Advancing predictions of Dimethylsulfide emissions and biogenic sulfur aerosol in the Mediterranean region via machine learning

Dimethylsulfide (DMS) is the main natural source of atmospheric sulfur and plays a critical role in marine aerosol formation (Mansour et al., 2020b; Mansour et al., 2020a; O'Dowd et al., 2004). It influences cloud radiative forcing, with feedback on regional and global climate (Charlson et al., 1987; Mansour et al., 2022). Despite its importance, the accurate representation of biogenic sulfur emissions in climate models remains a challenge (Mansour et al., 2023; Mansour et al., 2024a). We employed machine learning (ML) based approaches to characterize seawater DMS concentrations, sea-to-air DMS emission flux (F_DMS), as well as the atmospheric concentrations of marine biogenic methanesulfonic acid (MSA) and non-sea-salt sulfate (nss-SO₄^2–). This study focuses on the Mediterranean Sea, a warm, oligotrophic marine basin and a climate change hotspot with rapidly increasing temperatures.

In our methodology, a set of ML models (Mansour et al., 2024b) is trained and evaluated using nested cross-validation, forced by high-resolution satellite data (chlorophyll-a, sea surface temperature, photosynthetically available radiation) and Mediterranean physical reanalysis (mixed layer depth and seawater salinity) datasets, combined with in situ DMS measurements. The optimized model generates daily gridded fields of DMS and F_DMS at mesoscale resolution (0.083° × 0.083°, ~9 km) spanning 23 years (1998–2020). These high-resolution F_DMS estimates align with observational data of MSA and nss-SO₄^2–, secondary aerosol products from DMS oxidation, collected at the Lampedusa monitoring site in the central Mediterranean (Becagli et al., 2013). Compared to existing coarse-resolution global DMS datasets, the reconstructed F_DMS fields capture seasonal patterns of biogenic sulfur with much greater accuracy across the Mediterranean Sea.

Furthermore, the F_DMS outputs are integrated with high-resolution atmospheric datasets from the Copernicus European Regional Reanalysis (CERRA) to predict atmospheric concentrations of MSA and nss-SO₄^2–. The ML models produce daily time-series predictions over the same 23-year period, achieving finer temporal and spatial coverage than observational datasets alone.

This analysis demonstrates the potential of ML techniques to enhance the estimation of seawater DMS fluxes and associated sulfur aerosol concentrations, achieving outstanding predictive performance. The spatiotemporal dynamics of these variables over the 23 years are analysed to elucidate mesoscale oceanographic variability and its influence on sulfur cycling. Ongoing analyses of long-term trends and interannual variability aim to identify the main drivers of these patterns, with results to be presented and discussed in detail.

Funding:

This work was funded by the European Commission’s EU Horizon 2020 Framework program, project FORCeS (grant no. 821205), and the European Union’s Horizon, project CleanCloud (Grant No. 101137639).

References:

Becagli, et al. (2013), Atmospheric Environment, 79, 681-688, 10.1016/j.atmosenv.2013.07.032.

Charlson, et al. (1987), Nature, 326, 655-661, 10.1038/326655a0.

Mansour, et al. (2023), Science of The Total Environment, 871, 10.1016/j.scitotenv.2023.162123.

Mansour, et al. (2024a), npj Climate and Atmospheric Science, 7, 10.1038/s41612-024-00830-y.

Mansour, et al. (2022), Journal of Geophysical Research-Atmospheres, 127, 10.1029/2021jd036355.

Mansour, et al. (2024b), Earth System Science Data, 16, 2717–2740, 10.5194/essd-16-2717-2024.

Mansour, et al. (2020a), Atmospheric Research, 237, 10.1016/j.atmosres.2019.104837.

Mansour, et al. (2020b), Journal of Geophysical Research-Atmospheres, 125, 10.1029/2019jd032246.

O'Dowd, et al. (2004), Nature, 431, 676-680, 10.1038/nature02959.