(468c) Uncertainty Quantification and Parameter Constraint in Complex Environmental Models: A Case Study of Smoke Influencing Radiative Balance in Africa

Development of environmental models is key to understanding future climate and informing related policy. Constructing methods for improving the accuracy of these models can be just as important as their initial development. Validating environmental models with observations and measurements is necessary to reduce uncertainties and identify shortcomings in a model's capabilities.

Here, we use a case study of seasonal wildfires in Southern Africa to develop methods for uncertainty quantification and constraint of input parameters in the UK Earth System Model, developed by the Met Office in the UK. This case study poses a unique problem because the Southern African wildfires (1) release the largest portion of the planet’s biomass burning aerosols into the atmosphere and (2) the prevailing winds carry the smoke over the south Atlantic ocean where they interact with clouds. These wildfire emissions and their interactions with clouds have significant implications for the planet's radiation budget. Reducing large uncertainties associated with inputs to the model is key to developing a better understanding of the true magnitude of the importance of these seasonal fires in Southern Africa.

We constrain the uncertainty in the radiative effects of smoke from seasonal biomass burning in Southern Africa which we simulate in a climate model by considering the uncertainty in the input parameters to that model. We hypothesize that we can parameterize a large fraction of the uncertainty in smoke radiative effects by varying 12 uncertain parameters in the model. Some parameters scale emissions, both of smoke and of natural aerosols and precursors, and others scale parameters in uncertain processes such as the vertical velocities used to activate cloud droplets. Since satellites are not capable of measuring radiative effects, we start by comparing simulated aerosol optical depth (AOD) with satellite measured AOD to rule out combinations of parameters that produce implausible results. Prior work (e.g. Johnson et al, Atmos Chem Phys 2020) has shown AOD is capable of constraining model predictions involving the planet's radiative balance. We calculate the fire radiative effects by comparing simulations with and without smoke emissions. After AOD, we then apply the same methodology to simulated and satellite measured cloud droplet number concentration (CDNC) and cloud liquid water path (CLWP) to develop a stronger constraint on the radiative effect of these fires. Here, the variables for which values are compared between simulation and satellite will be called comparison variables.

To constrain our uncertain parameters, we built a perturbed parameter ensemble (PPE) consisting of 121 runs of the atmosphere-only UK Earth System Model spanning our space of 12 parameters and this was used as training data to build surrogate models. The values for the 121 parameter variants of the PPE were set using a Latin hypercube to ensure an adequate preliminary coverage of the parameter space. We simulate only the 2017 biomass burning season, and write data out from the simulations at a three-hourly time resolution, which allows us to understand the transport of individual smoke plumes. The surrogate models are trained using Gaussian Process Regression and allow us to approximately evaluate the model over the 12-D parameter space at a high coverage by utilizing 100,000 surrogate model parameter variants.

To rule out implausible parameter combinations that yield simulated comparison variables that are inconsistent with the observations, we use statistical methods developed in Carzon et al (Environmental Data Science, 2023). These allow us to quantify uncertainty in the parameters at a high level of statistical confidence, improving on previous related work using the method of 'history matching' (whose theoretical guarantees are not well understood). We build on the statistical methods of Carzon et al. which enable principled constraints on uncertain parameters with guarantees of frequentist coverage. The methods developed here can be applied to uncertainty analysis and parameter constraint across multiple variables in other complex chemical and atmospheric models. We will present the methods, the constraints on parameters we have achieved to date, and the implications for the smoke radiative effects.