James Murphy
Met Office Hadley Centre, Exeter, UK
During the 1960s, 70s and 80s the Synoptic Climatology Branch of the Met Office issued regular monthly forecasts, initially as a public service and from 1980 to commercial and corporate customers (Folland and Woodcock, 1986). During the 1980s, forecasts for days 6-15 and 16-30 ahead were based mainly on a multivariate statistical technique (Maryon and Storey, 1985). This predicted probabilities for six regional sea level pressure patterns (Fig. 1) using antecedent values of hemispheric circulation eigenvectors and worldwide SSTs. Derived using cluster analysis, the predictand patterns were assumed to represent spells of weather associated with persistent circulation regimes. Forecasts of surface temperature and rainfall for UK regions were derived from the most likely cluster, by combining further statistical processing with subjective assessments of confidence. The predictions were presented as probabilities for outcomes lying within quintiles (for temperature) and terciles (for rainfall) of the relevant climatological distributions.
In parallel, Tim, myself and colleagues were using early Met Office atmospheric general circulation models (AGCMs) to study dynamical extended-range forecasts, encouraged by early demonstrations of predictability at the monthly time scale (Shukla, 1981; Miyakoda et al., 1983). Initially, a hemispheric 5-level AGCM was used to demonstrate potential skill arising from SST anomalies (e.g. Palmer and Sun, 1985; Mansfield, 1986) and the use of ensembles (Murphy, 1988). Following the introduction of a more sophisticated 11-level global AGCM (Slingo, 1985), we contributed ensemble integrations to the regular monthly forecasts (Murphy and, Palmer, 1986). To our knowledge, this constituted the first use of monthly dynamical predictions in a quasi-operational context.
Our first real-time forecast was started from 15 September 1985, consisting of 7 members created using time-lagged operational analyses and persisted SST anomalies. The probabilistic nature of the statistical forecasts created a natural environment for dynamical ensembles expressed in a similar format. For example, our first forecast developed two distinct clusters in the Pacific/North American (PNA) sector, with four members resembling the PNA pattern (Wallace and Gutzler, 1981) while three predicted a broad cyclonic anomaly (Fig. 2).
Palmer et al. (1986) later improved the 11-level AGCM by introducing a parameterisation of gravity wave drag that removed a westerly bias in the mid-latitude northern hemisphere flow. Using this version, Murphy (1990) demonstrated probabilistic skill (relative to climatology) in a set of dynamical monthly predictions. Folland et al. (1986) found improved skill in operational forecasts issued during the 1980s, compared with their predecessors.
By the late 1980s, Tim and I had left the Synoptic Climatology Branch for ECMWF and the Met Office Hadley Centre respectively. The focus of my work switched to longer time scales, including initialised seasonal to decadal predictions (Smith et al., 2007) and multi-decadal climate change projections (Murphy et al., 2004). Since the 1990s, specifications of ensemble prediction systems for seasonal and longer time scales have grown in sophistication, accounting for modelling as well as initial state uncertainties. This has involved assembly of multi-model ensembles (e.g. Smith et al., 2013; Eyring et al., 2016) and design of single-model ensembles capturing parametric and stochastic uncertainties in sub-grid scale processes (e.g. Doblas-Reyes et al., 2009). As an example, the latest projections of 21st century UK climate use a suite of ensemble methods (Murphy et al., 2018; Kendon et al., 2021). These include global simulations that combine multi-model and perturbed parameter ensembles (MMEs and PPEs) to represent model uncertainties and limited-area ensembles derived from Met Office models. We also provide probabilistic projections based on MMEs and PPEs of earth system models (thus accounting for both physical and carbon cycle feedbacks) combined with regional simulations.