Attributing aerosol impacts on large-scale climate using detection and attribution and machine learning methods

Abstract

While greenhouse gases trap (outgoing) longwave radiation, heating the climate system, aerosol-associated cooling results from aerosols’ interactions with (incoming solar) radiation and clouds, referred to as aerosol-radiative (ARIs) and aerosol-cloud interactions (ACIs). The most recent IPCC report (IPCC, 2021a) assesses “the likely range of total human-caused global surface temperature increase" since the mid-twentieth century as 0.8◦C to 1.3◦C where “well-mixed GHG contributed a warming of 1.0◦C to 2.0◦C" while “other human drivers (principally aerosols) contributed a cooling of 0.0◦C to 0.8◦C". This shows that while there is confidence in the effects of greenhouse gases, aerosols impacts are still not fully understood and cause much larger uncertainties. These observational limitations of aerosols have led to large uncertainties in aerosol forcing, the corresponding responses in climate models and the quantification of aerosol contributions to observed climate changes. Thus, to improve the prediction of future temperatures where greenhouse gas (GHG) emissions will likely continue to increase, while future global aerosol emissions remain uncertain but are mostly likely decreasing, we can use past (historical) impacts of these two major climate forcers. However, this requires an understanding and quantification of the individual climate contributions by aerosols and GHG to the observed climate change, which is hindered by large uncertainties in aerosol forcing and responses across climate models. To estimate the historical aerosol impact, I apply detection and attribution methods to attribute a joint change in temperature and precipitation by combining signals of observed changes in tropical wet and dry regions, the interhemispheric temperature asymmetry, global mean temperature (GMT) and global mean land precipitation (GMLP) in Chapter 2. Fingerprints representing the climate response to aerosols (AER) and the remaining external forcings (noAER; mostly GHG) are derived from large-ensembles of historical single- and ALL-forcing simulations from three models in phase 6 of the Coupled Model Intercomparison Project and selected using a perfect model study. To test this method, I conduct perfect and imperfect model studies, and a hydrological sensitivity analysis, which support combining my choice of temperature and precipitation fingerprints into a joint study. I find that diagnostics including temperature and precipitation slightly better constrain the noAER signal than diagnostics based purely on temperature or GMT-only and allow for the attribution of AER cooling (even when GMT is not included in the fingerprint). These results are robust for fingerprints from different climate models. Estimated contributions for AER and noAER agree with estimates from the most recent IPCC report and I attribute a best estimate of 0.46 K ([-0.86, -0.05] K) of aerosolinduced cooling and of 1.63 K ([1.26, 2.00] K) of noAER warming in 2010-2019 relative to 1850-1900 using the combined signals of GMT and GMLP.

In Chapter 3, causal inference methods are used to investigate aerosol impacts on the diurnal temperature range (DTR), the difference between daily maximum (Tₘₐₓ) and minimum temperature (Tₘᵢₙ), which can provide additional information on climate change impacts related to the diurnal cycle and temperature extrema. Aerosols have been proposed as a main driver of the European DTR by reducing Tₘₐₓ and increasing Tₘᵢₙ, but limited understanding of aerosol effects have led to large uncertainties in the aerosol radiative forcing and its impact on the European DTR. From the causal analysis I can clearly identify aerosols as a main driver of the European DTR mediated through their effects on clouds and incoming shortwave radiation (SW). From the estimated total causal effect of aerosols on SW, I can further reconstruct the aerosol effective radiative forcing on surface SW (ERFₛ, ₛᵥᵥ) in Europe ([-1.9, -1.8] Wm⁻²). Evaluating the cloud and microphysical parameterisations of two CMIP6 models, I find that causal effects for aerosols and clouds are generally weaker in models than in observations. For the DTR, this is compensated by a higher historical AOD in the climate models, leading to similar reconstructed aerosol impacts while the modelled ERFₛ, ₛᵥᵥ estimates are weaker than those from observations ([-1.4, -1.0] Wm⁻²).

A climate change-induced rise in the frequency and size of wildfires in the western US has dramatically increased wildfire emissions over the past two decades. Because observational studies of ACIs are impacted by confounding and retrieval biases, “natural experiments”, such as wildfires, offer an opportunity to study ACIs while accounting for these limitations. Here, the unperturbed state (i.e., wildfire-free) is known to some extent and only minimal changes in meteorology occur. Based on this idea, in Chapter 4 I attempt to circumvent some of the existing limitations by introducing two approaches to apply a deep learning architecture to predict changes in cloud fraction. In the first approach I train a score-based latent diffusion model (LDM) to directly predict changes in cloud fraction as a function of reanalysis AOD and relative humidity. Although the model is not fully trained yet, my results suggest a wildfiredriven increase in cloud fraction at the location of the wildfire, and downwind thereof due to the atmospheric transport of the smoke. In a second approach the LDM is applied for data fusion to create a high spatio-temporal aerosol product to be able to study smoke cloud interactions using high resolution geostationary satellite observations of clouds. Preliminary results of merging a highly spatially resolved satellite product of UV aerosol index and a highly temporally resolved reanalysis product of aerosol optical depth, show the potential of the approach. The merged aerosol product resolves the temporal dispersion of the wildfire smoke.

In summary, this thesis promotes the importance of aerosols and their significant impact on the climate system by considering their influence on temperature and precipitation mediated through aerosol-cloud interactions. Especially, my research on drivers of the DTR and wildfire impacts highlights the influence aerosols have on clouds. I highlight that causal inference applications, when carefully designed, are possible to disentangle complex relationships in the Earth System and to attribute observed changes to environmental and anthropogenic drivers. The application of a deep learning architecture to study aerosol effects presents a promising approach but also shows that these models are sensitive to the scaling of the input data and require careful (often time intensive) training.