(566l) Adjoint Model for the Efficient Calculation of Atmospheric Nucleation Sensitivities

Adjoints can be used for computationally efficient calculation of the model output gradient in both linear and nonlinear systems. While model gradients may be calculated in a forward run using finite differences, the computational time in this case increases with the dimension of the input space. During a reverse run, however, the model state is stored after certain steps in the instruction sequence and then passed back through a piecewise-differentiated sequence so that additional inputs have no additional computational cost. These reverse-mode adjoints have a wide range of applications including sensitivity study, stability analysis, and optimization. Here, we present one environmentally-relevant use with the adjoint model of an atmospheric ice nucleation code by Barahona and Nenes (ABN09). ABN09 calculates the change in nucleated ice crystal number with size and composition of seed material as well as atmospheric temperature and vertical velocity in a single run. Running this gradient model with inputs from a global climate model, we can see in which geographic regions ice crystals nucleate heterogeneously versus homogeneously and where these regimes are most sensitive to addition of ice nuclei. An attribution analysis is performed to see whether more variability in ice crystal number comes from inherent model bias or from input fluctuations. We test three heterogeneous nucleation spectra to describe the portion of ice-nucleating seed material within ABN09 to understand the effect on sensitivity of different model formulations.