Industrial organisations involved

ARIANET is a consultancy company whose purpose is to contribute to the understanding of processes involving pollutants in the atmosphere.

ARIANET’s tools are aimed at meteorological reconstruction, treatment of emissions and simulation of processes involving pollutants in the atmosphere: transport and diffusion, chemical transformation and deposition on the ground.

ARIANET supports public and private clients in following fields: impact assessment of industrial plants, implementation of air quality forecasting and assessment systems, source apportionment analysis.

Technical/scientific Challenge:

The pollution inside urban area is determined by following contributions:

  • regional, due to the sources present in the surrounding areas (substantially uniform over a large portion of the agglomeration);
  • urban, relating to all sources of pollution within the city whose distribution can be considered uniform (e.g. building heating);
  • street-level, highly inhomogeneous, exceeding the previous two (due to traffic emissions and to the chemical-physical processes that take place inside the road canyons).

The first two contributions can be simulated by coupling atmospheric chemistry models and Machine Learning models. The third one can be evaluated by means of micro-scale dispersion models.

The Solution

  • QualeAria-Local consists of three elements: QualeAria Air Quality Forecast System (AQFS)
  • Random Forest Machine Learning (ML-RF) algorithm
  • μ-scale Parallel-Micro-Swift-Spray (PMSS) modelling suite.

The PM10, PM2.5, NO2 and O3 concentration fields produced by Qualearia for a test period (year 2021), together with a set of spatial-temporal predictors and available air quality observations were processed by the ML-RF algorithm to obtain corresponding higher resolution (1 km) concentration over the national territory. The resulting concentration fields provide the regional and urban background fields to the PMSS suite which is used to produce forecasts of atmospheric particulate matter and nitrogen oxides at very high spatial resolution (4 m) over a selected urban area (Milan conurbation). Figure 1 provides a schematic representation of QualeAria-Local.

Business impact:

The project confirmed the potential of ML models in the field of air pollution, to produce air quality forecasts at high spatial resolutions with a lower computational burden than that required for Chemical Transport Model (CTM) simulations.

The concentration fields produced by the CTM and the seasonal indicators are here the most important spatial-temporal predictors considered by the ML algorithm used (Random Forest).

The performance of the ML model highlighted, in some cases, a reduction of the systematic errors produced by the CTM, without worsening the results obtained.

The project confirmed the need for HPC resources to further increase the spatial resolution of air quality forecasts in urban areas. The constituent elements of QualeAria-Local are computationally intensive and largely already parallelized. For example, a 26-hour simulation, carried out using the PMSS modelling suite, on a domain of dimensions equal to 10×11 km2, which includes the urban area of Milan, at a resolution of 4 m. (2500×2750 grid points) required a calculation time equal to about 90 minutes using 391 cores on the Galileo100 HPC infrastructure.

Estimating the levels of pollution near roads and inside road canyons is important as in these areas
some pollutants reach particularly high concentration levels and can be harmful to the exposed

The benefits

  • Identification of the computational resources needed to evaluate the air quality within the Milan urban area up to the street level.
  • Procurement of forecasts of air quality over a urban area at a horizontal resolution of 4 m.
  • Procurement of regional air quality assessments at 1 km horizontal resolution for air quality assessment and for epidemiological surveys.

Images Courtesy: ARIANET