neuralsorcerer/air-quality

Synthetic Kolkata Air Quality & Meteorology 100M

A reproducible fully synthetic spatiotemporal benchmark inspired by broad Kolkata, West Bengal climatological and air-quality behavior. The release contains exactly 100,000,000 station-hour rows from 1,000 explicitly synthetic sensor sites.

This is not an official CPCB/WBPCB/IMD monitoring archive and is not a reconstruction of historical measurements. Synthetic coordinates, event effects and station classes are benchmark constructs.

Intended use

Suitable for:

• multivariate time-series forecasting
• graph/spatiotemporal forecasting
• missing-data imputation
• anomaly/event detection
• synthetic time-series benchmarking
• robustness experiments with variable-specific sensor outages

Not suitable for:

• epidemiological exposure estimates
• legal/regulatory compliance
• historical claims about actual Kolkata pollutant concentrations
• identifying real monitoring stations
• treating event multipliers as measured causal effects

Dataset grain and coverage

Property	Value
Grain	one synthetic site × one hourly timestamp
Sites	1,000
Hours/site	100,000
Total rows	100,000,000
Start	2015-01-01 00:00 Asia/Kolkata
End	2026-05-29 15:00 Asia/Kolkata
Train	first 80,000 hours/site = 80,000,000 rows
Validation	next 10,000 hours/site = 10,000,000 rows
Test	final 10,000 hours/site = 10,000,000 rows
Spatial graph	directed 10-nearest-neighbor Haversine graph

The regular hourly grid is always preserved. Missing sensor values remain as nulls rather than deleting timestamps.

Main variables

Field	Unit / encoding	Generation logic
timestamp	timezone-aware	Asia/Kolkata in Parquet export
station_id	uint16	synthetic site ID
temperature_c	°C	monthly climatology + diurnal cycle + persistent synoptic + microclimate
relative_humidity_pct	%	coupled to temperature and rainfall
wind_speed_ms	m/s	seasonal positive-skew process + cyclone gust regime
rainfall_mm	mm/hour	Markov wet/dry process + clustered storms + monthly interannual variability
pm25_ugm3	µg/m³	season + diurnal + ventilation + wet scavenging + site + dynamic spatial field
pm10_ugm3	µg/m³	PM2.5 + independently varying coarse component
no2_ugm3	µg/m³	source class + rush-hour/weekend + ventilation + dynamic combustion field
co_mgm3	mg/m³	combustion-linked but independently perturbed
so2_ugm3	µg/m³	industrial/source effects + ventilation + independent perturbation
o3_ugm3	µg/m³	solar/temperature photochemistry + NO2 titration + separate spatial field
event_code	uint8	0 normal, 1 Diwali stress, 2 lockdown, 3 cyclone, 4 heatwave
obs_mask	uint16	variable-specific observation bit mask
split_code	uint8	0 train, 1 validation, 2 test

See DATA_DICTIONARY.md for the exact bit layout of obs_mask.

Spatial structure

The generator uses 24 time-varying spatial latent fields so geographic proximity carries meaningful statistical structure rather than merely sharing one city-wide signal. Each station mixes its four nearest latent knots with Gaussian distance weights. Separate fields are used for fine particulate matter, coarse material, combustion-related pollutants and ozone.

A final 120-site × 6,000-hour diagnostic on the materialized dataset gives:

PM2.5 spatial diagnostic	Result
distance vs raw pairwise correlation	-0.7125
mean raw correlation <5 km	0.9546
mean raw correlation ≥30 km	0.9133
mean residualized correlation <5 km	0.2610
mean residualized correlation ≥30 km	-0.0570

The exact coefficients are properties of this synthetic design, not empirical claims about a universal Kolkata covariance law. Their purpose is to ensure that geographic proximity carries meaningful statistical information.

Structured, variable-specific missingness

Missingness combines whole-site telemetry outages with channel-specific clustered failures. Full-table rates are:

Variable	Missing rate
temperature	1.860%
relative humidity	2.051%
wind speed	2.427%
rainfall	2.524%
PM2.5	3.215%
PM10	3.772%
NO2	5.503%
CO	6.452%
SO2	7.459%
O3	6.000%

Missing periods occur in contiguous runs. For example, in the validator's 20-site sample, median missing-run lengths range from about 5 to 10 hours depending on channel, with longer high-percentile outages for gaseous-pollutant channels.

Physical and logical validation

The materialized 100M-row table passes all hard validation rules in validation_report.json:

• exact 100,000,000 rows
• 1,000 sites × 100,000 timestamps
• timestamp order consistency = 1.0
• timestamp uniqueness = 1.0
• hourly regularity = 1.0
• grid completeness = 1.0
• zero station/timestamp key errors
• zero train/validation/test structure errors
• zero invalid observation-mask bits
• zero mask/null mismatches
• zero negative pollutant values
• zero invalid temperature/RH/wind/rain range violations
• zero observed rows with PM10 < PM2.5
• released 10-nearest-neighbor graph has exactly 10 outgoing neighbors/site and no self edges

Across the full table:

Variable	Minimum	Mean	Maximum
temperature	7.743 °C	27.089 °C	42.330 °C
RH	30.986%	75.181%	100%
wind	0.030 m/s	1.805 m/s	21.910 m/s
rain	0	0.201 mm/h	69.722 mm/h
PM2.5	1.911	96.075	950 µg/m³
PM10	4.590	140.716	1,200 µg/m³
NO2	3.617	42.810	250 µg/m³
CO	0.170	1.396	6.321 mg/m³
SO2	1.046	9.292	50.318 µg/m³
O3	2.387	26.365	199.632 µg/m³

Safety caps are numerical backstops rather than major distributional drivers: only 15 observed PM2.5 values hit 950, 25 PM10 values hit 1,200, and one NO2 value hit 250 across 100M rows; CO, SO2 and O3 do not hit their upper caps.

The January temperature audit across 20 geographically spread synthetic sites is 8.36–31.35 °C, correcting the physically implausible >40 °C winter behavior found in the older dataset.

Relationship diagnostics

Station-0 full-history pairwise correlations include:

• PM2.5 vs PM10: 0.988
• PM2.5 vs wind: -0.381
• PM2.5 vs rain: -0.141
• NO2 vs O3: -0.212
• NO2 vs CO: 0.818

These are generator diagnostics, not target scientific constants. The validator checks signs and logical relationships rather than claiming exact empirical equivalence.

Temporal fidelity

The generator is not independent row sampling. Persistent city and local AR processes, calendar effects, rainfall episodes, events and spatial latent fields create long-range structure.

Example station-0 autocorrelation:

Variable	lag 1 h	lag 24 h	lag 168 h
temperature	0.983	0.976	0.903
PM2.5	0.967	0.888	0.835
NO2	0.901	0.728	0.681
O3	0.945	0.969	0.948

validation_report.json also stores first-difference summaries, monthly profiles, event-window ratios and spatial-correlation bins.

Synthetic event regimes

event_code supplies controlled stress regimes. Multipliers are deliberately documented as synthetic scenario parameters, not measured causal effects. The benchmark is therefore suitable for algorithmic stress tests but not for estimating historical policy or event effects.

Reproducibility

Generation seed: 20260917.

pip install -r requirements.txt
python generate.py --init --start-station 0 --end-station 1000
python validate.py
python export_parquet.py \
  --input air_quality_100m.npy \
  --stations stations.csv \
  --graph spatial_graph_knn10.csv \
  --out data

Generation is resumable by station range. The canonical NumPy materialization is deterministic for the same code, seed and dependency behavior.

Research rationale

Recent work supports evaluating synthetic sequential data beyond static marginals:

1. Kwon et al. (2026), Seq2Synth, arXiv:2607.15606 shows that static distribution fidelity can hide timestamp and trajectory failures and motivates direct checks of regularity, first differences, autocorrelation and cross-sectional dynamics.
2. Vapsi et al. (2026), Dynamic Linear Coregionalization, arXiv:2604.05064 motivates time-varying and lag-aware multivariate dependence rather than fixed cross-channel correlations.
3. Xu et al. (2026), AirQualityBench, arXiv:2605.05854 emphasizes explicit observation masks, physical-scale metrics, spatial graph structure and distance-aware correlation analysis for realistic air-quality forecasting benchmarks.

Those papers motivate the validation principles, not a claim that this synthetic generator reproduces their datasets.

See RESEARCH_VALIDATION.md for more detail.

Selected domain references

• India Meteorological Department, Kolkata/Alipore climatological and extreme-weather summaries.
• Central Pollution Control Board, National Ambient Air Quality Standards, India.
• Singh, V., Singh, S., & Biswal, A. (2021). Exceedances and trends of particulate matter (PM2.5) in five Indian megacities. Science of the Total Environment 750, 141461. DOI: 10.1016/j.scitotenv.2020.141461.
• Chatterjee, A. et al. (2013). Air Quality during Diwali Festival over Kolkata – A Mega-City in India. Aerosol and Air Quality Research 13, 1133–1144. DOI: 10.4209/aaqr.2012.03.0062.

Important limitation

This release passes internal synthetic-design validation, but that is different from proving empirical equivalence to a real Kolkata monitoring archive. A future external-fidelity study should compare against an authoritative, redistributable real reference set using matched seasonal/diurnal distributions, spectra, autocorrelation, cross-correlation, extreme-event statistics, spatial covariance and train-synthetic/test-real forecasting utility.

Accordingly, the correct description is literature-informed synthetic benchmark, not historical reconstruction or digital twin.

External calibration snapshot

A separate EXTERNAL_CALIBRATION.md compares selected synthetic summaries with published Kolkata evidence. In particular, a 100-site sample has a Jan–Feb PM2.5 mean of about 176.5 µg/m³, while a CPCB report gives 175 µg/m³ for Kolkata winter; the synthetic highest/lowest monthly mean ratio is 5.90, versus 6.98 in a six-year Kolkata study. These are broad plausibility checks only because the periods, stations and aggregation rules differ.

License

CC0-1.0 for the synthetic dataset and release metadata. Third-party references remain under their original terms.