Ben Harvey

Machine‐learning weather prediction (MLWP) has emerged as a powerful new forecasting tool, with the potential to complement or even replace numerical weather prediction (NWP) in early warning systems. MLWP models are computationally cheap once trained, offering more timely, accessible, and energy‐efficient forecasts than NWP. Yet their ability to capture weather extremes, particularly strong winds associated with extra‐tropical cyclones, remains insufficiently understood. This study provides the first comprehensive evaluation of MLWP performance for dynamically diverse UK windstorms, addressing a critical evidence gap before these models can be responsibly deployed in weather risk management. We evaluate eight state‐of‐the‐art MLWP models against reanalysis, NWP analysis, and an operational NWP ensemble for six contrasting UK cyclones from winter 2023/24. We find substantial inter‐storm and inter‐model variability in forecast skill at 24‐h lead time. Forecast accuracy is found to be linked to the underlying atmospheric dynamics and spatial scales of the wind‐producing features: wind maxima associated with mesoscale structures are systematically underestimated across all MLWP models. Significant variability is also found in the representation of mesoscale storm structure, and the highest‐resolution MLWP model (Aurora‐0.1) does not deliver uniform improvements across the cyclone set. In‐depth analysis of two mature Shapiro–Keyser cyclones reveals a consistent tendency for MLWP models to smooth sharp frontal gradients and misrepresent key mesoscale features, including bent‐back warm fronts and frontal‐fracture regions. These deficiencies are physically consistent with the underestimation of localised extreme winds. Overall, our findings highlight the current limitations of MLWP for underpinning UK wind warnings. While these models are not yet suitable to replace NWP for hazardous windstorm prediction entirely, their demonstrated skill in capturing storm tracks and synoptic‐scale evolution points to promising complementary roles in future national weather forecasting and resilience frameworks.

Extreme extratropical cyclones present major socio-economic risks in the mid-latitudes and are sensitive to anthropogenic climate change. Robust projections of the aggregate properties of extreme cyclones based on climate-model output have emerged in recent years. However, such projections group together cyclones with a range of contrasting dynamical characteristics potentially obscuring climate change effects on particular types of cyclones. Here, we adopt the cyclone track analogue approach to examine the influence of climate change on four contrasting historical cyclones impacting the UK, Ireland and France: the Great Storm in October 1987, Martin in December 1999, Ophelia in October 2017, and Arwen in November 2021. Analogues are identified in the recently-produced CANARI large ensemble simulations for both the present climate (1980–2010) and a high-emission future scenario (SSP3–7.0, 2070–2100). Averaged across the four cases, the overall number of analogue cyclones decreases in future while the intensity of the most extreme cyclones increase, in both precipitation rate and lower-tropospheric wind speed, aligning well with consensus cyclone projections. However, the four cases exhibit contrasting responses, indicating that cyclone-specific changes under anthropogenic warming can diverge from the aggregate signal. For example, there is a marked future increase in the number of cyclones with a path similar to the Great Storm. Such changes are likely driven by regional variations in the conditions for baroclinic growth. Since individual cyclones are typically associated with distinct meteorological hazards, accounting for cyclone-specific responses is critical for assessing regional impacts and developing adaptation strategies.

This paper introduces the CANARI HadGEM3 Large Ensemble (CANARI LE), a 40-member Single-Model Initial-condition Large Ensemble (SMILE) designed to advance the study of climate variability and change. It is based on HadGEM3-GC3.1, the physical climate model used by the UK Met Office and partners for submissions to CMIP6 (Climate Model Intercomparison Project Phase 6). The CANARI LE simulations span 1950–2100 and are driven by CMIP6 historical and SSP3-7.0 forcings. Compared to existing Large Ensembles, the CANARI LE offers several distinctive features: (i) its resolution (1/4° in the ocean, 60 km in the midlatitude atmosphere), (ii) sub-daily output at the surface and on atmospheric pressure levels, and (iii) the provision of boundary conditions for regional modelling. In this study, the design of the CANARI LE is documented and key aspects of historical model performance are evaluated, while applications of the SSP3-7.0 simulations are left for forthcoming studies. Model evaluations are carried out for global mean surface temperature, North Atlantic jets and storm tracks, Northern Hemisphere stratospheric circulation, Atlantic meridional overturning circulation and subpolar gyre, Arctic sea ice, and El Niño-Southern Oscillation teleconnections. In addition, an example is provided deriving analogues of the North Atlantic 2023 summer marine heatwave showing that the CANARI LE captures unprecedented events of this magnitude and spatial scale. The CANARI LE is a community resource that lends itself for the study of physical coupled climate processes and weather systems in a changing climate. The ability to drive regional models will potentially allow for risk assessments of unprecedented detail in climate impacts and adaptation studies.

High-impact weather occurring during periods of low forecast skill can lead to severe socio-economic consequences. This study examines the predictability of the causal chain of events leading to the formation of Medicane Daniel and its associated flooding impacts over Greece on 5 September and Northeast Libya on 10–11 September 2023. Daniel’s formation was preceded by Hurricane Franklin’s extratropical transition over the North Atlantic on 1 September 2023, the development of an Omega-type block over central Europe on 3 September 2023, and the detachment of a PV streamer over the eastern Mediterranean which led to Daniel’s formation as a cut-off low on 5 September 2023. Using a novel block error metric, we investigate the predictability of these events using operational ECMWF and Met Office ensemble forecasts and ERA5 reanalysis. Ensemble sensitivity analysis identifies Franklin’s location as a leading source of downstream uncertainty. Accurate prediction of Franklin’s track was critical for developing an upper-level ridge and European blocking. In ensemble members that best represented the European block, Franklin remained anchored to a downstream ridge, excited a Rossby wave packet over Europe, and initiated a wave breaking event that produced the cut-off leading to the development of Storm Daniel. Two distinct high error recurvature scenarios are identified across the forecasting systems, both of which lead to poor representation of blocking and subsequently to no Mediterranean cyclone development. The strength of Franklin's interaction with the midlatitude flow, and therefore the downstream impact of extratropical transition, is highly sensitive to small track errors at short forecast lead times. Together, these results demonstrate how small upstream errors can contribute to poor predictability of high-impact weather, reinforcing the need for improved representation of the physical processes that take place during tropical–extratropical interactions in operational forecasting to better protect vulnerable regions from impactful weather events.

Phrases like ‘a month’s worth of rain in a day’ make for dramatic news headlines. But how often do such events occur? Are they more likely in certain locations or seasons? And can we expect more in the future? Here we examine these questions for the UK using observations and high‐resolution model simulations. We find that almost all such events occur in summer in regions with the smallest monthly rainfall climatology and that future projections exhibit substantial increases in frequency due to a simultaneous intensification of extreme rainfall events coupled with a marked drying of UK summers.

Hazardous Mediterranean cyclones, including tropical-like medicanes, produce some of the most damaging wind and rainfall events in southern Europe, yet their prediction remains challenging because they are often small in scale and their evolution can be strongly controlled by diabatic processes. While previous research has assessed the role of diabatic processes in individual case studies and/or using resolutions that require parametrization of convection, it remains unclear whether the sensitivity of modelled Mediterranean cyclones to explicit representation of convection is robust across cyclone types. Using a consistent multi-case framework, we identify systematic rather than case-dependent sensitivities. Simulations of seven hazardous MedCys, with a range of characteristics, are performed using four limited-area configurations of the Met Office weather forecast model downscaled from ECMWF IFS analyses. The currently operational physics packages for the global (GAL9) and limited area (RAL3) versions of the model are used with 12-km and convection-permitting, 2.2-km grid spacing, respectively; an experimental package with the scale-aware convection scheme, CoMorph, is also used at both resolutions to cleanly identify resolution dependence. Sensitivity of dynamical variables and hazards to model configuration depends strongly on cyclone type, with medicanes exhibiting substantially greater and more robust dependence than more extratropical systems. The 2.2 km CoMorph configuration consistently over-intensifies medicanes, producing much lower minimum mean-sea-level pressures, and stronger winds and gusts than other simulations, especially for medicanes Ianos and Apollo. In contrast, the 12-km GAL9 configuration generally produces the weakest medicanes and tracks that is displaced eastward or southward; these biases are reduced using the 12-km CoMorph configuration. An in-depth analysis of medicane Ianos, which deepens by at least 20 hPa more in the 2.2-km CoMorph simulation than in the other simulations is presented. This CoMorph simulation yields a Ianos with a stronger potential vorticity tower, ascent and latent heat fluxes, and an eyewall structure resembling a tropical cyclone. These results demonstrate that the forecast skill of hazardous Mediterranean cyclones depends critically on the interaction between model resolution and physical parameterisations, implying that improvements in kilometre-scale forecasting require advances in model physics as well as increased resolution.

High-impact weather occurring during periods of low forecast skill can lead to severe socio-economic consequences. This study examines the predictability of the causal chain of events leading to the formation of Medicane Daniel and its associated flooding impacts over Greece on 5 September and Northeast Libya on 10–11 September 2023. Daniel’s formation was preceded by Hurricane Franklin’s extratropical transition over the North Atlantic on 1 September 2023, the development of an Omega-type block over central Europe on 3 September 2023, and the detachment of a PV streamer over the eastern Mediterranean which led to Daniel’s formation as a cut-off low on 5 September 2023. Using a novel block error metric, we investigate the predictability of these events using operational ECMWF and Met Office ensemble forecasts and ERA5 reanalysis. Ensemble sensitivity analysis identifies Franklin’s location as a leading source of downstream uncertainty. Accurate prediction of Franklin’s track was critical for developing an upper-level ridge and European blocking. In ensemble members that best represented the European block, Franklin remained anchored to a downstream ridge, excited a Rossby wave packet over Europe, and initiated a wave breaking event that produced the cut-off leading to the development of Storm Daniel. Two distinct high error recurvature scenarios are identified across the forecasting systems, both of which lead to poor representation of blocking and subsequently to no Mediterranean cyclone development. The strength of Franklin's interaction with the midlatitude flow, and therefore the downstream impact of extratropical transition, is highly sensitive to small track errors at short forecast lead times. Together, these results demonstrate how small upstream errors can contribute to poor predictability of high-impact weather, reinforcing the need for improved representation of the physical processes that take place during tropical–extratropical interactions in operational forecasting to better protect vulnerable regions from impactful weather events.

This paper introduces the CANARI HadGEM3 Large Ensemble (CANARI LE), a 40-member Single-Model Initial-condition Large Ensemble (SMILE) designed to advance the study of climate variability and change. It is based on HadGEM3-GC3.1, the physical climate model used by the UK Met Office and partners for submissions to CMIP6 (Climate Model Intercomparison Project Phase 6). The CANARI LE simulations span 1950–2100 and are driven by CMIP6 historical and SSP3-7.0 forcings. Compared to existing Large Ensembles, the CANARI LE offers several distinctive features: (i) its resolution (1/4° in the ocean, 60 km in the midlatitude atmosphere), (ii) sub-daily output at the surface and on atmospheric pressure levels, and (iii) the provision of boundary conditions for regional modelling. In this study, the design of the CANARI LE is documented and key aspects of historical model performance are evaluated, while applications of the SSP3-7.0 simulations are left for forthcoming studies. Model evaluations are carried out for global mean surface temperature, North Atlantic jets and storm tracks, Northern Hemisphere stratospheric circulation, Atlantic meridional overturning circulation and subpolar gyre, Arctic sea ice, and El Niño-Southern Oscillation teleconnections. In addition, an example is provided deriving analogues of the North Atlantic 2023 summer marine heatwave showing that the CANARI LE captures unprecedented events of this magnitude and spatial scale. The CANARI LE is a community resource that lends itself for the study of physical coupled climate processes and weather systems in a changing climate. The ability to drive regional models will potentially allow for risk assessments of unprecedented detail in climate impacts and adaptation studies.

Hazardous Mediterranean cyclones, including tropical-like medicanes, produce some of the most damaging wind and rainfall events in southern Europe, yet their prediction remains challenging because they are often small in scale and their evolution can be strongly controlled by diabatic processes. While previous research has assessed the role of diabatic processes in individual case studies and/or using resolutions that require parametrization of convection, it remains unclear whether the sensitivity of modelled Mediterranean cyclones to explicit representation of convection is robust across cyclone types. Using a consistent multi-case framework, we identify systematic rather than case-dependent sensitivities. Simulations of seven hazardous MedCys, with a range of characteristics, are performed using four limited-area configurations of the Met Office weather forecast model downscaled from ECMWF IFS analyses. The currently operational physics packages for the global (GAL9) and limited area (RAL3) versions of the model are used with 12-km and convection-permitting, 2.2-km grid spacing, respectively; an experimental package with the scale-aware convection scheme, CoMorph, is also used at both resolutions to cleanly identify resolution dependence. Sensitivity of dynamical variables and hazards to model configuration depends strongly on cyclone type, with medicanes exhibiting substantially greater and more robust dependence than more extratropical systems. The 2.2 km CoMorph configuration consistently over-intensifies medicanes, producing much lower minimum mean-sea-level pressures, and stronger winds and gusts than other simulations, especially for medicanes Ianos and Apollo. In contrast, the 12-km GAL9 configuration generally produces the weakest medicanes and tracks that is displaced eastward or southward; these biases are reduced using the 12-km CoMorph configuration. An in-depth analysis of medicane Ianos, which deepens by at least 20 hPa more in the 2.2-km CoMorph simulation than in the other simulations is presented. This CoMorph simulation yields a Ianos with a stronger potential vorticity tower, ascent and latent heat fluxes, and an eyewall structure resembling a tropical cyclone. These results demonstrate that the forecast skill of hazardous Mediterranean cyclones depends critically on the interaction between model resolution and physical parameterisations, implying that improvements in kilometre-scale forecasting require advances in model physics as well as increased resolution.

High-impact weather occurring during periods of low forecast skill can lead to severe socio-economic consequences. This study examines the predictability of the causal chain of events leading to the formation of Medicane Daniel and its associated flooding impacts over Greece on 5 September and Northeast Libya on 10–11 September 2023. Daniel’s formation was preceded by Hurricane Franklin’s extratropical transition over the North Atlantic on 1 September 2023, the development of an Omega-type block over central Europe on 3 September 2023, and the detachment of a PV streamer over the eastern Mediterranean which led to Daniel’s formation as a cut-off low on 5 September 2023. Using a novel block error metric, we investigate the predictability of these events using operational ECMWF and Met Office ensemble forecasts and ERA5 reanalysis. Ensemble sensitivity analysis identifies Franklin’s location as a leading source of downstream uncertainty. Accurate prediction of Franklin’s track was critical for developing an upper-level ridge and European blocking. In ensemble members that best represented the European block, Franklin remained anchored to a downstream ridge, excited a Rossby wave packet over Europe, and initiated a wave breaking event that produced the cut-off leading to the development of Storm Daniel. Two distinct high error recurvature scenarios are identified across the forecasting systems, both of which lead to poor representation of blocking and subsequently to no Mediterranean cyclone development. The strength of Franklin's interaction with the midlatitude flow, and therefore the downstream impact of extratropical transition, is highly sensitive to small track errors at short forecast lead times. Together, these results demonstrate how small upstream errors can contribute to poor predictability of high-impact weather, reinforcing the need for improved representation of the physical processes that take place during tropical–extratropical interactions in operational forecasting to better protect vulnerable regions from impactful weather events.