Sensitivity analysis and parameter tuning scheme for global sea-ice modeling

Automatic differentiation (AD) is used to perform a multiple parameter sensitivity analysis for the Los Alamos sea-ice model CICE. Numerical experiments are run by six-hourly, 1987 forcing data with a two-hour time step, and the AD-based sensitivity scheme is validated by comparison with derivatives calculated using the conventional finite-difference approach. Twenty-two thermodynamic and dynamic parameters are selected for simultaneous analysis. Of these, the simulated average sea-ice thickness is most sensitive to ice density; albedos and emissivity predominate in summer, while ice thickness is highly sensitive to the snow density in winter. Ice conductivity, the ice-ocean drag parameter, maximum ice salinity and ridging parameters significantly affect the simulation year-round. Gradient information computed by the AD-based sea-ice code is then used in an experiment designed to assess the efficacy of this technique for tuning the parameters against observational data. Preliminary results, obtained with a bound-constrained minimization method and with simulated observational data, show that satisfactory convergence is obtained.