The presence of environmental pollutants beyond desired levels in the atmosphere results in adverse effects on human health, infrastructure and economy. To mitigate this ever-growing problem, it is essential to build robust models using the voluminous data available through state-of-the-art air quality sensors and implement them in stringent policy-making. In this work, optimally designed recurrent neural networks (RNNs) are utilized to capture the nonlinearities of 15 pollutants measured in Taiwan. A novel evolutionary based neural architecture search algorithm balancing the variance-bias trade-off is proposed. While the dynamics in many environmental pollutants (including particulate matter of various sizes, hydrocarbons and Sulfur and nitrogen oxides) were found to be better emulated using univariate RNNs, pH of the rain necessitated the building of multivariate RNN models. Thus, to identify the most potent features effecting concentration of pollutants, Monte Carlo–based Global Sensitivity Analysis (GSA) using the optimally designed RNNs is performed. © 2023 Elsevier Inc. All rights reserved.