Stratocumulus adjustments to aerosol perturbations disentangled with a causal approach

Aerosols, airborne particles, modify the planetary radiative budget directly (absorbing or scattering radiation) and indirectly (acting as cloud condensation nuclei (CCN), altering cloud albedo). Between pre-industrial times and 2019, anthropogenic aerosols caused a negative effective radiative forcing (ERF) of −1.1[−1.7; −0.4] W m⁻², offsetting some global warming. However, the exact magnitude of aerosol-induced cooling is uncertain, largely due to the complexities of aerosol-cloud interactions (ACI). Liquid stratocumulus clouds, with their moderately high albedo (~40%) and extensive oceanic coverage, significantly contribute to this uncertainty. In liquid clouds, ERF_ACI combines the instantaneous Twomey effect (increased cloud droplet number concentration (N_d) leading to smaller effective radii (r_eff) and increased albedo at constant LWP) and rapid adjustments of LWP and cloud fraction (CF). The challenge lies in the counteracting effects of aerosol perturbations: reduced r_eff suppresses precipitation, increasing LWP and albedo, while simultaneously, enhanced cloud-top entrainment of warm, dry air leads to evaporation and LWP reduction. These processes occur simultaneously and involve feedback loops, making disentanglement difficult. The sign and magnitude of LWP adjustments vary with temporal and spatial scales, cloud regimes, and environmental conditions, hindering causal interpretation of correlations. Confounding variables (e.g., environmental factors influencing both aerosol properties and cloud properties) further complicate analysis. While randomized controlled trials are impractical, opportunistic experiments (ship tracks, volcanic eruptions) and climate model simulations offer alternatives, but suffer from limitations in representativity or computational cost. Many studies utilize satellite observations for their spatiotemporal coverage, but struggle to adequately account for confounding factors. Previous studies have attempted to account for confounders such as relative humidity (RH) effects on CF or rainfall's impact on convective clouds, and applied methods from causal inference, finding often a negative N_d-LWP sensitivity, implying a warming effect counteracting the Twomey effect's cooling. However, the strength of entrainment enhancement and precipitation suppression is modulated by environmental factors like lower tropospheric stability (LTS) and free tropospheric RH (RH_FT), affecting LWP adjustment responses. Environmental variables' roles (confounders, mediators, etc.) are often not explicitly defined, potentially leading to erroneous conclusions due to Simpson's paradox. Studies focusing on temporal development of LWP adjustments, using high temporal resolution geostationary satellite data, show promise in resolving causality. The temporal resolution of such data allows for the resolution of feedback loops and a focus on temporal development rather than spatial variability. This study uses a causal methodology to investigate LWP adjustments in stratocumulus clouds, applying a physical process-based causal graph to geostationary satellite data, moving beyond simple correlations and accounting for temporal developments and feedback loops.

Numerous studies have investigated the impact of aerosols on cloud properties, particularly focusing on the effect of aerosol perturbations on liquid water path (LWP). Early work established the Twomey effect, showing that increased aerosol concentrations lead to more numerous, smaller cloud droplets, increasing cloud albedo at a constant LWP. However, subsequent research highlighted the complexity of aerosol-cloud interactions (ACI) due to the interplay of various counteracting physical processes, such as precipitation suppression and cloud-top entrainment. The challenge of disentangling these processes has led to conflicting results in the literature regarding the net effect of aerosols on LWP. Some studies have reported a negative aerosol-LWP sensitivity, suggesting that cloud-top entrainment dominates, while others have found a positive or less negative sensitivity under specific meteorological conditions, indicating that precipitation suppression plays a crucial role. Many studies used correlations which may be affected by confounding variables. A number of studies applied advanced methods including causal inference, trying to disentangle direct and indirect effects of aerosols on clouds by using regression techniques, but still faced the challenges of confounding and complex feedback processes. Existing methodologies often lack a systematic approach to handling temporal dependencies and the diverse roles of meteorological variables within these complex causal relationships. The use of high temporal resolution data, and explicit models of causal interactions are still relatively under explored.

This study employs a causal approach to investigate LWP adjustments in stratocumulus clouds, combining physical knowledge encoded in a causal graph with geostationary satellite observations from the Namibian stratocumulus deck (2016-2017). The causal graph (Fig. 1b) explicitly represents the physical interactions between aerosols (represented by N_d), cloud microphysical properties (r_eff, LWP, cloud depth H), entrainment velocity (w_e), and precipitation rate (RR). The graph incorporates lagged effects (0 and 15-min lags) to capture feedback loops and temporal dependencies. Data included two years of SEVIRI satellite retrievals of cloud properties (r_eff, LWP), along with ERA5 reanalysis data for meteorological variables, and GPM IMERG data for precipitation rates. The variables N_d and H were calculated using the adiabatic assumption. Data were co-located, interpolated to 15-minute resolution, spatially averaged, and standardized by removing diurnal and seasonal cycles. Causal effect estimation employed Wright's path approach, a linear method that allows for the calculation of both direct and total causal effects. Direct causal effects (α_xy) represent the immediate impact of one variable on another, while total causal effects (β_xy) include both direct and indirect effects mediated through other variables in the graph. The sign and relative magnitude of these effects, rather than the absolute magnitude, were interpreted because of data standardization. The regime-dependence of causal effects (e.g., effects under different LTS or RH_FT conditions) was investigated using binary masking. Sensitivity analyses examined variations in causal effect estimates based on different versions of the causal graph, removing certain variables or connections to determine the robustness of the results and evaluate the impact of specific graph assumptions on causal effect estimations. Bootstrap methods were used to compute confidence intervals for causal effects.

The analysis confirms the physical plausibility of the proposed causal graph. The signs of direct causal effects largely agree with expectations based on physical processes (Table 1). The positive lag-0 arrows from N_d, r_eff, and H to LWP reflect the definition of LWP. The negative arrow from N_d to r_eff (Twomey effect) is observed, and the positive relationship between H and r_eff (condensation growth with height) is confirmed. The analysis also captures the cloud-top entrainment enhancement feedback loop: a negative effect of r_eff on w_e (larger droplets prevent entrainment) and a negative effect of w_e on r_eff and N_d (evaporation of cloud droplets due to entrainment). The precipitation suppression feedback shows a positive effect of r_eff on RR (larger droplets initiate precipitation), a negative effect of RR on N_d (immediate droplet removal), and a weak, insignificant effect of RR on N_d, possibly due to the interaction between wet scavenging and dynamical effects. Temporal development analysis of total causal effects reveals a time-dependent N_a-LWP relationship (Fig. 3). Initially positive, it quickly becomes negative, remaining so for up to 24 h. Precipitation suppression peaks at 4-6 h post-perturbation, while entrainment enhancement peaks around 12 h, persisting beyond 24 h. Entrainment enhancement is stronger under unstable, dry free tropospheric conditions, while under moist conditions, it is weak or even negative. Regime dependence of the causal effects is also observed. The long-lasting negative N_a-LWP sensitivity is primarily driven by entrainment enhancement, although the precipitation suppression effect is weak, potentially due to issues with precipitation rate retrievals. Sensitivity analysis of alternative causal graphs (Fig. 4) demonstrate the importance of considering covariations in r_eff and H. When neglecting these variables, an overly negative N_a-LWP response is obtained, indicating the underestimation of the cooling effect of ACI.

This study provides a novel physics-informed causal approach for quantifying the causal effect of aerosols on LWP in marine stratocumulus clouds. By disentangling the counteracting effects of entrainment and precipitation using a process-oriented causal graph, the study addresses the challenges of confounding. The findings show a time-dependent and regime-dependent LWP response, with entrainment enhancement being the dominant driver of the long-lasting negative response. This highlights that the effect of precipitation suppression on LWP is substantially weaker than implied in the literature. The sensitivity analysis reveals that neglecting covariations in cloud droplet size (r_eff) and cloud depth (H) can lead to an overly negative N_a-LWP sensitivity. This suggests that previous estimates of the cooling effect of ACI might have underestimated its strength. The study’s strength lies in its comprehensive approach, addressing not only the direct effects of aerosols but also the complex mediated effects through entrainment and precipitation. This framework represents an important advancement in quantifying the influence of aerosols on cloud properties and ultimately, climate.

This study presents a robust causal methodology for investigating LWP adjustments in stratocumulus clouds. Using a physics-informed causal graph and high-resolution geostationary satellite data, we disentangled the counteracting effects of entrainment enhancement and precipitation suppression. We found that failing to account for covariations in cloud droplet sizes and cloud depth leads to an overly negative aerosol-induced LWP response, potentially underestimating the cooling effect of aerosol-cloud interactions. Future research should focus on refining precipitation retrievals, investigating non-linear effects, and exploring the impact of spatial aggregation on causal estimates. Expanding this causal framework to other datasets and cloud regimes will enhance our understanding of ACI and its contribution to climate change.

The study's findings are contingent upon several assumptions: (1) validity of the causal graph, (2) linearity of causal effects, (3) absence of hidden confounders, (4) stationarity of causal effects, and (5) data trustworthiness. While a sensitivity analysis assessed the impact of graph assumptions, the potential for unobserved confounders remains. The linear assumption, while reasonable for short time lags, might not hold for longer timescales. The use of the adiabatic assumption for deriving cloud properties could introduce biases. Spatial averaging, although simplifying the analysis, could mask finer-scale variations in causal relationships. Potential data-related issues include limitations in precipitation retrievals, and uncertainties associated with the calculation of N_d and colocation of datasets.