From data to decisions: how neural network input structures propagate to air quality policies

Abstract

In air-quality planning, Artificial Neural Networks (ANNs) have emerged as surrogate models capable of capturing complex nonlinear relationships between emissions and pollutant concentrations with reduced computational time required by chemical and transport models. However, the extent to which surrogate model design influences policy optimization outcomes remains poorly investigated. This study analyzes how alternative ANN architectures affect both predictive performance and the identification of cost-effective air-quality policies at the regional scale. A set of ANN architectures differing in their spatial aggregation and weighting strategies is identified and applied within an Integrated Assessment Model aiming at designing efficient Air Quality policies for the Po Valley basin. The ANNs are trained by processing a set of scenarios simulated by a deterministic chemical-transport model and validated against a real scenario. Such ANNs have been implemented within a multi-objective optimization framework to identify cost-effective emission-reduction policies. The results show differences in ANN performance across different configurations. The largest difference appears between the single basin ANN, which achieves a relative mean absolute error of 6.4%, and regionalized ANNs, whose errors range from 2.4–3.3%. This difference propagates in the optimization results. At low implementation costs (50 M€–150 M€ per year), air-quality outcomes are comparable across architectures, with policies targeting three key macrosectors: domestic heating, transport, and agriculture, though with different resource allocation. At higher cost levels, regionalized architectures outperform the basin-wide model, highlighting how surrogate model design can influence both the fidelity of the simulation and the structure of optimal policy portfolios.