Inverse Modeling of Wildfire Smoke

A climate change-induced rise in the frequency and size of wildfires in the western US has dramatically increased the contribution of wildfire emissions to atmospheric aerosol concentrations over the past two decades. Exposure to biomass burning aerosols (BBAs) can have detrimental outcomes for human respiratory and cardiovascular systems, with a recent study estimating that increased wildfire smoke results in 4000 cases of premature mortality annually across the US. In the climate system, BBAs serve as a radiative climate forcer by impacting Earth's energy budget. Although in the global mean the IPCC has estimated the radiative impact of BBAs to have been around 0 ([-0.2, 0.2] Wm^-2) since pre-industrial times, regionally, wildfire heat and BBAs from wildfires have been found to affect the local meteorology changing the hydrological cycle and enhancing precipitation extremes. Regional radiative impacts of BBAs have further been found to impact modes of internal variability, such as ENSO, thus causing global impacts.

BBAs affect the Earth's radiative balance in different ways. Directly, they interact with incoming short-wave radiation through absorption or scattering (aerosol-radiative interactions, ARIs). Indirectly BBAs change cloud properties, referred to as aerosol-cloud interactions (ACIs), by serving as cloud condensation nuclei (CCN) for the formation of cloud droplets, thus modifying cloud lifetimes, cloud area fraction (CF), cloud albedo and precipitation. Finally, semi-directly rapid adjustments of ARI modify the atmospheric profiles of temperature and relative humidity (RH) also impacting cloud formation and the local moisture availability. BBAs predominantly consist of black carbon (BC) and organic carbon (OC) aerosols which have different radiative properties. While BC strongly absorbs shortwave (SW) radiation, OC mostly scatters SW in the visible spectrum (i.e., ~ 500nm) but can absorb in the UV spectrum. Therefore, the sign of BBA radiative forcing is dependent on BBA composition, location and the underlying surface albedo. Freshly emitted BC and OC are mostly hydrophobic and become hydrophilic during aging, eventually serving as CCN for cloud formation.

Observational studies of ACIs are limited by confounding biases, when looking at small-scale microphysical cloud processes, or the co-variability of aerosols and meteorology, i.e., temperature, specific humidity, turbulence, etc. Additional limitations of satellite-based studies arise from retrieval biases in both aerosols and cloud properties, such as overestimating aerosol optical depth (AOD) due to cloud contamination. Thus, studies have relied on “natural experiments'', such as volcanic eruptions or ship tracks, to study the effects of aerosols on clouds where the unperturbed state is known to some extent and only minimal changes in meteorology occur. Climate models offer an alternative approach to study ACIs, but constraining model estimates is limited due to observational uncertainties. ACIs remain one of largest sources of uncertainty in estimating the aerosol radiative forcing.

This study attempts to circumvent some of the listed limitations for estimating ACIs. Specifically, we explore the application of a score-based latent diffusion model (LDM) to fuse different observational datasets of aerosol loading to generate a spatiotemporal high-resolution aerosol dataset. While reanalysis datasets offer aerosol observations at an hourly resolution, highly spatially resolved aerosol observations from satellites often only provide daily observations and additionally suffer from a selection bias and confounding from aerosol swelling. Thus, by combining the different datasets, uncertainties and biases from either aerosol product can be corrected for. Additionally, for high spatial resolution we rely on highly resolved satellite observations of UV Aerosol Index which could be a more robust proxy of CCN and hence provide more insights into ACIs than AOD does. Such an aerosol dataset holds several benefits as it can leverage the resolution of highly resolved cloud observations because aerosol datasets of the same resolution are currently not available. From the highly-resolved aerosol dataset changes in cloud area fraction due to BBA perturbations after artificially “switching off’’ wildfire emissions can be predicted using, e.g., causal inference methods.

Initial results indicate that the LDM can ``learn'' the spatial structure as a background from UV aerosol index and the temporal aerosol dispersion during inference from the spatially coarse reanalysis AOD.