AIRWISE · Model & Methodology

How does the air-quality forecasting model work?

A practical guide to the science behind AIRWISE: the observations it uses, how weather and atmospheric chemistry are modelled, how forecasts are evaluated, and how source information can support decisions.

400 mDelhi inner-domain grid spacing
3 domainsIndia → NCR → Delhi
WRF-ChemCoupled weather and chemistry model
3DVAR–GSIChemical data assimilation

01. The idea in one minute

Air pollution depends on both emissions and the atmosphere. Winds transport pollutants; atmospheric stability and boundary-layer mixing influence dilution; chemical processes transform gases and particles. AIRWISE combines these processes in a numerical modelling system and uses observations to constrain its representation of current conditions.

1

Observe

Use satellite aerosol optical depth (AOD), surface air-quality measurements and information about fire activity.

2

Prepare inputs

Prepare meteorological fields, chemical boundary conditions and emissions for the model domains.

3

Assimilate

3DVAR–GSI combines the model background with observations while accounting for estimated errors.

4

Forecast and communicate

Run nested forecasts, process outputs, evaluate against observations and disseminate forecast products.

AQEWS produces air-quality forecasts and early-warning information. The Decision Support System (DSS) adds model-based source-attribution information and tests selected emission-reduction scenarios. Together, the paper refers to the framework as AIRWISE.

02. Model architecture

The following is a simplified, web-friendly schematic of the workflow described in Figure 1 of the paper. It is an original redraw, not a reproduction of the published figure.

AIRWISE FORECASTING & DECISION-SUPPORT WORKFLOWInput preparation → assimilation → nested forecasts → verification and use METEOROLOGYIITM-GFS initial andboundary conditions OBSERVATIONSMODIS satellite AODSurface PM₂.₅ / PM₁₀ EMISSIONS & FIRESInventories and biomass burningChemical boundary conditions CHEMICAL DATA ASSIMILATION3DVAR · Gridpoint Statistical Interpolation (GSI) D01 · INDIA10 km × 10 km10-day outer forecast D02 · NCR2 km × 2 kmRegional nested forecast D03 · DELHI400 m × 400 mHigh-resolution inner domain AQEWS FORECASTPM₂.₅ · PM₁₀ · AQIForecast maps and time series DSS ANALYSISSource contributionsEmission-reduction scenarios VERIFY & DISSEMINATECompare with observationsBulletins and decision support Simplified schematic based on Ghude et al. (2024), Figure 1 and Section 2.

03. Nested modelling domains

The paper describes three nested domains. Coarser regional simulations provide evolving conditions to finer domains, representing broad transport while adding spatial detail over Delhi.

DomainHorizontal grid spacingPurposeForecast period described
D0110 km × 10 kmIndian subcontinent; outer-domain meteorology and chemistry.10 days
D022 km × 2 kmDelhi–NCR and neighbouring states.Up to 96 hours in model description
D03400 m × 400 mHigh-resolution inner domain over Delhi.Up to 96 hours in model description

Grid spacing describes the model grid. It does not mean every grid cell contains a monitoring station or that every local forecast has equal certainty.

04. Inputs and physical / chemical processes

ComponentRoleConfiguration reported in the paper
MeteorologyDrives transport, mixing and weather-related changes.IITM-GFS; reported at 12.5 km grid spacing with updates every 3 hours.
Satellite AODProvides column-integrated aerosol information.MODIS Terra and Aqua; highest-quality QA flag 3 retrievals.
Surface observationsConstrain near-surface particulate concentrations and support verification.Quality-controlled PM₂.₅ / PM₁₀ observations from monitoring networks.
Anthropogenic emissionsRepresent estimated emissions from human activities.EDGAR-HTAP for D01, TERI for D02 and high-resolution SAFAR inventory for Delhi D03.
Fire emissionsRepresent episodic biomass-burning pollution.FINN-related processing and MODIS / VIIRS fire information, with acknowledged uncertainty.
Chemical boundariesRepresent chemical material entering the model domain.MOZART-4 climatology for the outer domain; chemical output is passed inward through nested domains.

Physical and chemical schemes reported

Morrison double-moment microphysicsRRTMG radiationNoah land surfaceMonin–Obukhov surface layerMYNN 2.5 boundary layerMOZART gas chemistryGOCART aerosols

The paper reports WRF-Chem version 3.9.1. It notes that the GOCART configuration used does not simulate nitrate and secondary organic aerosols (SOA), in part because of operational computational constraints.

05. What is chemical data assimilation?

Models begin from an estimated atmospheric state, but that state contains errors. Data assimilation combines the model background with observations, accounting for estimated background and observation errors. AIRWISE uses the 3DVAR method within the Gridpoint Statistical Interpolation (GSI) system.

Model background

The model's prior estimate of aerosol and related atmospheric conditions.

Observational constraint

Satellite AOD and surface particulate measurements provide complementary evidence about aerosols.

Important distinction: Satellite AOD is not the same quantity as surface PM₂.₅. The assimilation framework relates the model aerosol state to AOD and also uses surface observations to constrain near-surface pollution.

06. Forecasting versus source attribution

AQEWS: What may happen?

Forecasts pollutant concentrations, including PM₂.₅ and PM₁₀, and supports early identification of deteriorating air quality.

DSS: Why, and what could change?

Uses source tracers to estimate represented source contributions and test selected emission-reduction scenarios.

Source groups described in the paper

The DSS uses more than 400 three-dimensional passive tracers to represent source groups and reduction scenarios. The paper reports a 10 km horizontal grid for DSS and daily production of source-attribution and scenario information for the next five days.

How to interpret scenarios: A simulated 40% emissions reduction does not guarantee a 40% reduction in ambient PM₂.₅. Concentrations depend on chemistry, transport, weather and whether the assumed emissions change occurs.

07. Forecast verification and reported performance

Forecasts are evaluated against observations using concentration metrics and event-detection metrics. Each measures a different aspect of forecast quality.

Forecast dayRMSE (µg m⁻³)IOAMFBNMB
Day 171.230.720.04+1.35%
Day 273.760.710.01−1.32%
Day 3101.240.63−0.12−13.22%

These study-period mean PM₂.₅ forecast statistics are reported in Table 1 of Ghude et al. (2024). The paper reports stronger overall performance on Days 1–2 and moderate performance on Day 3. They are not a guarantee for an individual forecast.

Metrics explained

08. Limitations and uncertainty

Forecasts are evidence-based estimates, not certainties. The paper highlights several challenges, particularly during extreme pollution episodes.

Responsible use: Read forecast trends alongside the latest measured air quality. For operational decisions, follow current official bulletins and directions from relevant authorities.

09. Frequently asked questions

Is this an AI-only forecasting model?
No. The paper describes a numerical weather–chemistry modelling framework based on WRF-Chem, combined with chemical data assimilation. It does not describe the core forecast as an AI-only model.
What is WRF-Chem?
A coupled numerical model that simulates weather-related processes and atmospheric chemistry and aerosols within the same modelling framework.
Does 400 m mean every street has a separate measurement?
No. It is the horizontal grid spacing of the inner model domain. It does not mean every street has a monitoring station or that every grid-cell forecast has equal certainty.
How far ahead are forecasts available?
The paper describes a 10-day forecast for the outer 10 km domain and forecasts up to 96 hours for the 2 km and 400 m domains. Its high-resolution performance analysis focuses on a three-day forecast period.
Does a 40% emissions cut produce a 40% drop in PM₂.₅?
Not necessarily. DSS tests specified emission-reduction scenarios and estimates the resulting modelled change. Ambient concentrations respond to transport, chemistry, weather and interactions among sources.
Where can I read the original research?
Ghude et al. (2024), “Air Quality Warning and Integrated Decision Support System for Emissions (AIRWISE): Enhancing Air Quality Management in Megacities,” Bulletin of the American Meteorological Society. Read the paper via DOI.

Scientific reference

Ghude, S. D., et al. (2024). “Air Quality Warning and Integrated Decision Support System for Emissions (AIRWISE): Enhancing Air Quality Management in Megacities.” Bulletin of the American Meteorological Society. DOI: 10.1175/BAMS-D-23-0181.1.

Operational dissemination website identified in the paper: ews.tropmet.res.in.

Scope note: This page is an educational summary of the cited paper. Model configurations and performance values are attributed to that publication and may not describe subsequent operational updates.