Seasonal arctic sea ice predictability and prediction
- Autores: Rubén Cruz García
- Directores de la Tesis: Virginie Guemas (dir. tes.), Pablo Ortega Montilla (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2020
- Idioma: español
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Arctic sea ice plays a central role in the Earth’s climate. Changes in the sea ice on seasonal-to-interannual timescales impact ecosystems, populations and a growing number of stakeholders. A prerequisite for achieving better sea ice predictions is a better understanding of the underlying mechanisms of sea ice predictability.
In the first part, we investigate the seasonal-to-interannual Arctic sea ice predictability in perfect-model experiments performed with six different climate models. Similar pan-Arctic winter sea ice extent (SIE) reemergence is found for HadGEM1.2, GFDL-CM3 and E6F, while a sea ice volume (SIV) persistence from 1 to 3 years is confirmed for all models. Similarities in winter SIE predictability remergence in the GIN seas and Baffin Bay are found even though models have distinct sea ice states. A summer SIV skill reemergence is also found in the Barents, Kara and Chukchi seas. A regional analysis in EC-Earth2.3 suggests that Arctic basins can be classified according to three distinct regimes. The central Arctic drives most of the pan-Arctic SIV persistence. In peripheral seas, predictability for the SIE in winter is associated with ocean thermal anomalies persistence. The Labrador Sea exhibits the longest predictability (up to 1.5 years), the reemergence of predictability in winter being driven by the advection of heat content anomalies along the subpolar gyre.
In real predictions, forecast errors appear due to inconsistencies between the initial states of the different model components and to the development of the inherent model biases. We identify and quantify the contribution of initial condition (IC) inconsistencies and systematic model errors to the forecast model errors in two sets of seasonal forecasts (May and November initialized) produced with EC-Earth3.2 during the first forecast month. After 24 (19) days, the inherent model biases become the largest contributor to the forecast error for the May (November) initialized forecasts, while the initial inconsistency dominates in the previous days. This initial inconsistency is mostly associated to a mismatch between the sea ice and ocean ICs, with a marginal role associated to differences with the atmosphere. The development of both types of errors is sensitive to the month of initialization: the initial shock is more pronounced in November than in May because the initial ocean is warmer and less consistent with the initial sea ice cover, in both cases the shock leading to sea ice melting.
The last part compares three seasonal forecast systems based on EC-Earth and initialized through three different strategies: (1) using both the sea ice and ocean ICs from a native reconstruction that assimilates ORAS4 temperature and salinity with a weak surface restoring coefficient, (2) taking the sea ice ICs from the same reconstruction as in 1 and the oceanic ICs from ORAS4 and (3) the same as in 1 but using a stronger restoring coefficient. The objective is to assess the impact of these methods on the sea ice bias and skill. Strategy 2 induces an initial shock because of a too warm polar surface ocean in ORAS4 for the reconstructed sea ice ICs. For strategy 3, a strong ocean nudging towards ORAS4 produces a too warm ocean and a sea ice deficit. For the prediction initialized using method 1, no shocks are observed. The sea ice errors of strategies 2 and 3 converge in less than a month, demonstrating the role of the ocean ICs in driving the sea ice biases. Although no significant differences appear between the sea ice prediction skill of the three forecast systems, all three show significantly higher skill than the historical experiment for the pan-Arctic SIE during the first two months. The Labrador Sea is the only basin with a significant added value of initialization during the 7 months of prediction, whereas only about 3 months of added-value are detected in the peripheral basins.
