Thursday, 1 February 2024
Hall E (The Baltimore Convention Center)
Depositional ice growth, the process by which ice crystals grow from vapor, is an essential physical process in the formation of cirrus clouds. While the details of this process impact many important features, such as the radiative properties of cirrus clouds, we still lack understanding of the physics of depositional ice growth at the low temperatures (<235K) at which cirrus clouds form. Experiments based on the evolution of single ice-crystals reveal the surface attachment kinetics, parameterized as a deposition coefficient to be a complicated function depending on ambient conditions; however extrapolation to ice formation under atmospheric conditions is not straightforward. Previous work [Lamb et al., 2022, Atmos. Chem. Phys.] thus suggests the use of measurements from expansion experiments performed inside the Aerosol Interaction and Dynamics in the Atmosphere (AIDA) cloud chamber under realistic cirrus conditions to constrain proposed models of depositional ice growth. For a quantitative evaluation of these constraints, Bayesian analysis is a natural choice. Our goal is to infer microphysical model parameters while addressing uncertainty in observations, model parameters, and model structure in a comprehensive way.
To avoid a prohibitive number of model evaluations, we opt for an efficient adaptive Metropolis algorithm to approximate the posterior parameter distributions and use an ensemble of chains to check for mixing and convergence.
We embed this parameter inference in a larger Bayesian workflow, which allows for iterative model building, model checking, and model comparison for a variety of depositional ice growth models. In particular, we address the choice of priors by using model checks unconstrained by data (prior predictive checks) and prior sensitivity analysis, as well as simulation-based calibration - a method to check how well the model parameters can, in principle, be determined from the data at hand.
Model comparison based on predictive performance (using pointwise predictive densities) shows that a model with constant deposition coefficient and one with temperature-dependent critical supersaturations for step nucleation describe the data equally well.
To avoid a prohibitive number of model evaluations, we opt for an efficient adaptive Metropolis algorithm to approximate the posterior parameter distributions and use an ensemble of chains to check for mixing and convergence.
We embed this parameter inference in a larger Bayesian workflow, which allows for iterative model building, model checking, and model comparison for a variety of depositional ice growth models. In particular, we address the choice of priors by using model checks unconstrained by data (prior predictive checks) and prior sensitivity analysis, as well as simulation-based calibration - a method to check how well the model parameters can, in principle, be determined from the data at hand.
Model comparison based on predictive performance (using pointwise predictive densities) shows that a model with constant deposition coefficient and one with temperature-dependent critical supersaturations for step nucleation describe the data equally well.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner