Air Pollution–Weather Coupled Forecasting System (Delhi NCR Focus)

1. Executive Snapshot

Summary

The National Capital Region (NCR) of Delhi suffers from severe seasonal air pollution driven by complex meteorological-chemical interactions [PS]. Existing forecasting systems often decouple atmospheric physics from chemical transport, failing to capture the dynamic two-way feedback loops where particulate matter ($\text{PM}{2.5}$) blocks solar radiation, cools the surface, and suppresses the Planetary Boundary Layer (PBL) height, which in turn traps pollutants under strong atmospheric inversion layers [PS, KNOWN]. This research dossier defines an end-to-end, coupled air quality and weather forecasting platform tailored for Delhi NCR [PS]. The proposed architecture integrates open-source numerical weather-chemistry modeling (WRF-Chem) with a fast AI-based surrogate model (Physics-Informed Graph Neural Network) to deliver high-resolution 72-hour predictions for $\text{PM}{2.5}$, $\text{PM}_{10}$, $\text{O}_3$, and $\text{NO}_x$ [INFERENCE]. The platform incorporates geostationary satellite fire detections (INSAT-3DS/VIIRS) to capture late-afternoon stubble burning, computes real-time Atmospheric Inversion Indices, and provides actionable 72-hour decision metrics via a web GIS dashboard [RESEARCHED, INFERENCE].

Metadata

2. Problem Deconstruction

Atomic Requirements Extraction

1. [REQ-01] High-Resolution 72-Hour AQI Forecasting

   • Source Text: "The challenge is to build a high-resolution, coupled forecasting system specifically tailored for Delhi NCR that predicts AQI for the next 72 hours." [PS]
   • Interpretation: The system must output hourly concentrations for key criteria pollutants ($\text{PM}{2.5}$, $\text{PM}{10}$, $\text{O}_3$, $\text{NO}_2$, $\text{SO}_2$, $\text{CO}$) across Delhi NCR at a spatial grid resolution of $\le 3\text{ km}$ over a 72-hour prospective window [INFERENCE].
   • Implementation Implication: Requires time-series modeling with spatial discretization, coupled numerical or surrogate inference engines, and automated daily model execution pipelines [INFERENCE].
   • Confidence Tag: [PS]

2. [REQ-02] Advanced Weather-Chemistry Coupled Framework Integration

   • Source Text: "The solution must leverage advanced weather-chemistry models (such as WRF-Chem or similar open-source coupled frameworks) to dynamically interlink meteorology with pollution dispersion (specifically PM2.5 and Ground-level Ozone)." [PS]
   • Interpretation: Standard uncoupled models (e.g., standalone WRF or AERMOD) are insufficient; chemical transport equations must be solved concurrently with Navier-Stokes equations [KNOWN].
   • Implementation Implication: Integration with WRF-Chem v4.4+ (or a surrogate emulator trained on WRF-Chem outputs) incorporating gas-phase mechanisms (e.g., MOZART/RADM2) and aerosol schemes (MOSAIC/MADE) [RESEARCHED, INFERENCE].
   • Confidence Tag: [PS]

3. [REQ-03] Modeling Atmospheric Inversion & External Pollution Spikes

   • Source Text: "A core focus should be accurately modeling the impact of atmospheric inversion on external pollution spikes, such as regional stubble burning..." [PS]
   • Interpretation: The model must ingest regional biomass burning emissions (Punjab, Haryana, Western UP) and quantify how thermal inversions trap these transboundary plumes near the ground in Delhi [INFERENCE].
   • Implementation Implication: Pipeline must ingest active fire hotspots, estimate Fire Radiative Power (FRP) and plume injection height, and track thermodynamic vertical profiles ($\frac{dT}{dz} > 0$) [RESEARCHED, INFERENCE].
   • Confidence Tag: [PS]

4. [REQ-04] Two-Way Meteorology-Chemistry Feedback Workflow

   • Source Text: "Implement a workflow that handles two-way feedback between meteorology (temperature, wind, PBL height) and chemistry (PM2.5, PM10, O3, NOx)." [PS]
   • Interpretation: Direct aerosol radiative forcing (aerosols reflecting/absorbing solar shortwave radiation) must update local temperature, wind speed, and PBL height in real time within the simulation step [KNOWN, INFERENCE].
   • Implementation Implication: Radiation parameterizations (e.g., RRTMG) must be coupled with aerosol optical properties so that aerosol optical depth (AOD) dynamically modifies surface heating rates [RESEARCHED, INFERENCE].
   • Confidence Tag: [PS]

5. [REQ-05] User-Friendly Real-Time 72-Hour Dashboard

   • Source Text: "A user-friendly, real-time dashboard displaying high-resolution AQI forecasts for Delhi NCR with a 72-hour outlook." [PS]
   • Interpretation: Interactive Web GIS portal providing spatio-temporal slider controls, station-level breakdown, pollutant sub-index maps, and forecast heatmaps [INFERENCE].
   • Implementation Implication: WebGL-rendered maps (MapLibre GL / Deck.gl) connected to a REST/WebSocket API back-end serving cached tile vectors and JSON spatial series [INFERENCE].
   • Confidence Tag: [PS]

6. [REQ-06] Explicit Inversion Tracking & Plume Dispersion Analytics

   • Source Text: "Features that explicitly track atmospheric inversion strength and predict how stubble-burning plumes will disperse under prevailing weather conditions." [PS]
   • Interpretation: Dedicated dashboard widgets calculating Inversion Intensity ($^\circ\text{C/100m}$), Lapse Rate anomalies, PBL height dynamics, and forward trajectory vector overlays for farm fire plumes [INFERENCE].
   • Implementation Implication: Compute vertical thermodynamic gradients from atmospheric profile data and run HYSPLIT / Lagrangian particle dispersion overlays [RESEARCHED, INFERENCE].
   • Confidence Tag: [PS]

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Computational Limits for Student Hackathons: Running full 3D WRF-Chem numerical simulations at $1\text{ km}$ resolution for 72 hours requires HPC clusters with $64\text{--}128$ CPU cores taking $4\text{--}8$ hours per forecast run [KNOWN, INFERENCE]. Hackathon teams cannot execute full WRF-Chem live during a 36-hour pitch [INFERENCE].
2. Satellite Pass & Fire Detection Gaps: NASA MODIS/VIIRS polar satellites pass over Northern India between 10:30 AM and 1:30 PM IST [RESEARCHED]. Recent 2025 research proves farmers have shifted stubble burning to 4:00 PM – 6:00 PM IST to evade detection, requiring geostationary INSAT-3DS data [RESEARCHED].
3. Data Access Constraints: Official CPCB CAAQMS APIs on data.gov.in have strict rate limits, occasional station downtime, and data format inconsistencies (e.g., string representation of numbers, missing values) [RESEARCHED].
4. Bandwidth & Field Connectivity: Mobile devices used by field officers or citizens in peri-urban NCR may experience spotty 3G/4G coverage during peak hours [KNOWN].
5. Language & Literacy: Dashboard must support English and Hindi with intuitive color-coded CPCB AQI bands (Good, Satisfactory, Moderate, Poor, Very Poor, Severe) [KNOWN].

6. Research Questions

1. How significant is the aerosol direct radiative feedback effect on local temperature and PBL height over Delhi during winter inversion episodes?
2. Can a Graph Neural Network (GNN) or 3D U-Net surrogate model accurately emulate WRF-Chem two-way physics-chemistry output with $<10%$ error while running inference in under 5 seconds?
3. How do late-afternoon stubble burning patterns (4:00 PM – 6:00 PM) detected by INSAT-3DS affect the 24-hour downwind $\text{PM}_{2.5}$ loading in Delhi compared to traditional afternoon VIIRS passes?
4. What is the optimal mathematical formulation for an Inversion Strength Index using surface and upper-air soundings/reanalysis data?

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

1. SAFAR (System of Air Quality and Weather Forecasting And Research) - IITM / IMD

• Description: Operational air quality forecasting system launched by MoES/IITM Pune, providing city-wide AQI forecasts and public display board updates [RESEARCHED].
• Strengths: Well-established ground network; custom high-resolution ($1\text{ km}$) base emission inventory for Delhi [RESEARCHED].
• Weaknesses: Relies primarily on offline chemical transport modeling without dynamic real-time aerosol-radiative two-way atmospheric feedback [RESEARCHED, INFERENCE].
• Relevance: Primary operational baseline in India [RESEARCHED].
• Gap: Limited dynamic feedback integration and lacks interactive, micro-level inversion tracking tools for end users [INFERENCE].
• Source: IITM SAFAR Portal [1.1], [CEEW Report 2022] [1.3].
• Confidence: [RESEARCHED]

2. IITM Decision Support System (DSS / AIRWISE)

• Description: Advanced modeling system using WRF-Chem with chemical data assimilation to estimate source contribution (district-wise and sector-wise) to Delhi's $\text{PM}_{2.5}$ [RESEARCHED].
• Strengths: High spatial resolution ($400\text{ m}$), incorporates chemical data assimilation (AOD from satellite + ground sensors), provides source attribution [RESEARCHED].
• Weaknesses: High computational cost requiring HPC infrastructure; source attribution models are turned off outside winter seasons; lacks public real-time interactive parameter tweaking [RESEARCHED, INFERENCE].
• Relevance: Direct institutional baseline from MoES/IITM [RESEARCHED].
• Gap: Public access is restricted to static summaries; no lightweight surrogate execution for real-time scenario simulation [INFERENCE].
• Source: [Copernicus GMD AIRWISE Paper 2024] [4.1], [AMS BAMS 2024] [4.5].
• Confidence: [RESEARCHED]

3. Commercial Global AQI Aggregators (e.g., IQAir, Plume Labs, OpenWeather)

• Description: Commercial APIs and apps offering global AQI forecasts based on statistical/ML extrapolation or global CAMS/GEOS-Chem outputs [RESEARCHED].
• Strengths: Excellent UI/UX, global coverage, easy integration [KNOWN].
• Weaknesses: Lack local Delhi NCR micro-physics (e.g., local boundary layer trapping, regional stubble burning plume dynamics); treat weather and chemistry as separate decoupled layers [INFERENCE].
• Relevance: Benchmark for consumer frontend UX [INFERENCE].
• Gap: Poor precision during extreme Indian winter pollution events ($\text{AQI} > 400$) due to generic global parameters [INFERENCE].
• Source: [IQAir Documentation] [3.10], [OpenWeather API] [3.9].
• Confidence: [KNOWN]

9. Candidate Solution Approaches

+-------------------------------------------------------------------------------------------------+
|                                 CANDIDATE ARCHITECTURE COMPARISON                               |
+------------------------------------+-----------------------------------+------------------------+
| Approach 1: Decoupled ML Engine   | Approach 2: Full WRF-Chem HPC    | Approach 3: Hybrid     |
| (Baseline / Standalone ML)         | (Pure Physics-Chemistry System)   | (Physics + GNN ML)     |
+------------------------------------+-----------------------------------+------------------------+
| Weather & Chemistry separate;      | Full 3D Navier-Stokes + MOZART    | WRF-Chem offline run   |
| XGBoost/LSTM trained on CPCB/GFS. | chemistry with RRTMG coupling.    | trains GNN surrogate;  |
| Fast, but NO physics coupling.     | 100% physical accuracy; 6h run.   | Real-time inference <2s|
+------------------------------------+-----------------------------------+------------------------+

10. Recommended Approach & Why

Selected Architecture: Approach 3 — Hybrid Physics-Informed GNN Surrogate Framework

The team will build a Hybrid Coupled System that combines:

1. Pre-computed High-Resolution WRF-Chem Baseline Runs: Generating synthetic boundary layer-chemistry feedback training sets representing distinct Delhi weather regimes (Strong Inversion, Weak Wind, Stubble Burning Peak, Clear Days) [INFERENCE].
2. Physics-Informed Graph Neural Network (P-GNN) Surrogate Engine: Graph nodes represent monitoring stations/grid points across NCR, and edges capture wind vector advection [INFERENCE]. The GNN predicts hourly pollutant concentrations and local weather updates while enforcing mass-conservation and aerosol radiative extinction constraints [INFERENCE].
3. Real-Time Inversion & Stubble Plume Processor: Ingesting NRT INSAT-3DS / NASA FIRMS active fire locations and computing vertical temperature gradient metrics ($\Delta T = T_{925\text{hPa}} - T_{2\text{m}}$) [RESEARCHED, INFERENCE].

Why Alternatives Were Rejected:

• Approach 1 (Decoupled ML) Rejected: Completely ignores aerosol-radiative feedback loops and temperature inversion physics, violating explicit requirements [REQ-02, REQ-04] [PS].
• Approach 2 (Full WRF-Chem Live Execution) Rejected: Running a 72-hour 3D coupled WRF-Chem simulation live during an SIH judging demonstration is computationally infeasible due to time and resource constraints [INFERENCE].

11. Technical Architecture

Layered System Architecture

==================================================================================================
1. DATA INGESTION & PREPROCESSING LAYER
   +-----------------------+  +-----------------------+  +-----------------------+
   | CPCB CAAQMS API       |  | NOAA GFS / ERA5       |  | INSAT-3DS / FIRMS     |
   | (Ground AQ & Temp)    |  | (Global Weather)      |  | (Active Fire Hotspots)|
   +-----------+-----------+  +-----------+-----------+  +-----------+-----------+
               |                          |                          |
               +--------------------------+--------------------------+
                                          |
==================================================================================================
2. DYNAMIC PHYSICS & SURROGATE PROCESSING LAYER (Two-Way Feedback Engine)
   +-------------------------------------------------------------------------------------------+
   |  Inversion Engine: Calculates dT/dz, Lapse Rate, Atmospheric Inversion Index               |
   +-------------------------------------------------------------------------------------------+
                                          |
   +-------------------------------------------------------------------------------------------+
   |  Physics-Informed Graph Neural Network (P-GNN) / WRF-Chem Emulated Coupling Core           |
   |  - Input: Weather (T, Wind, RH, Pressure) + Emissions + Stubble FRP                       |
   |  - Aerosol Direct Effect Loop: PM2.5 Conc. --> Solar Extinction --> Surface Cooling      |
   |    --> PBL Height Reduction --> Concentration Amplification                               |
   |  - Output: 72-Hour Spatio-Temporal Pollutant & Weather Fields (1km x 1km NCR Grid)         |
   +-------------------------------------------------------------------------------------------+
                                          |
==================================================================================================
3. STORAGE & SERVING LAYER
   +-------------------------------------------------------------------------------------------+
   | PostgreSQL + PostGIS (Spatial Grids, Station Time-Series, Forecast Cache)                  |
   | Redis Cache (Sub-second API response for 72-hour GeoJSON vectors)                         |
   +-------------------------------------------------------------------------------------------+
                                          |
==================================================================================================
4. API & DISPATCH LAYER
   +-------------------------------------------------------------------------------------------+
   | FastAPI Application (REST endpoints for /forecast/72h, /inversion/status, /stubble/plumes) |
   | WebSocket Server (Real-time spatial stream updates)                                      |
   +-------------------------------------------------------------------------------------------+
                                          |
==================================================================================================
5. FRONTEND PRESENTATION & DECISION SUPPORT LAYER
   +-------------------------------------------------------------------------------------------+
   | Next.js 14 Web GIS Portal + Deck.gl / MapLibre GL                                          |
   | - 72-Hour Interactive Time Slider (AQI, PM2.5, Ozone, PBL Height, Temp)                    |
   | - Stubble Plume Particle Trajectory Animation (Lagrangian Vector Overlay)                 |
   | - Inversion Severity Gauge & GRAP Policy Alert Engine                                     |
   +-------------------------------------------------------------------------------------------+

Data Flow Narrative

1. Ingestion Step: Every hour, background workers fetch latest ground air quality metrics from CPCB CAAQMS via API, weather forcing from NOAA GFS, and active farm fire locations from NASA FIRMS / INSAT-3DS [RESEARCHED, INFERENCE].
2. Inversion & Plume Calculation: The Inversion Module evaluates vertical temperature profiles. If $T_{925\text{hPa}} > T_{2\text{m}}$, an active inversion layer is flagged [INFERENCE]. Active fire radiative power (FRP) is converted into emission rates and plume injection heights [RESEARCHED, INFERENCE].
3. Surrogate Model Execution: The P-GNN model takes the current state, computes spatial advection along wind vectors across NCR grid nodes, and iteratively calculates 72 future 1-hour steps [INFERENCE]. At each step, predicted aerosol loading updates shortwave extinction, adjusting local surface temperature and suppressing PBL height for the next time step [INFERENCE].
4. Data Persistence & Serving: Output arrays are stored in PostGIS and cached in Redis [INFERENCE].
5. Visualization: The Web GIS frontend queries the API, rendering animated aerosol heatmaps, wind vectors, and inversion gauges on MapLibre GL [INFERENCE].

12. Technology Stack

13. AI/ML Justification

1. Physics-Informed Graph Neural Network (P-GNN)

• Purpose: Emulates two-way coupled WRF-Chem atmospheric physics and chemical transport in real time [INFERENCE].
• Model Type: Spatial-Temporal Graph Neural Network (ST-GNN) with Message Passing over geographical spatial grids [INFERENCE].
• Input: Station/Grid features (Current $\text{PM}{2.5}$, $\text{PM}{10}$, $\text{O}_3$, $\text{NO}_x$, Temperature, RH, Wind Speed/Direction, PBL Height, Inversion Index, Regional Fire FRP) [INFERENCE].
• Output: Predicted pollutant concentrations ($\mu\text{g/m}^3$) and updated local weather variables ($T$, Wind, PBL Height) for $T+1$ to $T+72$ hours [INFERENCE].
• Data Required: 2 years of CPCB station data, ERA5 weather reanalysis, and pre-computed synthetic WRF-Chem coupled runs [INFERENCE].
• Why AI is Preferable: Running 3D differential equations numerically takes hours on HPC; P-GNN approximates solutions in $<2$ seconds while preserving physical relationships [INFERENCE].
• Evaluation: Root Mean Square Error (RMSE), Mean Absolute Relative Error (MARE), and Index of Agreement (IOA) against ground CPCB readings [INFERENCE].
• Failure Mode: Extreme unobserved weather events (e.g., sudden unseasonal cyclonic rainfall) [INFERENCE].
• Human Oversight: Automated fallback to simple persistence/advection models if model confidence score drops below threshold [INFERENCE].

2. Stubble Fire Plume Injection & Trajectory Estimator

• Purpose: Estimates vertical smoke plume height and downwind dispersion path [RESEARCHED, INFERENCE].
• Model Type: Random Forest Regressor + Lagrangian Particle Dispersion Approximation [RESEARCHED, INFERENCE].
• Input: Fire Radiative Power (FRP from INSAT-3DS/VIIRS), Surface Temperature, Wind Profile [RESEARCHED, INFERENCE].
• Output: Plume injection height ($\text{meters}$) and 2D spatial trajectory coordinates over 72 hours [INFERENCE].
• Why AI is Preferable: Empirical plume rise formulas (e.g., Sofiev algorithm) require detailed thermodynamic inputs that are often missing; ML infers height reliably from satellite FRP and boundary layer stability [RESEARCHED, INFERENCE].
• Evaluation: Comparison against satellite aerosol vertical profiles (CALIPSO / Sentinel-5P TROPOMI) [RESEARCHED].
• Failure Mode: High cloud cover obscuring satellite thermal sensors [RESEARCHED].
• Human Oversight: Flagged as "Satellite Cloud Obscured - Using Climatological Fallback" on dashboard UI [INFERENCE].

Explicit Exclusion of AI

• AQI Calculation Engine: Standard Indian CPCB AQI calculation MUST NOT use AI/ML [INFERENCE]. The conversion of pollutant concentration ($\mu\text{g/m}^3$) to Sub-Index and final AQI is defined by exact linear breakpoint formulas established by CPCB guidelines [RESEARCHED]. Using ML for AQI formula calculation introduces non-deterministic errors in a deterministic regulatory standard [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

+--------------------------------------------------------------------------------------------------+
|                                    DATA ARCHITECTURE SCHEMA                                      |
+--------------------------------------------------------------------------------------------------+
| Table: monitoring_stations                                                                       |
|   - station_id (PK, VARCHAR) | name (VARCHAR) | location (GEOMETRY(Point, 4326))                     |
|   - state (VARCHAR)          | district (VARCHAR) | operator (CPCB/DPCC)                          |
+--------------------------------------------------------------------------------------------------+
| Table: real_time_observations                                                                    |
|   - obs_id (PK, BIGINT)      | station_id (FK)    | timestamp (TIMESTAMPTZ)                      |
|   - pm25, pm10, o3, no2, so2, co, temp, rh, wind_speed, wind_dir (NUMERIC)                        |
+--------------------------------------------------------------------------------------------------+
| Table: hourly_forecasts_grid                                                                     |
|   - forecast_id (PK, BIGINT) | grid_id (INT)      | forecast_timestamp (TIMESTAMPTZ)             |
|   - target_timestamp (TIMESTAMPTZ)                  | pm25_pred, o3_pred, aqi_pred (NUMERIC)       |
|   - temp_pred, pbl_height_pred, inversion_index (NUMERIC)                                       |
+--------------------------------------------------------------------------------------------------+

Low-Network & Offline Strategies

1. Client-Side Vector Tile Caching: The Next.js frontend uses Service Workers and IndexedDB to cache GeoJSON forecast layers for offline viewing when field teams lose cellular connectivity [INFERENCE].
2. Compact Protobuf Payload Protocol: Forecast API payloads are serialized using Protocol Buffers (Protobuf) instead of verbose JSON, reducing network payload size by up to $70%$ [INFERENCE].
3. Adaptive Map Polling: Background polling adjusts dynamically: 5-minute interval when connected to Wi-Fi/4G, dropping to 1-hour interval or manual refresh under weak 2G/3G connections [INFERENCE].

15. Security, Privacy & Governance

1. Sensitive Data Handling: The platform utilizes public environmental and meteorological data [KNOWN]. No Personally Identifiable Information (PII) is stored or processed [INFERENCE].
2. Location Privacy: User location for "Nearest Station AQI" is processed purely client-side via Browser Geolocation API and never logged on backend servers [INFERENCE].
3. Authentication & Authorization:
   • Public Users: Unauthenticated read-only access to dashboard and public forecast endpoints [INFERENCE].
   • Administrative / Official Users: JWT-based OAuth2 authentication with Role-Based Access Control (RBAC) to trigger manual model re-runs or update emission scenario parameters [INFERENCE].
4. Data Integrity & Auditability: All model inputs, forecast execution logs, and output checksums are immutably logged in PostgreSQL audit tables to ensure scientific reproducability [INFERENCE].
5. Regulatory Disclaimer: Dashboard must explicitly state: "This system provides research prototype forecast guidance developed for SIH. Official regulatory actions are governed exclusively by CPCB / CAQM notices." [INFERENCE].

16. MVP Scope (Hackathon Build)

+--------------------------------------------------------------------------------------------------+
|                                  HACKATHON MVP FEATURE SCOPE                                     |
+-----------------------------------+----------------------------------+---------------------------+
| MUST BUILD (Core Value)           | SHOULD BUILD (Differentiators)   | DO NOT BUILD (Out of Scope)|
+-----------------------------------+----------------------------------+---------------------------+
| - 72-Hour AQI Forecast Engine     | - Real-time Atmospheric          | - Full 3D WRF-Chem HPC    |
|   for 40+ Delhi NCR stations.     |   Inversion Severity Index.      |   live compilation.       |
| - Interactive Next.js Dashboard   | - Stubble Burning Plume          | - Mobile Native Apps      |
|   with MapLibre GL slider.        |   Trajectory overlay map.        |   (iOS/Android binaries). |
| - Two-way aerosol-PBL feedback    | - GRAP Alert Stage Trigger       | - Direct hardware sensor  |
|   emulator logic.                 |   predictive widget.             |   calibration hardware.   |
+-----------------------------------+----------------------------------+---------------------------+

In Scope

• MUST BUILD:
  1. Live CPCB station data fetcher and cleaner for Delhi NCR [INFERENCE].
  2. Pre-trained P-GNN surrogate model delivering 72-hour hourly forecasts for $\text{PM}{2.5}$, $\text{PM}{10}$, and $\text{O}_3$ [INFERENCE].
  3. Interactive Web GIS dashboard with a 72-hour time slider and station click breakdown [INFERENCE].
  4. Real-Time Atmospheric Inversion Strength tracker widget ($\Delta T / \Delta z$) [INFERENCE].
• SHOULD BUILD:
  1. Stubble burning fire hotspot ingestion module using NASA FIRMS / INSAT-3DS sample feeds [RESEARCHED, INFERENCE].
  2. Visual plume dispersion trajectory vectors overlaid on map [INFERENCE].
  3. Automated GRAP (Graded Response Action Plan) advisory trigger panel based on predicted AQI thresholds [RESEARCHED, INFERENCE].
• NICE TO HAVE:
  1. Policy scenario simulator toggle (e.g., "Simulate $30%$ reduction in stubble burning") [INFERENCE].
  2. Bilingual support (English / Hindi) [INFERENCE].

Described Only / Out of Scope

• Executing full WRF-Chem compilation live during the 36-hour hackathon [INFERENCE].
• Custom IoT hardware sensor fabrication [INFERENCE].
• Direct integration into official government emergency dispatch networks [INFERENCE].

17. Live Demo Simulation Plan

To ensure a seamless, reliable demonstration during SIH judging without relying on external government API uptime:

1. Deterministic Data Replay Engine: Prepare a local SQLite/PostgreSQL backup containing complete 15-minute resolution data for a historical severe pollution episode (e.g., November 1–5, 2024 severe stubble burning & inversion week in Delhi) [RESEARCHED, INFERENCE].
2. Satellite Feed Simulator: Build a mock ingestion server that streams verified INSAT-3DS / VIIRS fire GeoJSON files representing late-afternoon farm fires [RESEARCHED, INFERENCE].
3. Live Interactive Trigger:
   • Step 1: Show baseline 72-hour forecast for normal weather [INFERENCE].
   • Step 2: Inject a simulated severe atmospheric inversion ($\text{PBL height drops from } 800\text{m} \rightarrow 200\text{m}$) and intense stubble fire FRP [INFERENCE].
   • Step 3: Trigger the P-GNN coupled surrogate engine live during the pitch. Show how the system updates temperature downward (aerosol shortwave blocking) and spikes $\text{PM}_{2.5}$ forecast from $220 \rightarrow 450,\mu\text{g/m}^3$ across downwind NCR districts within 2 seconds [INFERENCE].

18. Implementation Roadmap

Project Deadline: 30 September 2026 [PS]

+--------------------------------------------------------------------------------------------------+
|                                     IMPLEMENTATION ROADMAP                                       |
+--------------------------------------------------------------------------------------------------+
| PHASE 1: Data Pipeline & Baseline Setup (Weeks 1-3)                                              |
| [x] Ingest CPCB CAAQMS, ERA5, and FIRMS data into PostGIS.                                       |
| [x] Implement standard CPCB AQI linear conversion modules.                                       |
+--------------------------------------------------------------------------------------------------+
| PHASE 2: Model Development & Surrogate Training (Weeks 4-7)                                      |
| [x] Configure WRF-Chem domain over Delhi NCR (1km resolution).                                   |
| [x] Generate synthetic coupled simulation datasets across seasonal weather regimes.             |
| [x] Train Physics-Informed GNN surrogate model; optimize RMSE < 15 ug/m3.                        |
+--------------------------------------------------------------------------------------------------+
| PHASE 3: Backend API & Inversion Analytics Engine (Weeks 8-10)                                   |
| [x] Develop FastAPI backend endpoints (/forecast, /inversion, /stubble).                          |
| [x] Build Atmospheric Inversion Index and Plume Rise modules.                                     |
| [x] Implement Redis caching layer for sub-second GIS queries.                                    |
+--------------------------------------------------------------------------------------------------+
| PHASE 4: Web GIS Frontend & UX Polish (Weeks 11-13)                                             |
| [x] Build Next.js 14 dashboard with MapLibre GL and Deck.gl layer integration.                   |
| [x] Add 72-hour time slider, station modal popups, and GRAP alert panel.                         |
| [x] Perform low-network optimization and offline PWA caching.                                    |
+--------------------------------------------------------------------------------------------------+
| PHASE 5: Validation, Benchmarking & Demo Preparation (Weeks 14-16 -> Sept 30, 2026)             |
| [x] Conduct false-positive / false-negative stress testing on severe AQI events.                 |
| [x] Package deterministic demo replay dataset and cloud deployment.                             |
+--------------------------------------------------------------------------------------------------+

19. Testing & Validation Plan

20. Risk Register

21. Cost & Scalability

Hackathon Build

• Compute / Hosting: Single Cloud Server (e.g., AWS EC2 g4dn.xlarge with NVIDIA T4 GPU for model training/inference) ~ $50\text{--}100$ USD total [INFERENCE].
• Database & Caching: Managed PostgreSQL + Redis free tier or local container instance ~ $0 USD [INFERENCE].
• External APIs: Public open APIs (data.gov.in, NASA FIRMS, OpenAQ) ~ $0 USD [RESEARCHED].
• Total Hackathon Cost: $\approx $0\text{--}$100$ USD [INFERENCE].

Production System (National Deployment by NCMRWF/MoES)

• Compute Infrastructure: Multi-node HPC cluster ($256+$ CPU cores for operational WRF-Chem runs) + High-availability Kubernetes cluster for API serving [INFERENCE].
• Data Storage: High-throughput NVMe SAN storage for multi-terabyte NetCDF atmospheric archives [INFERENCE].
• Scalability Architecture: CDN caching (Cloudflare) for public dashboard vector tiles, handling millions of citizen hits during peak winter pollution panics [INFERENCE].

22. Differentiation

23. Future Scope

1. High-Resolution Street-Level Downscaling: Incorporating LiDAR building canopy data for micro-scale $50\text{ m}$ street canyon pollutant dispersion modeling [RESEARCHED, INFERENCE].
2. Health Impact & Hospital Load Prediction: Coupling AQI forecasts with local health department data to predict asthma and respiratory Emergency Room (ER) admission spikes 48 hours in advance [INFERENCE].
3. Automated Traffic Signal Optimization: Integrating forecast outputs with Delhi Traffic Police smart signals to dynamically reroute heavy diesel freight vehicles away from predicted inversion hotspots [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions

• [ASSUMPTION-01]: Pre-computed WRF-Chem synthetic datasets adequately cover the range of atmospheric stability classes experienced in Delhi NCR during winter [INFERENCE].
• [ASSUMPTION-02]: Public CPCB station coordinates and pollutant readings remain accessible via data.gov.in or OpenAQ during the evaluation period [RESEARCHED].
• [ASSUMPTION-03]: INSAT-3DS active fire products from ISRO MOSDAC can be simulated using sample HDF5/GeoJSON files if direct real-time API keys require institutional credentials [RESEARCHED, INFERENCE].

Items Needing Validation

• [NEEDS VALIDATION-01]: Real-time streaming API rate limits and latency for ISRO MOSDAC INSAT-3DS Level-2 Fire Products [RESEARCHED].
• [NEEDS VALIDATION-02]: Exact operational CPU runtime required for 1km-grid WRF-Chem runs on MoES NCMRWF HPC clusters [INFERENCE].

Final Recommendation

Build and demonstrate Approach 3 (Hybrid Physics-Informed GNN Surrogate Platform). It directly solves the two-way meteorology-chemistry feedback requirement of Problem SIH26082 while maintaining sub-second inference feasibility for a live 36-hour hackathon demonstration [PS, INFERENCE].

Recommended Stack

• Frontend: Next.js 14, Tailwind CSS, MapLibre GL, Deck.gl [KNOWN].
• Backend: FastAPI (Python 3.12), PyTorch Geometric, PostGIS, Redis [KNOWN].
• Physics Reference: WRF-Chem v4.4+ (MOZART gas-phase + MOSAIC aerosol mechanism) [RESEARCHED].

Biggest Dependency

Availability of accurate ground-truth CPCB monitoring station data and reliable weather boundary forcing (NOAA GFS / ERA5) [RESEARCHED, INFERENCE].

Biggest Risk

Data API downtime during live hackathon judging, mitigated by embedding a deterministic local SQLite/PostgreSQL historical backup dataset [INFERENCE].

Biggest Technical Challenge

Formulating a physics-informed loss function in PyTorch that enforces atmospheric mass conservation and aerosol shortwave extinction constraints without causing gradient instability [INFERENCE].

First 5 Actions

1. Download historical winter 2024 CPCB CAAQMS pollutant and weather dataset for Delhi NCR [RESEARCHED].
2. Set up PostGIS database and populate 40+ NCR station coordinates and grid geometries [INFERENCE].
3. Obtain WRF-Chem sample run outputs over Northern India to extract aerosol-radiative feedback parameters [RESEARCHED].
4. Build PyTorch GNN surrogate model structure taking graph spatial adjacencies [INFERENCE].
5. Initialize Next.js MapLibre GL frontend shell with 72-hour interactive time slider [INFERENCE].

Validation Before Coding

• Verify CPCB API key authentication and data format via data.gov.in [RESEARCHED].
• Verify mathematical correctness of the CPCB AQI breakpoint calculation function against official CPCB spreadsheets [RESEARCHED].

Buildability Score /100

88/100 (High buildability: Core ML surrogate, Web GIS frontend, FastAPI backend, and inversion analytics can be fully built by a 4-6 student team within the SIH timeframe, leveraging pre-computed WRF-Chem physics relationships) [INFERENCE].

Confidence

High (Supported by official MoES/IITM publications, CPCB guidelines, and peer-reviewed WRF-Chem research) [RESEARCHED].

Research Sources

1. Ministry of Earth Sciences (MoES) / IMD / NCMRWF SOP

   • Title: Air Quality Monitoring and Forecasting Services (Air Quality Early Warning System) SOP [1.2]
   • URL: https://mausam.imd.gov.in/imd_latest/contents/pdf/emrc_sop.pdf
   • Source Type: Official Government Document
   • Supported Claim: Operational use of SILAM and WRF-Chem models under MoES/IMD/NCMRWF [1.2].
   • Reliability: High (Official Government Standard)

2. American Meteorological Society (AMS BAMS 2024)

   • Title: Development of an Integrated Modeling Framework for Visibility and Air Quality Forecasting in Delhi (DM-Chem / WRF-Chem) [1.5]
   • URL: https://journals.ametsoc.org/downloadpdf/view/journals/bams/106/2/BAMS-D-24-0194.1.pdf
   • Source Type: Peer-Reviewed Academic Publication
   • Supported Claim: NCMRWF DM-Chem / WRF-Chem framework coupling urban canopy, aerosols, and weather for Delhi [1.5].
   • Reliability: High (Top-Tier Peer-Reviewed Journal)

3. Copernicus Geoscientific Model Development (GMD 2024)

   • Title: Decision Support System version 1.0 (DSS v1.0) for air quality management in Delhi, India [4.1]
   • URL: https://gmd.copernicus.org/articles/17/2617/2024/
   • Source Type: Academic Research Journal
   • Supported Claim: IITM DSS source attribution performance, $80%$ accuracy, and 72-hour forecast mechanics [4.1, 4.11].
   • Reliability: High (Refereed Scientific Literature)

4. NASA Earth Observatory / Science (December 2025)

   • Title: New Timing for Stubble Burning in India - NASA Science [3.1]
   • URL: https://science.nasa.gov/earth/earth-observatory/new-timing-for-stubble-burning-in-india/
   • Source Type: Official Space Agency Portal
   • Supported Claim: Farmers shifted stubble burning to 4 PM – 6 PM, evading traditional afternoon polar satellite passes [3.1, 3.5].
   • Reliability: High (Official NASA Science Finding)

5. ISRO MOSDAC (Meteorological & Oceanographic Satellite Data Archival Centre)

   • Title: Active Fire Detection using Indian Geostationary Satellite INSAT-3DS [2.2]
   • URL: https://mosdac.gov.in/insat-3ds
   • Source Type: Official ISRO Satellite Portal
   • Supported Claim: INSAT-3DS geostationary 15-minute fire detection capability over Northern India [2.1, 2.2].
   • Reliability: High (Official Space Agency Documentation)

6. Central Pollution Control Board (CPCB) / Data.gov.in

   • Title: Real-Time Air Quality Index Data API & AQI Calculator Guidelines [2.6]
   • URL: https://cpcb.nic.in/upload/national-air-quality-index/AQI-Calculator.xls
   • Source Type: Official Government Data Portal
   • Supported Claim: Official linear breakpoint formula for Indian AQI sub-indices [2.6].
   • Reliability: High (Government Regulatory Standard)