneous case was established by dividing the model domain into four hydrogeologic zones based on local geological records (Li et al., 2023), with the hydraulic conductivity of each zone determined from pumping test results (Xu, 2022) to represent the regional hydraulic properties of the study area. To investigate the influence of stratigraphic heterogeneity on solute migration, reactive transport simulations were performed for 20 heterogeneous realizations and the ensemble-averaged results were compared with those from the homogeneous case.

Scenario combinations are constructed by considering the presence or absence of artificial recharge, denitrification, and geological heterogeneity, thereby enabling a comprehensive assessment of individual and interactive effects. To assess the impact of artificial recharge on regional groundwater level recovery, Case 1 and Case 2 are compared, while the role of artificial recharge in nitrate removal is analyzed by comparing Case 1 with Case 2 and Case 5 with Case 6 under different geological structures. The denitrification contribution is analyzed by comparing Case 1 with Case 3 and Case 5 with Case 7 to quantify the role of denitrification in nitrate reduction. The synergistic effect is discussed by comparing Case 1 with Case 4 and Case 5 with Case 8 to assess the combined impact of artificial recharge and denitrification reactions and their additive or antagonistic effects.

To reduce bias caused by concentration magnitude difference among scenarios and to highlight relative changes, the concentration differences (ΔC) for each scenario are logarithmically transformed. This transformation helps visualize the distribution patterns of high and low variation zones in the spatial distribution analysis.

To verify the model's reliability, simulated groundwater levels were compared against observations from four observation wells (Points 1–4) over 2017–2019 (Fig. 3a–d). The results show that the simulated water level trends at the four observation points are generally consistent with the measured values (R2=0.95) and accurately reflect the dynamic changes in water levels during both the recharge and non-recharge periods. In particular, during the first recharge phase from August to December 2018, rise in water levels was observed at Point 1 and Point 2, which are located near the recharge area, and the model was able to accurately capture this sudden increase. In contrast, the more muted responses at distal wells (Points 3 and 4) were also reproduced. Additionally, a domain-wide groundwater budget was computed, with major inflow and outflow components summarized in the Supplementary Information (Table S4). Collectively, these results indicate that the constructed groundwater flow model reliably represents the spatiotemporal variations in groundwater levels under recharge and non-recharge conditions, providing a robust hydrodynamic basis for solute transport analysis (Fig. 3e).

We assess basin-scale groundwater level responses to MAR during 2023–2035 by comparing simulations with and without recharge (Fig. 4). Without MAR, the cone of depression persists with limited natural recovery (Fig. 4a–d). MAR systematically elevates water levels, weakens the depression, and propagates higher heads inward from the river margins (Fig. 4e–h).The corresponding difference field (Δh, defined as head_MAR - head_no MAR), shows a persistent high-Δh corridor along the northeastern river margin in all periods (Fig. 4i–l), with maximum rises up to 7.5 m and a mean gain of 1.11 m. Influence extends ∼ 5 km from the recharge reach, decaying toward the depression center where recovery remains limited. From January 2023 to January 2031, the high-Δh belt expands slightly inland and the gradient relaxes (Fig. 4i–k); by January 2035 (Fig. 4l), the pattern stabilizes with only modest additional propagation. Taken together, MAR elevates heads most strongly near the river and transmits effects inward with distance- and connectivity-controlled attenuation, consistent with the area's permeability structure (Fig. 4). A quantitative zonal analysis further confirms this control: by 2035, Zone 1 acts as a hydraulic barrier with 100 % of cells showing minimal recovery (Δh<1 m). In contrast, the high permeability corridors facilitate significant recovery, with Zone 3 showing extensive coverage (64 % cells > 1 m) and Zone 4 achieving the highest mean rise (∼ 3.18 m) and peak values (∼ 7.5 m) (Fig. S2). This demonstrates that strong MAR benefits are spatially confined to conductive pathways (Zones 3–4) rather than being uniformly distributed.

Time-series head from five monitoring points (Fig. S3) further illustrate distance- and permeability-dependent responses. Points 1 and 2 are located near the recharge area, with Point 1 in a high-permeability zone and Point 2 in a low-permeability zone. Points 3 and 4 are progressively farther from the recharge location, with the distance between Point 4 and the recharge location being approximately twice that of Point 3. Point 5 lies within the cone of depression. Near the river, Point 1 (high-permeability) and Point 2 (low-permeability) exhibit rapid onsets during recharge windows and persistent head gains thereafter. Although the two sites respond nearly synchronously, their cumulative water-level rise varies with permeability: Point 1 rises by ∼ 4–7 m, while Point 2 increases by ∼ 3–5 m across the projection horizon (Fig. 5a, b). Superimposed on the step-like elevation is a distinct seasonal cycle – peaks that coincide with recharge pulses and troughs that reflect background abstraction – indicating that MAR augments, rather than overrides, the natural intra-annual variability.

With increasing distance from the river, the MAR signal becomes weaker and arrives later. Inland points (Points 3 and 4) exhibit delayed and dampened responses (∼ 2–4 m and ∼ 1–2 m, respectively), while the depression-center well (Point 5) shows only minor gains. Seasonal cycles remain evident at all sites, with recharge pulses superimposed on background abstraction. The amplitude hierarchy (Point 1>2>3>4>5) is consistent across years, demonstrating that permeability controls recovery magnitude near the river, while distance and connectivity govern inland attenuation.

Under homogeneous geology, MAR induces a smooth and continuous nitrate reduction pattern, strongest along the river corridor and extending inland over time, (Fig. 6a–d). From 2023 to 2035, the high-reduction belt expands landward while preserving strong spatial continuity and year-to-year persistence. By 2035, cells exhibiting a discernible reduction cover 67.85 % of terrestrial model cells, indicating that in homogeneous media the recharge signal propagates rapidly and relatively uniformly and sustains a basin-scale decline in nitrate.

In contrast, under heterogeneous conditions, the same reduction appears, but it is strongly patchy and anisotropic (Fig. 6e–h). High-permeability corridors form reduction hotspots, while low-permeability and poorly connected zones show weaker or negligible changes. Inland expansion is faster and more fragmented compared to the homogeneous case with 73.43 % of cells affected by 2035, 5.58 % higher than under homogeneous conditions, and pronounced reductions remain concentrated within ∼ 15 km of the river. This contrast reflects permeability-controlled flow partitioning and residence times that high-permeability pathways accelerate dilution and flushing, whereas low-permeability zones restrict response.

The spatial responses can be summarized by area-contribution analysis of signed log-difference classes relative to the baseline (Fig. 7a–c). Negative values indicate decreases in nitrate concentration; within this negative range, values closer to zero represent larger absolute decreases. The response classes correspond to orders of magnitude: -8 to -7 (very small decrease), -7 to -6 (small), -6 to -5 (moderate to large), and >-5 (the most distinct decreases, at least on the order of 10⁻⁵); while the absolute magnitude (10⁻⁵) appears minor, it represents the primary zones of cumulative mass removal in the system. Cells labeled NaN lie below the magnitude threshold. Both homogeneous and heterogeneous settings show a growing footprint of meaningfully responding cells (non-NaN) and a shift toward stronger response classes, but heterogeneity case achieves greater magnitude. In the homogeneous field (Fig. 7a), the non-responding cells (NaN) declines from 39.0 % (2023) to 32.2 % (2035), the moderate–large class (-6 to -5) expands from 5.4 % to 19.6 %, and the strongest class (>-5) from 1.28 % to 3.28 %, while the small-reduction class (-7 to -6) remains modal at roughly 31–33 %. In the heterogeneous field (Fig. 7b), the same pattern holds, but the magnitude is enhanced: NaN decreases from 35.2 % to 26.57 %, and the expansions into -6 to -5 and >-5 reach 22.43 % and 3.79 %, respectively, bigger gains than in the homogeneous case (Fig. 7c). Thus, MAR consistently expands the footprint of nitrate reductions, while heterogeneity enhances the intensity of reductions but weakens their spatial uniformity. To isolate the role of denitrification, recharge was disabled and only the reaction process was simulated. Under homogeneous geology, denitrification yields a spatially continuous nitrate decreaseacross the basin (Fig. 8a–d), which intensifies over time but retains a uniform pattern, indicating reaction control. Under heterogeneous geology, the reduction remains domain-wide but slightly non-uniform, with near-river high-permeability zones showing weaker declines due to shorter residence times that limit reaction (Fig. 8e–h).

Area-contribution analysis confirms this systematic trend (Fig. 9a–c). Both homogeneous (Fig. 9a) and heterogeneous (Fig. 9b) fields show a coherent order-of-magnitude shift and remain closely aligned in magnitude. The domain evolves from being entirely <-6.9 in 2023 (100 %) to being dominated by -6.7 to -6.6 in 2035, accounting for ∼ 90.0 % under the homogeneous field and 89.03 % under the heterogeneous field, with -6.9 to -6.8 prevailing in 2027 (∼ 95 %). This trajectory indicates a systematic amplification and narrowing of absolute decreases toward the 10^−6.7–10^−6.6 range and shows limited sensitivity to permeability heterogeneity overall. Notably, the overall removal efficiency (Fig. 9c) is ∼ 1 % lower under the heterogeneous setting, consistent with its slightly smaller share in the strongest response classes. The modeled negligible and uniform denitrification is fundamentally driven by the hydrogeochemical regime of the study site. Field data (Table 2) indicates that the aquifer is generally oxidizing and carbon-limited. This outcome validates the suitability of the adopted two-step reduction scheme (Eq. 5), which effectively captures the suppression of denitrification rates under such donor-poor and oxidizing conditions. Under such conditions, the reaction kinetics are globally suppressed by oxygen inhibition and electron donor starvation. Consequently, the influence of physical heterogeneity (i.e., variability in residence time) becomes secondary to this overwhelming chemical limitation. Even in zones with long residence times, the reaction rates remain low due to the lack of favorable redox conditions, resulting in a spatially uniform distribution of minimal denitrification.

When MAR and denitrification act together, nitrate declines basin-wide relative to the no-recharge, no-reaction baseline (Fig. 10). Under homogeneous geology (Fig. 10a–d), the combined effect produces a smooth near-river belt that expands inland from 2023 to 2035, yielding a broad footprint with 95.58 % of terrestrial cells affected. The spatial pattern indicates recharge-driven dilution dominates within ∼ 5 km of river, where concentrations drop rapidly, while denitrification governs the slower decay in the 5–15 km inland zone. Under heterogeneous geology (Fig. 10e–h), the combined effect also lowers nitrate but forms a patchy, anisotropic mosaic due to preferential flow along high-permeability corridors. The overall affected fraction is 94.38 %, and spatial heterogeneity is more evident than in the homogeneous case, with stronger contrasts between well-connected and poorly connected regions.

Area-class analysis (Fig. 11a–c) shows that the combined forcing yields the largest and most persistent expansion of favorable classes in both settings, with the heterogeneity case advancing further. Under the homogeneous field (Fig. 11a), the small-reduction band (-7 to -6) increases from 49.8 % in 2023 to 60.3 % in 2035, the moderate–large band (-6 to -5) from 6.0 % to 27.7 %, and the strongest band (>-5) from 1.28 % to 3.31 %. In the heterogeneous field (Fig. 11b), the small-reduction band is nearly unchanged, while the moderate–large (2.9 % to 31.0 %) and strongest (1.35 % to 3.83 %) bands grow more than in the homogeneous case. Unlike the MAR-only runs, the heterogeneous setting also retains a larger share of non-responding cells (NaN increasing from 2.96 % to 5.61 %, compared with 1.67 % to 4.42 % in the homogeneous case). Overall, dilution coupled with reaction maximizes nitrate decrease; the homogeneous configuration secures greater gains in the moderate–large class, whereas heterogeneity acts as a brake on the strongest tail and accentuates spatial unevenness.

Along the transect at y=13 km, four monitoring points (x=28, 25, 22 and 19 km, listed from nearest to farthest from the recharge reach) were used to evaluate the MAR effect as the concentration difference between the MAR and no-MAR scenarios, where positive values indicate that MAR lowers nitrate and larger values denote stronger reductions. In the homogeneous field (Fig. 12a), all sites exhibit step-like year-to-year increases that align with the August–December recharge period, evidencing a clear multi-year cumulative effect; spatially, the response decays monotonically with distance from the recharge reach (28 km > 25 km > 22 km > 19 km), indicating distance control under near-uniform media. In the heterogeneous field (Fig. 12b), The near-river site (28 km) shows the largest amplitudes, with higher post-season steps and faster cumulative growth than under homogeneity. The two mid-distance sites (x=25 and 22 km) have nearly identical onset times and amplitudes, implying that preferential high-permeability pathways transmit the recharge signal and offset simple distance control. By contrast, the distal site (x=19 km) has a weak mean response but an uncertainty band second only to the near-river site (28 km), indicating large variability in far-field connectivity across realizations. Overall, heterogeneity amplifies the near-field response while flattening differences between the two mid-distance sites, and increases spatiotemporal variance-patterns consistent with the area-class analysis.

Average concentration analysis indicates that by 2035, nitrate reductions from recharge alone, denitrification alone, and their combined effect are 2.55 × 10⁻⁶, 2.62 × 10⁻⁷ and 2.81 × 10⁻⁶ mol L⁻¹, respectively, under homogeneous conditions, and 2.75 × 10⁻⁶, 2.61 × 10⁻⁷ and 3.01 × 10⁻⁶ mol L⁻¹ under heterogeneous conditions (Fig. 13a). These correspond to nitrate reductions of 30.72 %, 3.16 %, and 33.8 % in homogeneous, and 31.73 %, 3.01 %, and 34.75 % in heterogeneous settings (Fig. 13b). These findings suggest that: (i) recharge plays a dominant role in nitrate removal, accounting for about 91 % of the total reduction, significantly more than denitrification (about 9 %); (ii) geological heterogeneity slightly enhances the recharge effect (1.01 %), but has little to no impact on denitrification (p>0.05); (iii) the combined effect is nearly equal to the sum of the individual effects (difference <0.5 %).