Can AI really forecast India’s air quality this accurately?
A peer-reviewed model reported an R² of 0.9994 on 435,742 historical records, but the target was a three-pollutant AQI proxy dominated by one of its inputs, severe episodes were scarce and the data ended in 2015. The result is promising method work, not evidence that today’s official AQI can be forecast almost perfectly.
The short version
- What happened
- The study used 435,742 daily CPCB records from 1990–2015 and a chronological 70/15/15 split, combining a 14-day LSTM with XGBoost before retraining on 10 SHAP-selected features.
- Who it affects
- Better forecasts could support earlier health and school precautions, but the published system was not tested in those decisions.
- What remains unproven
- The target was a three-pollutant historical AQI proxy rather than India’s current eight-pollutant official index.
Keep in mind
The target was a three-pollutant historical AQI proxy rather than India’s current eight-pollutant official index.
RSPM correlated 0.99 with the derived target, creating a strong definitional relationship between an input and the outcome.
The records ended in 2015, several modern pollutants and meteorological variables were absent, and missingness reached 54% in one column.
Poor, Very Poor and Severe observations were scarce, and winter peaks were underpredicted by four to eight AQI units.
Validation was internal; there was no prospective deployment, independent modern city network or measured public-health outcome.
Source analysis · 1 linked source. Primary analysed source date: 7 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The direct answer: the model was accurate on its proxy, not proven on today’s AQI
The reported score is real within the study’s test design, but it is narrower than the headline number suggests. The final ensemble achieved a root mean squared error of 0.3817, mean absolute error of 0.2773 and R² of 0.9994 on a held-out chronological test set. Yet the model did not predict the full eight-pollutant National Air Quality Index used by India today. It predicted a historical proxy calculated from three pollutants available consistently in records spanning 1990 to 2015: respirable suspended particulate matter, sulphur dioxide and nitrogen dioxide.
That proxy was itself strongly determined by one of the input variables. The paper reports a 0.99 correlation between RSPM and the target AQI because the index is defined as the maximum of pollutant sub-indices and RSPM dominated most observations. The authors ran control experiments, including removing RSPM, and still reported R² of 0.9851. That supports genuine pattern learning, but it does not remove the central circularity risk: forecasting a constructed index from the measurements used to construct it is easier than forecasting future pollution from independent upstream signals.[1]
The dataset was large, historical and heavily incomplete
The researchers used 435,742 daily measurements from Central Pollution Control Board monitoring stations across Indian states, with particularly dense coverage in Maharashtra, Uttar Pradesh, Andhra Pradesh, Punjab and Rajasthan. The records contained SO₂, NO₂, RSPM and suspended particulate matter. Missingness was substantial: 28% for SO₂, 31% for NO₂, 41% for RSPM and 54% for SPM. The gaps were not random; they were associated with station age and monsoon months, when outages were more common.
To fill gaps, the team used linear interpolation plus Gaussian noise scaled to 2% of each column’s observed standard deviation, followed by non-negative clipping. The method preserved aggregate variance closely and was cheaper than several comparison imputers. But interpolation cannot recreate an unobserved pollution spike inside a long outage. The authors flag gaps longer than 30 days, about 3.8% of imputed values, and acknowledge that the injected noise prevents unrealistically flat segments without recovering the missing physical dynamics.[1]
Two complementary models were retrained after explanation
The pipeline engineered 26 predictors, including short lags, rolling summaries, calendar variables and pollutant ratios. One branch was a two-layer long short-term memory network receiving a 14-day sequence. The other was XGBoost, a tree ensemble suited to nonlinear interactions in tabular data. Their predictions were combined using weights based on validation R². The measured residual correlation of 0.31 suggested the branches made meaningfully different errors rather than duplicating one another.
The unusual step was to use SHAP explanations from XGBoost as a feedback mechanism. The ten highest-ranked features were retained, and both models were trained again on that reduced set. The authors report that this cut input dimensionality by 62% and reduced RMSE by a further 38% compared with the unreduced ensemble. Gradient sensitivity rankings for the LSTM broadly agreed with the tree explanations, with a Spearman correlation of 0.87 and the same top three variables.[1]
The validation was careful, but mostly internal
The paper did more than report one favourable split. Models were retrained across 30 random seeds, and paired t-tests and Wilcoxon signed-rank tests were applied with Bonferroni correction. A five-fold rolling-origin evaluation preserved the model ranking. The final system also outperformed a Temporal Fusion Transformer in this dataset. Monte Carlo dropout produced nominally well-covered prediction intervals, while residual checks found little remaining autocorrelation apart from heavier tails at extreme AQI values.
These are useful safeguards against random-seed luck and an arbitrary split. They remain internal validation on the same historical source and target definition. The model was not tested prospectively on incoming measurements, on an independent modern city network or on the full pollutant mix used in current official reporting. Statistical significance across repeated runs cannot establish operational reliability when the target, population and sensors change.[1]
The cases that matter most were the least represented
Air-quality warnings matter most during severe episodes, yet the test data contained few observations in the Poor, Very Poor and Severe bands; the paper says the Severe category had roughly 196 test samples. The model systematically underpredicted winter peaks by about four to eight AQI units. The authors attribute that gap to missing meteorological variables such as wind speed, temperature, boundary-layer height and humidity, as well as episodic sources that the historical pollutant series alone cannot explain.
The dataset also predates routine coverage of PM2.5, carbon monoxide, ozone, ammonia and lead at most stations. RSPM is treated as an approximate historical proxy for PM10, although measurement protocols differ. A model that works on these records may require different features and calibration in a PM2.5-dominated city, an ozone-dominated region or a network with newer instruments. The architecture may transfer; the published performance number should not.[1]
What this could mean for people
A fast, transparent forecast could help hospitals anticipate respiratory demand, schools decide whether to limit outdoor activity and authorities target inspections or temporary controls. The study reports CPU inference of 7.9 milliseconds per sample and offers feature-level explanations. Those properties could make the method practical as one component in an operational service, especially where compute and data quality are constrained.
But no hospital, school, regulator or community decision was tested. An error concentrated in rare pollution peaks could be more harmful than a larger average error spread across clean days. Before people rely on the tool, evaluators would need to report threshold-specific sensitivity, missed-alert rates, false alarms, calibration by city and season, and performance when sensors fail in the patterns seen during real emergencies.[1]
What would change the assessment
The decisive next test is a genuinely forward-looking evaluation on post-2015 CPCB measurements using the official eight-pollutant AQI, with co-located weather and emissions information. Results should be separated by city, season and severity band, compared with persistence, deterministic AQI calculations, established statistical forecasts and modern deep-learning baselines. A locked model should then be run prospectively before any deployment claim is made.
External replication in other countries would show whether SHAP-guided retraining is the transferable contribution or whether the exceptional score depended mainly on this proxy target. The authors report no research funding, no competing interests and university support for the publication charge. For now, the study is best read as a well-tested modelling pipeline and a warning about missing historical pollution data—not as proof that India’s current public-health AQI can be predicted almost perfectly.[1]