AQBench: Benchmarking Neural Surrogates for Air Quality Forecasting

AQBench supports the evaluation of neural surrogate models for air quality forecasting over the Indian subcontinent. It combines hourly simulations from the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem), meteorological and emissions fields, and gridded ground station observations. Both data products use a 140 × 124 grid with 25 km × 25 km cells.

The simulation release contains 32 stored months across 2016–2019. The 2019 simulation file ends in May. Coverage is discontinuous, and some stored months have missing hours. Use the stored timestamps to identify gaps and construct contiguous forecasting windows.

Prediction targets

Task	Targets
Single pollutant	PM₂.₅
Multiple pollutants	PM₂.₅, NO₂, CO

Dataset structure

AQBench/
├── wrfchem_simulations/
│   ├── wrfchem_2016.h5
│   ├── wrfchem_2017.h5
│   ├── wrfchem_2018.h5
│   └── wrfchem_2019.h5
└── ground_station_observations/
    ├── ground_station_2018.h5
    └── ground_station_2019.h5

WRF-Chem simulations

wrfchem_YYYY.h5
├── feature_names       # (26,), in stored channel order
├── audit_json          # Build audit, including missing hours
└── month_XX/
    ├── data            # (timesteps, 140, 124, 26), float32
    └── time            # (timesteps,), one timestamp per data row

Feature	Description
pm25, no2, co	Simulated pollutant concentrations used as prediction targets
q2, t2	Water vapor mixing ratio and temperature at 2 m
u10, v10	Eastward and northward wind components at 10 m
rainc, rainnc	Convective and nonconvective precipitation
swdown	Downward shortwave flux at the surface
pblh	Planetary boundary layer height
psfc	Surface pressure
NOx_e, NH3_e, PM25_e, NMVOC_e, SO2_e	Anthropogenic emissions by species
NOx_finn, NH3_finn, PM25_finn, NMVOC_finn, SO2_finn	Fire emissions by species
bio	Biogenic isoprene emissions
Kyy	Diffusivity coefficient used in the advection–diffusion residual
no, pm10	Additional stored concentrations excluded from the modeling inputs and targets

Ground station observations

The station files contain hourly PM₂.₅ observations from 55 monitoring stations, mapped to the simulation grid. Use station_mask to identify station cells and account for missing observations when evaluating predictions.

ground_station_YYYY.h5
├── feature_names       # (1,), pm25
├── station_mask        # (140, 124), bool
└── month_XX/
    ├── data            # (timesteps, 140, 124, 1), float32
    └── time            # (timesteps,), one timestamp per data row

Read a simulation month

import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="m3rg-iitd/AQBench",
    repo_type="dataset",
    filename="wrfchem_simulations/wrfchem_2018.h5",
)

with h5py.File(path, "r") as file:
    features = file["feature_names"].asstr()[:].tolist()
    month = file["month_01"]
    times = month["time"].asstr()[:]
    pm25_first_hour = month["data"][0, :, :, features.index("pm25")]