Massimo Bonavita
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Massimo Bonavita has a total of 94 co-authors. Here are the top co-authors by number of shared publications:
- Alan Jon Geer
ECMWF, Reading, RG2 9AX, UK
6 shared publications - Patrick Laloyaux
ECMWF
6 shared publications Lars Isaksen
European Centre for Medium-Range Weather Forecasts, Reading, United Kingdom
5 shared publications- Alban Farchi
ECMWF, Shinfield Park, Reading, UK
4 shared publications Elias Holm
4 shared publicationsMarcin Chrust
ECMWF Reading UK
4 shared publications
Latest Publications
- Open Access
- 21 October 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- p. e70331
Data‐driven emulation of flow‐dependent background‐error variance in variational data assimilation
Accurate representation of flow‐dependent background‐error covariance is essential for modern ensemble–variational data assimilation systems, but estimating ensemble covariance from sufficiently large ensembles remains computationally expensive. In this work, we propose a machine‐learning approach to emulate ensemble‐derived background‐error variance fields from severely undersampled ensembles. Machine‐learning models are trained to map variance estimates derived from five‐member ensembles to those obtained from operational 50‐member European Centre for Medium‐Range Weather Forecasts (ECMWF) ensemble of data assimilations (EDA). The trained models are deployed within a hybrid ensemble four‐dimensional variational data assimilation system and evaluated over an independent three‐month period in an operational numerical weather prediction environment. In these experiments, the full ensemble is retained for the non‐variance components of the background‐error covariance, so the study tests machine‐learning emulation of the variance component rather than finding a complete low‐cost replacement for the operational ensemble system. Results show that the machine‐learning‐emulated variance fields reproduce much of the structure and analysis impact of the full 50‐member variance estimate, yielding comparable observation‐space diagnostics and forecast‐error characteristics for this component while using only a small fraction of the ensemble size. To our knowledge, this represents the first demonstration in an operational environment of machine‐learning emulation of ensemble‐derived background‐error variance statistics within a fully cycled numerical weather prediction data assimilation framework.
- 21 October 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- p. e70331
Open AccessData‐driven emulation of flow‐dependent background‐error variance in variational data assimilation
Accurate representation of flow‐dependent background‐error covariance is essential for modern ensemble–variational data assimilation systems, but estimating ensemble covariance from sufficiently large ensembles remains computationally expensive. In this work, we propose a machine‐learning approach to emulate ensemble‐derived background‐error variance fields from severely undersampled ensembles. Machine‐learning models are trained to map variance estimates derived from five‐member ensembles to those obtained from operational 50‐member European Centre for Medium‐Range Weather Forecasts (ECMWF) ensemble of data assimilations (EDA). The trained models are deployed within a hybrid ensemble four‐dimensional variational data assimilation system and evaluated over an independent three‐month period in an operational numerical weather prediction environment. In these experiments, the full ensemble is retained for the non‐variance components of the background‐error covariance, so the study tests machine‐learning emulation of the variance component rather than finding a complete low‐cost replacement for the operational ensemble system. Results show that the machine‐learning‐emulated variance fields reproduce much of the structure and analysis impact of the full 50‐member variance estimate, yielding comparable observation‐space diagnostics and forecast‐error characteristics for this component while using only a small fraction of the ensemble size. To our knowledge, this represents the first demonstration in an operational environment of machine‐learning emulation of ensemble‐derived background‐error variance statistics within a fully cycled numerical weather prediction data assimilation framework.
- Open Access
- 6 September 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- p. e70309
Error in ERA5 2‐m temperature identified using GraphCast
Reanalyses such as the European Centre for Medium‐Range Weather Forecasts (ECMWF) Reanalysis Version 5 (ERA5) have long been foundational for weather and climate science. They have also found a new use case, as training and verification data for machine‐learnt weather prediction (MLWP) models. Here we compare short‐lead‐time (6 h) forecasts from the MLWP model GraphCast against ERA5. In doing so, we identify a recurrent, spatially coherent error in 2‐m temperature centred on the Ethiopian Highlands, which occurs predominantly at 0600 UTC. We find that the same error feature is present in other MLWP models trained on ERA5. We show that these error events arise from errors in ERA5 that are also present in the ECMWF operational analysis. They arise from the two‐dimensional (2D) optimal interpolation procedure, when surface reports that are temporally displaced compared with the background forecast are assimilated. This produces spuriously warm analysis increments over Ethiopia on approximately 7% of dates at 0600 UTC across the reanalysis record. The spread from the ensemble of data assimilation partially flags these cases but is underdispersive. We assess the impact on a MLWP system trained on ERA5. While the MLWP model can largely ignore these unphysical error events, a small systematic degradation in forecast skill over the region is observed. We discuss the implications for using reanalysis as truth in machine‐learning training and verification, and recommend simple changes to reduce such artefacts in future analyses.
- 6 September 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- p. e70309
Open AccessError in ERA5 2‐m temperature identified using GraphCast
Reanalyses such as the European Centre for Medium‐Range Weather Forecasts (ECMWF) Reanalysis Version 5 (ERA5) have long been foundational for weather and climate science. They have also found a new use case, as training and verification data for machine‐learnt weather prediction (MLWP) models. Here we compare short‐lead‐time (6 h) forecasts from the MLWP model GraphCast against ERA5. In doing so, we identify a recurrent, spatially coherent error in 2‐m temperature centred on the Ethiopian Highlands, which occurs predominantly at 0600 UTC. We find that the same error feature is present in other MLWP models trained on ERA5. We show that these error events arise from errors in ERA5 that are also present in the ECMWF operational analysis. They arise from the two‐dimensional (2D) optimal interpolation procedure, when surface reports that are temporally displaced compared with the background forecast are assimilated. This produces spuriously warm analysis increments over Ethiopia on approximately 7% of dates at 0600 UTC across the reanalysis record. The spread from the ensemble of data assimilation partially flags these cases but is underdispersive. We assess the impact on a MLWP system trained on ERA5. While the MLWP model can largely ignore these unphysical error events, a small systematic degradation in forecast skill over the region is observed. We discuss the implications for using reanalysis as truth in machine‐learning training and verification, and recommend simple changes to reduce such artefacts in future analyses.
- Open Access
- 4 June 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 152 (780)
Spatial observation‐error correlations for AMSU‐A in all‐sky assimilation: An ECMWF and UK Met Office intercomparison
Observations from microwave temperature sounders such as the Advanced Microwave Sounding Unit‐A (AMSU‐A) provide some of the largest contributions to forecast skill in numerical weather prediction. Currently, AMSU‐A radiances are assimilated under all‐sky conditions at operational centres without accounting explicitly for spatial observation‐error correlations. To mitigate the impact of unrepresented correlated errors, strategies like spatial thinning (reducing observation density) and inflated observation‐error variances are commonly used. Here we present new estimates of spatial observation‐error correlations for the AMSU‐A all‐sky systems from the European Centre for Medium‐Range Weather Forecasts (ECMWF) and the Met Office, using diagnostics from background and analysis departures. Results are presented for three different cases: when all data are considered and when data are separated by surface type and cloud cover. High spatial observation‐error correlations are seen particularly for tropospheric channels (4–8) over land, with correlation length‐scales ranging from 75 to 125 km. We hypothesise that these correlations originate primarily from inadequacies in the modelling of surface emissivity, surface skin temperature, clouds, or precipitation. Our findings suggest that an increase in forecast skill could be achieved by following a pragmatic strategy of increasing assimilated observation density for stratospheric channels (9–14) and all channels 4–14 over the ocean, due to the negligible error correlations in these situations. In contrast, for tropospheric channels over land, exploiting the available data fully through reduced thinning requires accounting for spatial observation‐error correlations.
- 4 June 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 152 (780)
Open AccessSpatial observation‐error correlations for AMSU‐A in all‐sky assimilation: An ECMWF and UK Met Office intercomparison
Observations from microwave temperature sounders such as the Advanced Microwave Sounding Unit‐A (AMSU‐A) provide some of the largest contributions to forecast skill in numerical weather prediction. Currently, AMSU‐A radiances are assimilated under all‐sky conditions at operational centres without accounting explicitly for spatial observation‐error correlations. To mitigate the impact of unrepresented correlated errors, strategies like spatial thinning (reducing observation density) and inflated observation‐error variances are commonly used. Here we present new estimates of spatial observation‐error correlations for the AMSU‐A all‐sky systems from the European Centre for Medium‐Range Weather Forecasts (ECMWF) and the Met Office, using diagnostics from background and analysis departures. Results are presented for three different cases: when all data are considered and when data are separated by surface type and cloud cover. High spatial observation‐error correlations are seen particularly for tropospheric channels (4–8) over land, with correlation length‐scales ranging from 75 to 125 km. We hypothesise that these correlations originate primarily from inadequacies in the modelling of surface emissivity, surface skin temperature, clouds, or precipitation. Our findings suggest that an increase in forecast skill could be achieved by following a pragmatic strategy of increasing assimilated observation density for stratospheric channels (9–14) and all channels 4–14 over the ocean, due to the negligible error correlations in these situations. In contrast, for tropospheric channels over land, exploiting the available data fully through reduced thinning requires accounting for spatial observation‐error correlations.
- Open Access
- 14 March 2026
- Published by Copernicus GmbH
Reaching far and wide: accounting for spatially correlated observational errors in an ensemble of 4D-Vars system
Data assimilation (DA) seeks to provide the most likely estimate of the true state of the atmosphere or ocean by combining a background estimate from a numerical model with observations, each weighted by their respective error covariances. For computational reasons, diagonal covariances together with variance inflation have traditionally been favoured for the observational error covariance. However, recent studies have shown that variance inflation can degrade DA performance when the correlation length scales of observational errors are comparable to, or exceed, those of the background errors—a situation frequently encountered when assimilating wind vectors derived from Atmospheric Motion Vectors. These observations are assimilated, alongside many others, in the ensemble DA system at the European Centre for Medium-Range Weather Forecasts (ECMWF). In this system, each ensemble member undergoes an independent 4D-Var minimisation after perturbing its observations and model parameters to represent observational and model uncertainties. This work presents results from efforts to explicitly account for spatial correlations in observational errors within the ensemble DA framework. In particular, it demonstrates the positive impact of introducing spatially correlated perturbations to assimilated observations on ensemble spread, offering a pathway to improved representation of uncertainty in operational forecasting.
- 14 March 2026
- Published by Copernicus GmbH
Open AccessReaching far and wide: accounting for spatially correlated observational errors in an ensemble of 4D-Vars system
Data assimilation (DA) seeks to provide the most likely estimate of the true state of the atmosphere or ocean by combining a background estimate from a numerical model with observations, each weighted by their respective error covariances. For computational reasons, diagonal covariances together with variance inflation have traditionally been favoured for the observational error covariance. However, recent studies have shown that variance inflation can degrade DA performance when the correlation length scales of observational errors are comparable to, or exceed, those of the background errors—a situation frequently encountered when assimilating wind vectors derived from Atmospheric Motion Vectors. These observations are assimilated, alongside many others, in the ensemble DA system at the European Centre for Medium-Range Weather Forecasts (ECMWF). In this system, each ensemble member undergoes an independent 4D-Var minimisation after perturbing its observations and model parameters to represent observational and model uncertainties. This work presents results from efforts to explicitly account for spatial correlations in observational errors within the ensemble DA framework. In particular, it demonstrates the positive impact of introducing spatially correlated perturbations to assimilated observations on ensemble spread, offering a pathway to improved representation of uncertainty in operational forecasting.
- Open Access
- 8 February 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 152 (776)
Forecast verification using information and noise
Numerical weather prediction (NWP) centres evaluate forecast quality using statistical assessments of error and skill, commonly referred to as scores. Traditional forecast verification relies on metrics such as root‐mean‐squared error, anomaly correlation coefficient, Brier score, and so forth. However, these measures can be sensitive to changes in bias and activity, as well as to intrinsic changes in forecast skill, complicating direct comparisons between deterministic models, ensemble averaging approaches, and machine‐learning‐based forecasts. A clear separation between intrinsic forecast skill and post‐processing enhancements, such as calibration, is essential for accurately assessing the predictive capability of a forecast system. In this work we take forecast reliability and resolution as the fundamental attributes characterising forecast performance, with resolution representing the intrinsic predictive capability of a system—its ability to distinguish among observed events. Recent work introduced information and noise as new metrics designed to provide an unambiguous assessment of statistical resolution. This study aims to introduce these novel scores in an accessible manner, relating them to traditional verification metrics, and to tackle some of the limitations of the original formulation. Additionally, we demonstrate their practical implementation for routine forecast verification in an operational NWP environment and provide examples of their use in the standard NWP research workflow. Examples of application of these new verification metrics to ensemble forecasting and to machine‐learning forecast models are also provided.
- 8 February 2026
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 152 (776)
Open AccessForecast verification using information and noise
Numerical weather prediction (NWP) centres evaluate forecast quality using statistical assessments of error and skill, commonly referred to as scores. Traditional forecast verification relies on metrics such as root‐mean‐squared error, anomaly correlation coefficient, Brier score, and so forth. However, these measures can be sensitive to changes in bias and activity, as well as to intrinsic changes in forecast skill, complicating direct comparisons between deterministic models, ensemble averaging approaches, and machine‐learning‐based forecasts. A clear separation between intrinsic forecast skill and post‐processing enhancements, such as calibration, is essential for accurately assessing the predictive capability of a forecast system. In this work we take forecast reliability and resolution as the fundamental attributes characterising forecast performance, with resolution representing the intrinsic predictive capability of a system—its ability to distinguish among observed events. Recent work introduced information and noise as new metrics designed to provide an unambiguous assessment of statistical resolution. This study aims to introduce these novel scores in an accessible manner, relating them to traditional verification metrics, and to tackle some of the limitations of the original formulation. Additionally, we demonstrate their practical implementation for routine forecast verification in an operational NWP environment and provide examples of their use in the standard NWP research workflow. Examples of application of these new verification metrics to ensemble forecasting and to machine‐learning forecast models are also provided.
- Open Access
- 30 June 2025
- Published by Copernicus GmbH
Towards higher resolution data assimilation in ECMWF IFS
Enhancing the spatial resolution of global analyses and forecasts enables weather prediction systems to more accurately capture rapidly evolving mesoscale and convective-scale atmospheric phenomena. The Integrated Forecast System (IFS) of ECMWF consists of an Earth-system model coupled with an incremental 4D-Var data assimilation (DA) system, which has recently been experimentally upgraded to higher spatial resolution.This upgrade involved higher resolution (4.4 km) 4D-Var trajectories and higher resolution (20 km) 4D-Var minimizations using the tangent linear model (TLM) and its adjoint (ADM). This configuration requires less observation thinning and resulted in greater use of high-resolution observations, such as those from multispectral imagers aboard Himawari and Meteosat satellites. Shorter time steps in TLM and ADM further enabled the use of shorter observation time slots (400 seconds), allowing for more accurate comparison between observations and their model equivalents.We demonstrate that this setup leads to significant improvements in the accuracy of initial conditions and enhances tropospheric forecast skill, extending medium-range predictability by 6 to 12 hours. We argue that the improved forecast skill can be attributed to a better representation of large-scale features in the stratosphere. Higher resolution DA and forecasts also lead to improved forecasting of extreme precipitation events, as well as the track and intensity of tropical cyclones. Furthermore, the improved initial conditions significantly increase the predictability machine-learning-based forecasts.The benefits of the higher resolution DA system are demonstrated using the case of Tropical Cyclone Otis, which made landfall as a Category 5 cyclone, but was predicted by most global and regional NWP models to peak only at tropical storm intensity – a major forecasting bust. By employing higher resolution 4D-Var DA, we were able to replicate the observed rapid intensification of TC Otis.
- 30 June 2025
- Published by Copernicus GmbH
Open AccessTowards higher resolution data assimilation in ECMWF IFS
Enhancing the spatial resolution of global analyses and forecasts enables weather prediction systems to more accurately capture rapidly evolving mesoscale and convective-scale atmospheric phenomena. The Integrated Forecast System (IFS) of ECMWF consists of an Earth-system model coupled with an incremental 4D-Var data assimilation (DA) system, which has recently been experimentally upgraded to higher spatial resolution.This upgrade involved higher resolution (4.4 km) 4D-Var trajectories and higher resolution (20 km) 4D-Var minimizations using the tangent linear model (TLM) and its adjoint (ADM). This configuration requires less observation thinning and resulted in greater use of high-resolution observations, such as those from multispectral imagers aboard Himawari and Meteosat satellites. Shorter time steps in TLM and ADM further enabled the use of shorter observation time slots (400 seconds), allowing for more accurate comparison between observations and their model equivalents.We demonstrate that this setup leads to significant improvements in the accuracy of initial conditions and enhances tropospheric forecast skill, extending medium-range predictability by 6 to 12 hours. We argue that the improved forecast skill can be attributed to a better representation of large-scale features in the stratosphere. Higher resolution DA and forecasts also lead to improved forecasting of extreme precipitation events, as well as the track and intensity of tropical cyclones. Furthermore, the improved initial conditions significantly increase the predictability machine-learning-based forecasts.The benefits of the higher resolution DA system are demonstrated using the case of Tropical Cyclone Otis, which made landfall as a Category 5 cyclone, but was predicted by most global and regional NWP models to peak only at tropical storm intensity – a major forecasting bust. By employing higher resolution 4D-Var DA, we were able to replicate the observed rapid intensification of TC Otis.
- Open Access
- 30 January 2025
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 151 (768)
Development of an offline and online hybrid model for the Integrated Forecasting System
In recent years, there has been significant progress in the development of fully data‐driven global numerical weather prediction models. These machine learning weather prediction models have their strengths, notably accuracy and low computational requirements, but also their weaknesses: they struggle to represent fundamental dynamical balances, and they are far from being suitable for data assimilation experiments. Hybrid modelling emerges as a promising approach to address these limitations. Hybrid models integrate a physics‐based core component with a statistical component, typically a neural network, to enhance prediction capabilities. In this article, we propose to develop a model‐error correction for the operational Integrated Forecasting System (IFS) of the European Centre for Medium‐Range Weather Forecasts using a neural network. The neural network is initially pre‐trained offline using a large dataset of operational analyses and analysis increments. Subsequently, the trained network is integrated into the IFS within the Object‐Oriented Prediction System (OOPS) so as to be used in data assimilation and forecast experiments. It is then trained further online using a recently developed variant of weak‐constraint 4D‐Var. The results show that the pre‐trained neural network already provides a reliable model‐error correction, which translates into reduced forecast errors in many conditions, and that the online training improves the accuracy of the hybrid model further in many conditions.
- 30 January 2025
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 151 (768)
Open AccessDevelopment of an offline and online hybrid model for the Integrated Forecasting System
In recent years, there has been significant progress in the development of fully data‐driven global numerical weather prediction models. These machine learning weather prediction models have their strengths, notably accuracy and low computational requirements, but also their weaknesses: they struggle to represent fundamental dynamical balances, and they are far from being suitable for data assimilation experiments. Hybrid modelling emerges as a promising approach to address these limitations. Hybrid models integrate a physics‐based core component with a statistical component, typically a neural network, to enhance prediction capabilities. In this article, we propose to develop a model‐error correction for the operational Integrated Forecasting System (IFS) of the European Centre for Medium‐Range Weather Forecasts using a neural network. The neural network is initially pre‐trained offline using a large dataset of operational analyses and analysis increments. Subsequently, the trained network is integrated into the IFS within the Object‐Oriented Prediction System (OOPS) so as to be used in data assimilation and forecast experiments. It is then trained further online using a recently developed variant of weak‐constraint 4D‐Var. The results show that the pre‐trained neural network already provides a reliable model‐error correction, which translates into reduced forecast errors in many conditions, and that the online training improves the accuracy of the hybrid model further in many conditions.
- 16 January 2025
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 151 (768)
The role of observations in ECMWF tropical cyclone initialisation and forecasting
In a unique undertaking, 13 Observing System Experiments (OSEs) were conducted at the European Centre for Medium‐range Weather Forecasts (ECMWF) during a 37‐day period in August–September 2020, using the same 9‐km resolution as the operational data assimilation. These OSEs provided a unified framework to evaluate and understand the role of different types of observational data for tropical cyclone (TC) forecasts, keeping in mind that removing or adding individual types is expected to have a limited effect due to resilience in the full observation system. The strongest impact came from withholding of microwave radiance observations, which on average degraded the TC position for all evaluated lead times and central‐pressure forecasts out to three days, with some further improvement from all‐sky assimilation of Advanced Microwave Sounding Unit (AMSU‐A) temperature sounders to the TC position forecasts. Scatterometer data improved the forecasts of both quantities, with further improvements from reduced spatial thinning. The impact for other satellite data such as Global Navigation Satellite System (GNSS)‐RO, atmospheric motion vectors (AMVs) and Aeolus were mixed/neutral. While the aircraft‐borne data (dropsonde, flight‐level data) were mostly beneficial, further investigations on how to best use this data in the TC core are needed. The assimilation of interpolated hourly central pressure and position data estimated by forecasters had an overall positive impact on both position and central pressure. While this study has illustrated the difficulties in obtaining strong results for observation impact on TCs for an operational data assimilation system, including new observations and improvements to the use of existing ones are necessary to keep improving the accumulated positive effect from the observation system. Additionally, future work is needed to advance diagnostic tools to judge the impact of individual observation systems on TCs.
- 16 January 2025
- Published by Wiley in Quarterly Journal of the Royal Meteorological Society
- Vol. 151 (768)
The role of observations in ECMWF tropical cyclone initialisation and forecasting
In a unique undertaking, 13 Observing System Experiments (OSEs) were conducted at the European Centre for Medium‐range Weather Forecasts (ECMWF) during a 37‐day period in August–September 2020, using the same 9‐km resolution as the operational data assimilation. These OSEs provided a unified framework to evaluate and understand the role of different types of observational data for tropical cyclone (TC) forecasts, keeping in mind that removing or adding individual types is expected to have a limited effect due to resilience in the full observation system. The strongest impact came from withholding of microwave radiance observations, which on average degraded the TC position for all evaluated lead times and central‐pressure forecasts out to three days, with some further improvement from all‐sky assimilation of Advanced Microwave Sounding Unit (AMSU‐A) temperature sounders to the TC position forecasts. Scatterometer data improved the forecasts of both quantities, with further improvements from reduced spatial thinning. The impact for other satellite data such as Global Navigation Satellite System (GNSS)‐RO, atmospheric motion vectors (AMVs) and Aeolus were mixed/neutral. While the aircraft‐borne data (dropsonde, flight‐level data) were mostly beneficial, further investigations on how to best use this data in the TC core are needed. The assimilation of interpolated hourly central pressure and position data estimated by forecasters had an overall positive impact on both position and central pressure. While this study has illustrated the difficulties in obtaining strong results for observation impact on TCs for an operational data assimilation system, including new observations and improvements to the use of existing ones are necessary to keep improving the accumulated positive effect from the observation system. Additionally, future work is needed to advance diagnostic tools to judge the impact of individual observation systems on TCs.
- Top Cited PapersOpen Access
- 14 June 2024
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 51 (12)
On Some Limitations of Current Machine Learning Weather Prediction Models
Machine Learning (ML) is having a profound impact in the domain of Weather and Climate Prediction. A recent development in this area has been the emergence of fully data‐driven ML prediction models which routinely claim superior performance to that of traditional physics‐based models. We examine some aspects of the forecasts produced by three of the leading current ML models, Pangu‐Weather, FourCastNet and GraphCast, with a focus on their fidelity and physical consistency. The main conclusion is that these ML models are not able to properly reproduce sub‐synoptic and mesoscale weather phenomena and lack the fidelity and physical consistency of physics‐based models and this has impacts on the interpretation of their forecasts and their perceived skill. Balancing forecast skill and physical realism will be an important consideration for future ML models.
- 14 June 2024
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 51 (12)
Top Cited PapersOpen AccessOn Some Limitations of Current Machine Learning Weather Prediction Models
Machine Learning (ML) is having a profound impact in the domain of Weather and Climate Prediction. A recent development in this area has been the emergence of fully data‐driven ML prediction models which routinely claim superior performance to that of traditional physics‐based models. We examine some aspects of the forecasts produced by three of the leading current ML models, Pangu‐Weather, FourCastNet and GraphCast, with a focus on their fidelity and physical consistency. The main conclusion is that these ML models are not able to properly reproduce sub‐synoptic and mesoscale weather phenomena and lack the fidelity and physical consistency of physics‐based models and this has impacts on the interpretation of their forecasts and their perceived skill. Balancing forecast skill and physical realism will be an important consideration for future ML models.
- Open Access
On some limitations of data-driven weather forecasting models
As in many other areas of engineering and applied science, Machine Learning (ML) is having a profound impact in the domain of Weather and Climate Prediction. A very recent development in this area has been the emergence of fully data-driven ML prediction models which routinely claim superior performance to that of traditional physics-based models. In this work, we examine some aspects of the forecasts produced by an exemplar of the current generation of ML models, Pangu-Weather, with a focus on the fidelity and physical consistency of those forecasts and how these characteristics relate to perceived forecast performance. The main conclusion is that Pangu-Weather forecasts, and possibly those of similar ML models, do not have the fidelity and physical consistency of physics-based models and their advantage in accuracy on traditional deterministic metrics of forecast skill can be at least partly attributed to these peculiarities. Balancing forecast skill and physical consistency of ML-driven predictions will be an important consideration for future ML models. However, and similarly to other modern post-processing technologies, the current ML models appear to be already able to add value to standard NWP output for specific forecast applications and combined with their extremely low computational cost during deployment, are set to provide an additional, useful source of forecast information. .
Open AccessOn some limitations of data-driven weather forecasting models
As in many other areas of engineering and applied science, Machine Learning (ML) is having a profound impact in the domain of Weather and Climate Prediction. A very recent development in this area has been the emergence of fully data-driven ML prediction models which routinely claim superior performance to that of traditional physics-based models. In this work, we examine some aspects of the forecasts produced by an exemplar of the current generation of ML models, Pangu-Weather, with a focus on the fidelity and physical consistency of those forecasts and how these characteristics relate to perceived forecast performance. The main conclusion is that Pangu-Weather forecasts, and possibly those of similar ML models, do not have the fidelity and physical consistency of physics-based models and their advantage in accuracy on traditional deterministic metrics of forecast skill can be at least partly attributed to these peculiarities. Balancing forecast skill and physical consistency of ML-driven predictions will be an important consideration for future ML models. However, and similarly to other modern post-processing technologies, the current ML models appear to be already able to add value to standard NWP output for specific forecast applications and combined with their extremely low computational cost during deployment, are set to provide an additional, useful source of forecast information. .