Paleoclimate Data Assimilation with and without a Forecast Model

Abstract:

Data assimilation (DA) has emerged as a promising technique for combining information from paleoclimate proxy data and climate models. Research on this topic has progressed to the point where an operational grid reconstruction project is underway using an ensemble approach (the Last Millennium Reanalysis; LMR). For problems on weather timescales, ensemble DA typically utilizes a “cycling” process, where an ensemble of forecasts provides the prior estimate to be combined with observational information. For the paleoclimate problem, cycling faces dual challenges of very large computational cost, and weak predictive skill on annual-decadal timescales. As a result, recent work in this area has used a “no cycling” approach where the prior ensemble is instead drawn randomly from a long climate simulation. Here we investigate the viability of adding cycling by means of a low-cost alternative for climate forecasting known as a linear inverse model (LIM).

LIMs have been shown to have forecast skill comparable to coupled global climate models on annual time scales and due to their simplicity have low computational expense. In this study, we assess the reconstruction skill of ensemble DA with cycling relative to a control no-cycling reconstruction. Each reconstruction uses a random draw from a pre-industrial climate simulation as its initial (cycling) or annual (no-cycling) prior estimate, and assimilates observations from the PAGES 2k proxy dataset. Reconstructions for the period from 1000-2000 CE are performed, and both the correlation and coefficient of efficiency (CE) values for global averages and spatial fields are calculated against observational datasets during the instrumental record. Preliminary results for global mean temperature show that while correlations are high with the cycling approach (>0.8), they are slightly lower than results for the no-cycling reconstruction.