Weather Research and Forecasting (WRF) (Skamarock et al., 2008) is configured to perform a high-resolution simulation with five interactive nested domains (27, 9, 3, 1, and 0.2 km) and 72 vertical levels (Fig. 1a and b). The initial and boundary conditions for these simulations are obtained from the Indian Monsoon Data Assimilation and Analysis (IMDAA) data (Rani et al., 2021), available at a 12 km resolution every 6 h. The Grell-Freitas “scale-aware” cumulus scheme (Grell and Freitas, 2014) is applied in the 27 and 9 km domains, while the inner three domains explicitly resolve convection at 3, 1, and 0.2 km. The Mellor-Yamada Nakanishi and Niino (MYNN) level 2.5 closure scheme (Nakanishi and Niino, 2009) is implemented for the planetary boundary layer (PBL) processes in the outer four domains. However, the innermost domain of 0.2 km is configured in the large-eddy simulation (LES) mode, without any PBL parameterization scheme. In this LES configuration, subgrid-scale turbulent motions are represented using the 1.5-order turbulent kinetic energy (TKE) closure, while sixth-order monotonic horizontal diffusion is applied to suppress grid-scale numerical noise. Additional model physics includes the RRTMG scheme for both longwave and shortwave radiative transfer (Mlawer et al., 1997) and MM5 similarity theory for estimating surface heat and moisture fluxes.

An updated two-moment Thompson aerosol-aware microphysics scheme (Thompson and Eidhammer, 2014) is employed to incorporate the influence of aerosols via cloud condensation nuclei (CCN) and ice-nucleating (proxy for IN) aerosols into the model. Land surface processes across all domains are represented using Noah-LSM (Chen and Dudhia, 2001) coupled with a single-layer urban canopy model (Kusaka et al., 2001; Kusaka and Kimura, 2004). The CCN concentrations are estimated from the aerosol mixing ratios obtained from Modern-Era Retrospective analysis for Research and Applications (MERRA-2) (Buchard et al., 2017) reanalysis data at a relatively fine resolution of 0.5° × 0.625 °. The MERRA-2 reanalysis data of mixing ratios for sea salts and dust, along with other hydrophilic and hydrophobic aerosol components, were converted into number concentrations (Lu et al., 2022) and subsequently into CCN emission (Thompson and Eidhammer, 2014). In our analysis, we considered 15 years (2003–2017) of climatological mean CCN concentrations, which provided CCN emissions to the WRF model. To generate spatial heterogeneity in CCN emissions over Chennai, we integrated MERRA2-derived CCN with a spatial pattern obtained from mean 1-km Aerosol Optical Depth data from the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm (Lyapustin et al., 2018). This data, averaged for November-December months of 2014–2016, undergoes Radial Basis Function interpolation (Zou et al., 2013) using the SciPy package (Virtanen et al., 2020). This approach ensures realistic spatial heterogeneity in the magnitude of CCN emissions.

We performed three aerosol sensitivity experiments, with sequentially reducing CCN emission and concentrations (Table S1 in the Supplement). The baseline case (CNTRL) was followed by low-CCN case (LCCN10) with reduction factor of 10 and a very-low-CCN (LCCN100) case at one-hundredth with all other model configurations held constant. The motivation for these aerosol sensitivity experiments is that the climatological CCN values used in the CNTRL scenario (Fig. 1d) might not ideally represent conditions prior to the EPF event in Chennai. In multiple studies, the period from 29 November to 3 December 2015, is identified as the precipitation period for the 2015 event (Konduru et al., 2023; Narasimhan et al., 2016). Thus, by 29 November, the influence of the “below-cloud scavenging” (BCS) mechanism must have resulted in relatively cleaner conditions (compared to the climatology mean). Based on the Laakso et al. (2003) scavenging parameterization, continuous rainfall of 2–4 mm h⁻¹ over 24 h is estimated to reduce accumulation-mode aerosol concentrations by factors of ∼ 26–180, corresponding to 96 %–99 % removal via wet scavenging. Furthermore, several raingauge stations recorded extreme rainfall exceeding 30–40 mm h⁻¹ on 1 December 2015, indicating even higher scavenging efficiency. Thus, the LCCN sensitivity experiments were designed to simulate cleaner, more representative pre-event atmospheric conditions.

To assess the robustness of the aerosol-sensitivity results and to account for IMV within WRF, an ensemble approach was employed at 1-km resolution (d4). Rainfall was simulated for the period 30 November 00:00 UTC to 2 December 00:00 UTC, with three ensemble members generated for each aerosol scenario using different lead times of 27, 24, and 21 h, resulting in nine simulations in total. This ensemble framework (CNTRL-Ens, LCCN10-Ens, LCCN100-Ens) enables a more reliable evaluation of how aerosol loading affects rainfall over the urban basin during the EPF event.

Chennai city is located in the Adyar basin (Fig. 1c), through which the Adyar River flows and discharges into the Bay of Bengal. The hydrological model, HEC-HMS, is set up for the Adyar basin to simulate inflow for different scenarios into the Chembarambakkam reservoir that is located upstream of the city, and also the upstream boundary condition (marked as “hydraulic domain inflow” in Fig. 1c) for the hydraulic model, HEC-RAS. The Curve Number (CN) and kinematic wave routing methods are used to estimate the effective runoff from precipitation and to route the riverine flow in the basin. The CN method utilizes soil and land use/land cover data from the National Bureau of Soil Sciences, Pune, India, and Landsat 8, respectively. The two-dimensional HEC-RAS model routes the inflow hydrographs from the HEC-HMS model through the Adyar River and obtains the flood inundation extents in the city limits. To represent the complex urban terrain of Chennai, the 2 m × 2 m high-resolution LIDAR DEM data for Chennai city is used for hydraulic modeling. For a better representation of flooding in the urban area of Chennai, the building footprints are obtained from the Open Street Map (https://www.openstreetmap.org, last access: 15 January 2026) and according to the Building Block (Brown et al., 2007; Hunter et al., 2008; Schubert and Sanders, 2012) method, the values of DEM pixels corresponding to the footprints are increased by 5 m to maintain no-flow condition through the buildings. More details of the hydrological-hydraulic model set-up and its calibration for the 2015 EPF can be found in earlier studies (Nithila Devi et al., 2019). For the hydrological and hydraulic analyses, which are highly sensitive to rainfall magnitude, a single ensemble member was selected based on its spatio-temporal performance. LES simulations (d5) for rainfall were then conducted for all aerosol scenarios (CNTRL-LES, LCCN10-LES, LCCN100-LES) and were one-way coupled to the hydrologic-hydraulic setup (i.e., HEC-HMS and HEC-RAS). The detailed description of all simulations is provided in Table S1.

Figures 2a–d present the spatiotemporal distribution of ensemble-mean accumulated precipitation for the sensitivity experiments, and its corresponding evaluation against IMERG satellite observations. IMERG revealed that this heavy rainfall was mostly centered over the southern half of the Adyar basin, extending to regions further south (Fig. 2a). The Adyar basin experienced significant rainfall on 1 December, with in situ rain gauge data recorded precipitation of 494 mm at Kattupakkam (ka) and 427 mm at Chembarambakkam (cb), respectively (Fig. 2).

In the CNTRL-Ens case, the maximum accumulated precipitation over the basin is approximately 270 mm, with a significant fraction falling outside the basin boundary in the north near 13.5° N (Fig. 2b). In comparison, LCCN10-Ens simulation demonstrates improvement in spatial distribution, with an increased Adyar basin-maximum of ∼ 295 mm (Fig. 2c). LCCN100-Ens exhibits further enhancement of the ensemble-mean rainfall, producing a maximum precipitation of ∼ 370 mm within the basin (Fig. 2d). The latitudinal rainfall distribution (Fig. 2e) captures this gradual improvement, with the peak rainfall region strengthening over the south of Adyar basin as we move from CNTRL-Ens to LCCN10-Ens to LCCN100-Ens, and the northern peak weakens correspondingly. Figure 2f presents the basin-averaged accumulated precipitation from IMERG along with the ensemble-mean values and associated ensemble spread for all the sensitivity experiments. IMERG indicates a basin-mean accumulation of 334 mm. The ensemble-mean basin-averaged precipitation is 242 mm for CNTRL-Ens, 260 mm for LCCN10-Ens, and 297 mm for LCCN100-Ens. The maximum basin-averaged totals within the ensemble members are 289, 352, and 370 mm for CNTRL-Ens, LCCN10-Ens, and LCCN100-Ens, respectively, demonstrating a systematic enhancement in the representation of rainfall over the Adyar basin across the three scenarios.

At the same time, the temporal variability of precipitation within the Adyar basin is also validated using in situ rain gauges (Fig. 2g). Rainfall records averaged across four gauge stations (ka, cb, mb, and au; locations shown in Fig. 2) indicate that the most intense rainfall phase in the basin occurred between 09:00 and 21:00 UTC. The ensemble-mean temporal profiles from all sensitivity experiments show a distinct bimodal distribution during this period, consistent with the IMERG observations within the basin (Fig. 3a). LCCN100-Ens exhibits the lowest temporal bias (-8.25 mm), followed by LCCN10-Ens (-10.51 mm) and CNTRL-Ens (-11.67 mm) (Fig. 2f). Although all simulations show a bimodal structure, the enhanced performance of LCCN10-Ens and LCCN100-Ens arises from their improved representation of the temporal peaks, particularly the second peak during the period of intense rainfall. The first peak at 09:00 UTC is comparable across simulations, with 3-hourly rainfall values of ∼ 57 mm (CNTRL-Ens), ∼ 46 mm (LCCN10-Ens), and ∼ 42 mm (LCCN100-Ens). However, a clear distinction appears at the second peak around 18:00 UTC, where LCCN10-Ens and LCCN100-Ens (∼ 53 mm) outperform CNTRL-Ens (∼ 38 mm). This advantage persists and increases during later hours, resulting in accumulated ensemble-mean rainfall differences of ∼ 18 mm for LCCN10-Ens and ∼ 55 mm for LCCN100-Ens relative to CNTRL-Ens. Figures S1 and S2 show that these rainfall differences are consistent across all ensemble members. To better understand the spatiotemporal rainfall improvement in LCCN100-Ens over the Adyar Basin, we analyze the aerosol-induced changes in simulated storm morphology and cloud microphysical characteristics relative to CNTRL-Ens during the entire rainfall spell on 1 December (Fig. 3).

IMERG surface rainfall observations on a latitudinal-temporal axes highlights a northward propagation of the storm during this event, with peak precipitation over the Adyar basin at 06:00 and 15:00 UTC on 1 December (Fig. 3a). Both CNTRL-Ens (Fig. 3b) and LCCN100-Ens (Fig. 3c) successfully capture this northward shift. However, LCCN100-Ens better represents the rainfall magnitudes in and around the basin with peak rainfall observed during 09:00–12:00 and 18:00–21:00 UTC, contributing to relatively improved simulation of the observed bimodal temporal pattern in IMERG. In contrast, CNTRL-Ens simulates rainfall more to the north of the basin, with maxima near 13.5° N at 12:00 and 21:00 UTC, consistent with the accumulated pattern shown in Fig. 2b and e.

Comparison of altitude-averaged (surface-500 hPa) cloud water content (CWC) over the basin is consistently higher in CNTRL-Ens than in LCCN100-Ens throughout the event (Fig. 3d and f). During 00:00–06:00 UTC on 1 December, CNTRL-Ens simulates CWC values of ∼ 0.12–0.15 g kg⁻¹, nearly twice that in LCCN100-Ens (∼ 0.06–0.09 g kg⁻¹). This contrast continues as the storm enters the basin during 06:00–09:00 UTC and continues through the later stages of the event (18:00–21:00 UTC). The enhanced CWC in CNTRL-Ens relative to LCCN100-Ens is accompanied by comparatively lower rainwater content (RWC), which remains around 0.3 g kg⁻¹ with localized higher values in the basin at 06:00 and 15:00 UTC (Fig. 3e). In contrast, LCCN100-Ens exhibits values exceeding 0.4 g kg⁻¹ gradually propagating into the southern part of the basin by 09:00 UTC. A similar evolution was observed in later hours, with higher RWC south of the basin at 15:00 UTC and subsequent expansion into the basin between 15:00–18:00 UTC (Fig. 3g).

As expected, lower CCN concentrations in LCCN100-Ens simulates larger droplets and enhance collision-coalescence (Sarangi et al., 2018b). This accelerates raindrop growth and supports the early RWC enhancement observed during 06:00–09:00 UTC as the storm approaches the southern basin boundary. In contrast, CNTRL-Ens exhibits a larger CWC to RWC ratio under higher aerosol loading, favoring northward propagation of convection during the same period. In LCCN100-Ens, the southward enhancement of RWC during 06:00–09:00 UTC intensifies cold pool activity around the basin, strengthening cold pool-driven updrafts and moisture content (Fig. S3). These processes collectively maintain peak rainfall within and near the basin through 12:00 UTC without any northward movement.

The intensified rainfall at 12:00 UTC further amplifies cold pool strength and moisture-updraft interactions around the basin. At 12:00 UTC, the spatial distribution of cold pool intensity (colored shading indicates cold pool regions; Fig. 3h) shows a domain-wide band of strong cold pool signatures between 79.6–79.8° E in LCCN100-Ens, whereas CNTRL-Ens concentrates stronger cold pools farther north. Consequently, LCCN100-Ens produces greater moisture-flux convergence in and around the basin at 12:00 UTC (Fig. 3k), while CNTRL-Ens exhibits higher convergence primarily north of the basin (Fig. 3j). This setting facilitates RWC and surface rainfall south of the basin at 15:00 UTC in LCCN100-Ens, followed by northward progression into the basin by 18:00 UTC as cold pool strength remains elevated (Fig. 3i). Moisture convergence correspondingly intensifies around the basin at 18:00 UTC (Fig. 3m), in contrast to the more scattered and weaker signals in CNTRL-Ens (Fig. 3l). These processes collectively explain the stronger rainfall during 18:00–21:00 UTC in LCCN100-Ens. Overall, the reduced CCN concentrations in LCCN100-Ens enhances microphysics-dynamics interaction in the region, improving the basin's rainfall representation.

Hydrological and hydraulic simulations are carried out to determine whether the improved spatio-temporal rainfall distribution from the aerosol-sensitivity experiments are reflected into corresponding changes in reservoir inflow and flood simulations. Since the Adyar basin is ungauged, hydrological simulation errors are evaluated by comparing simulated inflows into the Chembarambakkam (cb) reservoir (Fig. 1c) and the Chennai city (marked as “hydraulic domain inflow” in Fig. 1c) with observed inflows during the 2015 EPF (details in Nithila and Kuiry, 2025). Among the WRF ensemble simulations, the 24-h lead-time member (CNTRL-24h, LCCN100-24h, LCCN10-24h) is selected based on the best spatio-temporal performance in 1-km runs (Figs. S1–S2; Table S2). Aerosol-sensitivity LES simulations are then conducted using the same model configuration. These simulations improve both the basin-wide cumulative rainfall on 1 December and the temporal evolution of rainfall for LCCN10-LES and LCCN100-LES relative to the corresponding 1-km simulations (LCCN10-24h and LCCN100-24h), with LCCN100-LES showing the closest alignment with rain gauge observations (Fig. 4a).

The improved rainfall representation influences the cumulative inflow volume into the cb reservoir (Fig. 4b) and Chennai city (Fig. 4c). For the cb reservoir, cumulative inflow errors decreased to 0.18 % and 2.40 % in the LCCN10-LES and LCCN100-LES, respectively, while errors for Chennai city reduced to 20.95 % and 8.74 %. These improvements are substantial compared with errors of 20.5 % and 25.38 % obtained in the CNTRL-LES. The timing of peak inflows is also more accurately captured in the LCCN experiments during the EPF event (Fig. S4). Such improvements in inflow prediction had an influence in the downstream flooding during the event. The inundation simulated based on IMERG rainfall is closer to that from the rain gauge data (Nithila Devi et al., 2020) and is considered as the reference for evaluating the inundation extents from CNTRL-LES, LCCN10-LES and LCCN10-LES rainfall inputs till 2 December 00:00 UTC. Although we were unable to secure satellite-derived inundation extents, the hydraulic simulation was validated using post-flood survey depths collected closer to the Adyar River. Therefore, it can be used as a reference for accuracy assessment (Nithila Devi et al., 2020). Figure 4d showcases the confusion matrix of the area and depth distribution for the experiments. It categorizes the simulated results as correctly simulated (wet in both observed and simulated), underpredicted (wet in observed but dry in simulated), and overpredicted (dry in observed but wet in simulated) relative to the reference in the corresponding computational cells. LCCN10-LES and LCCN100-LES exhibit flood extents of 10.33 and 12.07 km² respectively demonstrating the closest alignment with the reference maximum inundation area of 15.37 km². These represent ∼ 29 % and ∼ 51 % improvements over the CNTRL-LES's 8.01 km² simulated flood extent. We have quantified the accuracy of predicted flood inundation extent using fitness index scores (Kuiry et al., 2010). It reported values of 0.52 for CNTRL-LES, 0.67 for LCCN10-LES, and 0.78 for LCCN100-LES. These results emphasize the enhanced accuracy of LCCN experiments in simulating flood inundation depth and extent.