The research team, led by Mahesh Bankar of Dr. Babasaheb Ambedkar Technological University in Lonere, together with Vinayak Patki, Sachin Pore, and Mithun B. Patil, focused on the three pollutants that dominate the air quality picture at the Solapur Municipal Corporation monitoring station: sulphur dioxide, nitrogen oxides, and respirable suspended particulate matter. These pollutants are the primary inputs used to compute the Air Quality Index in India, a single number that condenses complex atmospheric chemistry into a scale that officials and the public can act upon. Rather than treating the index as a purely statistical artifact, the team set out to learn the underlying nonlinear relationships between pollutant concentrations and the resulting index, relationships that simpler models routinely miss.

The technical pipeline behind the model reflects the messy reality of real-world environmental data. Monitoring stations in developing cities often suffer from gaps caused by instrument downtime, calibration issues, and communication failures, and the Solapur dataset was no exception. The researchers applied interpolation to fill in missing values, a technique that estimates absent readings from the surrounding data points. They then eliminated outliers, the anomalous readings that can arise from sensor malfunctions or transient local events, and normalized the features so that all input variables operated on comparable scales. Each of these preprocessing steps matters: unhandled gaps and outliers can destabilize neural network training, while unscaled features can cause the optimizer to weight certain pollutants disproportionately simply because their units produce larger numbers.

At the heart of the study is a three-layer deep neural network trained with the Adam optimizer, a widely used adaptive gradient descent algorithm that adjusts learning rates individually for each parameter. The network was trained to minimize the mean squared error between its predictions and the observed Air Quality Index values, a standard loss function that penalizes large deviations heavily and thereby pushes the model toward accuracy on the most consequential errors. Three-layer architectures of this kind strike a practical balance: they possess enough depth to capture nonlinear interactions among pollutants, such as the way particulate levels respond differently to precursor gases depending on season and meteorology, while remaining small enough to train efficiently on the modest datasets typical of a single monitoring station.

The performance figures reported by the team are striking. The model achieved a mean absolute error of 3.88 index points and a root mean squared error of 5.92, both small relative to the range of values the Air Quality Index can span. The coefficient of determination of 0.983 indicates that the network’s predictions track the observed values almost perfectly across the evaluation period. For context, an error of under four index points means the model can generally distinguish not just between broad categories such as moderate and poor air quality, but between fine gradations within those categories, which is precisely the level of fidelity needed for meaningful public health advisories.

Equally important is the comparative analysis the researchers conducted against baseline models. Linear regression, the workhorse of classical statistics, assumes that the Air Quality Index responds linearly to pollutant concentrations, an assumption that atmospheric science has repeatedly shown to be false. Support vector regression, a more flexible machine learning method that maps inputs into higher-dimensional spaces, performs better but still struggled to match the deep network. The deep neural network’s superiority over both baselines demonstrates that the relationships embedded in air pollution data are genuinely nonlinear and that architectures capable of representing those nonlinearities deliver tangible gains in predictive skill rather than merely theoretical elegance.

The significance of this work extends well beyond a single city in Maharashtra. Cities across India and the wider developing world operate sparse monitoring networks, often with just one or two reference-grade stations per urban area, and the data those stations produce are frequently incomplete. Sophisticated forecasting systems built on numerical atmospheric models demand extensive computational resources, detailed emissions inventories, and meteorological inputs that many municipalities cannot assemble. A neural network trained on pollutant concentrations from a single station, by contrast, requires only data that municipal corporations already collect and report publicly. The Solapur study used air quality data freely available from the Maharashtra Pollution Control Board, underscoring how accessible such systems have become.

The study also situates itself within a rapidly expanding global literature on machine learning for air quality. Recent years have seen researchers apply recurrent architectures, convolutional networks, hybrid models combining gradient boosting with deep learning, and even quantum-inspired neural networks to pollutant forecasting in cities from Delhi to Kampala. Systematic reviews of the field conclude that artificial intelligence methods now consistently outperform traditional statistical approaches for both monitoring and forecasting tasks. What the Solapur study adds is a demonstration that this revolution reaches down to second-tier cities, not just megacities with dense sensor networks and abundant funding, and that a carefully preprocessed dataset from a single station can support a genuinely high-performing model.

There remain important caveats and directions for future work. A model trained on data from one station captures the pollution dynamics of that location and may not generalize to other neighborhoods or cities without retraining, particularly where emission sources differ, as they do between industrial corridors and residential districts. The study also predicts the index from concurrent pollutant concentrations rather than forecasting it days in advance, and true early-warning systems require lead time. Nonetheless, the framework the researchers describe, from interpolation and outlier removal through normalization, network design, and rigorous evaluation with multiple error metrics, provides a replicable template that other municipalities can adapt to their own monitoring records.

The broader stakes could hardly be higher. Exposure to particulate matter and pollutant gases is linked to respiratory disease, cardiovascular illness, and premature death, and accurate, timely air quality information is a prerequisite for protective measures ranging from school closures to traffic restrictions. If deep learning models of the kind developed for Solapur can be deployed across the hundreds of Indian cities that monitor air quality, they could convert routine regulatory data into actionable forecasts at almost no additional cost. The authors report no funding or conflicts of interest, and they state that the data and code supporting the findings are available from the corresponding author upon reasonable request, an openness that may accelerate exactly that kind of adoption. In a field where the gap between data collection and public benefit has often been wide, this study shows how a modest neural network, trained on freely available measurements, can begin to close it.

Subject of Research: Deep neural network prediction of the Air Quality Index from pollutant concentrations at an urban monitoring station in Solapur, India

Article Title: Deep neural network-based prediction of air quality index (AQI) for municipal corporation station, Solapur, Maharashtra State, India

Article References: Bankar, M., Patki, V., Pore, S., & Patil, M. B. (2026). Deep neural network-based prediction of air quality index (AQI) for municipal corporation station, Solapur, Maharashtra State, India. Theoretical and Applied Climatology, 157(10), Article 651. https://doi.org/10.1007/s00704-026-06588-y

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06588-y

Keywords: air quality index, deep neural network, air pollution, Solapur, Maharashtra, machine learning, particulate matter, sulphur dioxide, nitrogen oxides, environmental monitoring, Adam optimizer, regression baselines