Storm track regimes and medium range predictability.
The mid-latitude storm track is essentially a breeding ground for mid-latitude storms: strong horizontal temperature gradients on the western side of the ocean basins provide the required energy reservoir for midlatitude storms to grow. Each storm depletes some, or perhaps all of this energy reservoir. After such a depletion event, the energy reservoir needs building up again. Recent work has shown that the energy reservoir and the storm-track activity are related like a nonlinear predator-prey model.
The storm track and the jet stream form a complex interlinked system that is ultimately steered by this predator-prey system of storms feeding on upstream temperature gradients. It is expected that the dynamical core of weather forecast models should do well for such processes, yet some persistent model biases remain: the mean jet latitudes are often biased in a way that is consistent with too little upstream storm track activity. Furthermore, onset and decay of large-amplitude patterns, such as blocking patterns, are still notoriously hard to predict, thus degrading medium range forecast skill.
How good are current models in following the observed chain of processes from the predator-prey stage of nascent storms to the decay phase along the highly deformed jet stream? To what extent does this chain of processes rely on accurate representation of diabatic processes in the storm-track? Can we build this dynamical-system based framework for storm track variability into a physics-informed machine learning system to enhance medium range predictability?
Large ensembles of machine learning forecasts for advanced nonlinear filters in atmospheric data assimilation
Recently, machine learning (ML) weather forecasting models have shown deterministic forecast skill approaching that of physics-based models, at a small fraction of the computational cost. This provides the opportunity to create very large ensembles of ML forecasts, with the potential to improve data assimilation (DA), the process of optimally combining forecasts and observations to estimate the state of the atmosphere.
Many atmospheric processes are known to be highly nonlinear and non-Gaussian, especially at convective scales. However, DA algorithms currently used in operational centres are only optimal in the linear, Gaussian setting. Advanced nonlinear DA algorithms such as particle filters can estimate the true state of the system without such restrictions, and there is thus great interest in their application to the atmosphere. However, their use has so far been considered unfeasible due to the large ensembles required.
The aim of this project is to test the use of ML ensembles for enabling advanced nonlinear DA methods, following previous work that has shown that a large ML ensemble can augment a small ensemble of physics-based forecasts. This will involve:
• Assessing the probabilistic skill of ML ensembles, and investigating how this skill can be improved during training or with transfer learning.
• Developing a particle filter algorithm which incorporates both physics-based and ML forecasts, following previous work on the multi-model ensemble Kalman filter.
• Experiments with this particle filter algorithm and large ML ensembles on models of increasing complexity, ultimately testing in a close-to-operational setting at the ECMWF.
Improving subseasonal-to-seasonal precipitation forecasts using hybrid and ML methods
Two important sources of predictability for S2S rainfall forecasts are the monsoon intraseasonal oscillation (MISO) and the Madden–Julian oscillation (MJO). Recent work has shown that MISO and MJO can be skillfully predicted using data-driven methods, and that these predictions can be used to significantly improve physics-based model forecasts (Bach et al., 2024). This project will apply new ML methods to improve upon the best data-driven predictions of MISO and MJO, and then use these forecasts within a multi-model data assimilation framework (Bach & Ghil, 2023) to correct global physics-based precipitation forecasts at S2S timescales. This project will also test pure ML-based predictions for S2S precipitation prediction.
An important challenge in improving the skill of global flood forecasts produced by hydrological models is the improvement of precipitation forecasts that are used as drivers. For drought, the TAMSAT-ALERT framework provides a method for combining meteorological observations and forecasts into skillful assessments of soil moisture deficit (Black et al., 2024). This project will use the improved precipitation predictions to drive flood and drought models, and assess the potential improvement for operational prediction.
Bach, E., & Ghil, M. (2023). A multi-model ensemble Kalman filter for data assimilation and forecasting. Journal of Advances in Modeling Earth Systems.
Bach et al. (2024). IImproved subseasonal prediction of South Asian monsoon rainfall using data-driven forecasts of oscillatory modes | PNAS. Proceedings of the National Academy of Sciences.
Black, E., Ellis, J., & Maidment, R. I. (2024). A computationally lightweight model for ensemble forecasting of environmental hazards: General TAMSAT-ALERT v1.2.1. Geoscientific Model Development.
Prediction of high impact extratropical cyclone clustering events
Clusters of extratropical cyclones, which follow one another in close succession, can result in rain falling on already saturated ground leading to costly flooding events. For example, in the UK, in December 2015 with storms Desmond, Eva and Frank, and more recently in February 2022 with storms Dudley, Eunice and Franklin. These cyclones cannot be treated as isolated meteorological events, because they have the potential to interact with each other – either directly or indirectly through feedbacks onto the background flow.
The PhD project aims firstly to assess the predictability of extratropical cyclone clusters by comparing global ECMWF ensemble forecasts with reanalysis data. The project will then focus on their interaction and feedback mechanisms, which are often underestimated in models. The recent introduction of a 100-member extended range ensemble at ECMWF provides an opportunity to test the hypothesis that model inadequacies in representing these feedbacks weaken responses to local and remote processes, ultimately limiting predictability. Using data from the ensemble will enable discernment of differences between well-performing and underperforming members. Additionally, the research will evaluate model errors in a cyclone-centred framework, offering an innovative approach to identifying poorly represented physical processes in the forecast model. There is also potential to evaluate forecasts against observations taken by aircraft during the January 2026 NAWDIC field campaign which will focus on mid-latitude atmospheric dynamics and their relation to high impact weather in the North Atlantic region.
Improving forecasts of high impact weather using new satellite radar observations
Deep convection is a key driver of high impact weather such as thunderstorms and floods. These cause considerable loss of life and damage. Poor forecasts are often caused by the modelled representation of interactions between convection and larger-scale dynamics. To improve forecasts and save lives, we need to use convective-scale observations to both initialize the forecast model and help reveal the causes of convection-related model biases.
In 2024, the new satellite mission, EarthCARE (Earth Cloud Aerosol and Radiation Explorer) was launched. EarthCARE provides unprecedented synergistic observations of the vertical profile of clouds and precipitation using radar and lidar. These observations can be used in a physics-constrained machine learning approach known as data assimilation (DA), to provide improved initial conditions for forecasts. The European Centre for Medium-range Weather Forecasts (ECMWF) already has a prototype DA system for these observations. This project will use this system to address questions that are critical for the best use of EarthCARE and other cloud-related observations:
• What are the uncertainties in the comparison between the forecast model and observations? How can these uncertainties be reduced? What is the best way to represent remaining uncertainties in the DA?
• What are the impacts on forecasts from assimilating EarthCARE and other cloud-related observations? How do convective system characteristics (size, lifetime) change due to the DA?
The project can be further developed according to the student’s interests. For instance, they could develop physics-based approaches for assimilating data from other EarthCARE instruments or investigate mathematical/computational questions relating to convergence of the DA.
Short range sea ice forecasting, using advanced models and new data
Over recent decades, the sea ice cover of the Arctic Ocean has reduced in extent, area, thickness and age. Climate models suggest the Arctic will become seasonally ice free within 20-30 years. Navigating the Arctic Ocean can considerably shorten trade routes, reducing carbon emissions and saving money. Shipping navigability has improved in recent years and this will continue, with increased interest in sea ice weather forecasts being inevitable.
This PhD project will investigate and develop a prototype for an unprecedentedly skilful sea ice forecast system, up to subseasonal timescales at various spatial resolutions (<50 km). This project will combine the latest developments in sea ice-ocean modelling such as melt ponds, form drag and floe size distribution (largely from UoR) with the latest satellite sea ice products, including CryoSat-2, SMOS, IceSat-2 and, for the first time, estimates of summer sea ice thickness. We will use a data assimilation scheme we developed based on the existing Parallel Data Assimilation Framework.
We will (i) improve the underlying sea ice-ocean model using assimilation of new, ground-truthed satellite sea ice data. This is done through examination of reanalysis increments in individual terms in the sea ice mass and momentum budgets, allowing model shortcomings to be identified and rectified; (ii) identify sources of predictive skill throughout the year but particularly in the navigable summer and shoulder seasons, building on our previous success using melt pond fraction to predict the sea ice minimum; and (iii) develop and assess a prototype sea ice forecast system.
The Roles of Tropical Waves in the Sub-seasonal Prediction of Tropical Cyclones
Tropical cyclones (TCs) are among the most hazardous weather systems worldwide. They pose a major threat to life, properties and ecosystems in coastal regions. As climate change amplifies exposure and vulnerability to natural hazards, gaining a better understanding of TC activity has never been more critical. However, the sub-seasonal timescale is still considered as a “predictability desert” for extreme weather due to limited understanding of the sources of S2S variability. The current prediction models have very limited or no skill in predicting the deviation of TC activity (i.e., anomaly) from seasonal climatology beyond a week.
A recent study (Feng et al., 2023) has shown that pre-existing tropical waves, including equatorial waves (EQWs), can explain the sub-seasonal variability of TCs from a global perspective. However, the nature of the wave-TC interaction and how the wave-associated dynamical and thermodynamical processes influence TC genesis and intensification is largely unknown. Furthermore, it is critically important to diagnose and understand how these interactions are represented in global forecast models, to better understand the large-scale drivers and potentially improve the sub-seasonal forecasts in practice.
The main research questions are: (1) what dynamical and convective processes are responsible for the modulation of TC activity (e.g. TC seeds, generation, and intensification) by tropical waves? (2) How well are these processes represented in the k-scale climate models and operational forecasts (e.g., Met Office and ECMWF systems)? (3) How sensitive is the TC-wave interaction to large-scale drivers (e.g., regionality, seasonality, El Niño Southern Oscillation)?
Feng, X., Yang, G. Y., Hodges, K. I., & Methven, J. (2023). Equatorial waves as useful precursors to tropical cyclone occurrence and intensification. Nature communications, 14(1), 511.
Influence of kilometre-scale atmospheric variability on the oceanic submesoscale eddy field
A new frontier in understanding and modelling oceans is the submesoscale range, i.e. small eddies (typically 1-10 km wide) which live in the surface mixed layer of the ocean. These eddies are known to play a critical role in controlling the depth of the mixed layer, regulating heat and CO2 uptake in the ocean interior, and bringing nutrients back into the surface layer, with long-term effects on climate.
Submesoscale eddies are generated by a family of instabilities that feed on the energy stored in larger-scale fronts and filaments (10-100 km). Previous studies have ignored the role of atmospheric variability in enhancing or dissipating the submesoscale field, in part because only the synoptic variability, which is much larger than 1-10 km, has been considered.
As coupled ocean-atmosphere models move toward higher resolutions, the scales of oceanic submesoscale eddies and of the newly resolved (down to kilometre-scale) atmospheric variability will converge. Key questions arise:
– What is the impact of this newly resolved variability on submesoscale eddies?
– Whether this new variability enhance or damp the submesoscale variability, which compensating mechanisms maintain the observed level of submesoscale variability?
– How these processes represented in current parameterizations used in the ocean components of climate models?
To address these questions, the successful student will use a combination of idealized submesoscale-resolving ocean simulation forced by kilometre-scale atmospheric forcing to explore fluid dynamic process, and state-of-the art coupled model simulations to work toward the real system. Depending on the student’s taste and abilities, the PhD work could take a more theoretical or a more modelling direction.
LLM-Forecast-Ensemble: LLMs as Components of Ensemble Forecasts
Atmospheric data assimilation (DA) plays a critical role in improving the accuracy of numerical weather prediction (NWP) by integrating observational data with physical models. Ensemble-based DA methods, such as the Ensemble Kalman Filter (EnKF) and Particle Filters (PF), are widely used to handle uncertainties. However, these traditional methods face limitations in managing highly nonlinear and non-Gaussian error distributions, particularly in complex atmospheric dynamics like extreme weather events. Recent advancements in machine learning (ML), particularly large language models (LLMs), present an opportunity to enhance these ensemble forecasting methods. LLMs, which have demonstrated a remarkable ability to capture intricate patterns and relationships in data, can be adapted to atmospheric forecasting. The novelty of this approach lies in using LLMs not only to process historical weather data but also to integrate contextual information from meteorological reports, localized text data, and computer vision analysis of satellite imagery. While prior work has explored integrating LLM into ensemble methods, this project will focus on employing LLMs as ensemble components for forecasting by leveraging contextual text and satellite imagery. The aim is to develop a robust method that enhances forecast accuracy by combining physical models with data-driven insights from LLMs. The specific objectives are:
• Integrate LLMs with regional contextual data and computer vision for ensemble forecasting: Adapt large language models to process historical weather data alongside regional contextual information, such as meteorological reports, news, and social media feeds, which provide localized insights into atmospheric conditions. Additionally, the LLM will use computer vision techniques to analyse satellite imagery in real-time, extracting visual features such as cloud formations and storm structures. This will allow the LLM to generate detailed, data-driven forecasts that incorporate both textual and visual inputs.
• Develop and validate an LLM-based model for atmospheric forecasting: Investigate the potential of using an LLM-based model as a standalone forecasting tool. The LLM will integrate data from various sources (textual and visual) and learn from historical weather patterns to produce forecasts independently of traditional physical models. The goal is to assess whether the LLM can generate accurate, data-driven predictions without the aid of numerical models and to evaluate its performance in handling nonlinearities in complex atmospheric phenomena.
• Hybrid ensemble forecasting with LLM-generated forecasts: Develop a hybrid ensemble system where the LLM-generated forecasts are combined with traditional physics-based models within an ensemble framework. The LLM will serve as one of the ensemble members, contributing its predictions based on contextual text and satellite imagery alongside forecasts from numerical models. The integration of data-driven forecasts with physics-based models will be evaluated for its ability to improve forecast accuracy, particularly in nonlinear and chaotic weather conditions.
Smoothing and Machine Learning for improving reanalyses through 20th Century
“Reanalyses” of historical atmospheric and climate properties (including eg. ocean, sea-ice, land surface) rely on assimilation methods which combine available (possibly sparse) observational data with a time evolving computer model, to produce the complete state of the system, – “maps-without-gaps”. Weather forecasting assimilation methods incorporate only a short window of recent observations (the model then projects the forecast). However a better reconstruction of past weather and climate events (including weather extremes) should also account for “future” observations made after the events of interest have occurred, especially in the pre-satellite era when fewer observations were available. Improved understanding of past events is critical for understanding weather and climate trends.
Rather than running dynamics models backwards in time “Smoothing” algorithms incorporate future observations statistically using cross-time error covariances derived from the model, but this is highly memory and computationally intensive and rarely used. However machine learning methods could make smoothing much more straightforward. You will investigate and develop new methods of “post-processed” smoothing, which are efficient and computationally cheap. Initial tests with simple systems refining methods will be extended to reconstructions of historical storms incorporating more hours-to-days of observational data, or improved ocean circulation states by using more months-to-years of observations (as ocean memory timescales are longer). Stored data from the operational weather centres can be used in these procedures. You will work with an experienced team of assimilation scientists at Reading, the Met Office and ECMWF, having access to state-of-the-art reanalysis systems and their products using local and national computing infrastructures.
Unleashing the potential of high-resolution sub-seasonal tropical cyclone predictions
Flooding and strong winds associated with tropical cyclones in the South-West Indian Ocean (SWIO) severely impact vulnerable Southern African communities every year. Access to reliable, actionable early warnings on sub-seasonal timescales are essential for supporting anticipatory actions and building resilience. This ambitious project seeks to address a critical question towards this: Do sub-seasonal predictions of tropical cyclones improve when harnessing regional kilometre-scale model configurations?
High-resolution modelling has significantly advanced our ability to forecast the intensity and track of tropical cyclones on daily timescales. However, whilst sub-seasonal forecasts can identify regions at risk of cyclones, they struggle to accurately predict observed intensification and longevity of cyclones. Addressing this gap, the project will first quantify forecast skill of existing sub-seasonal predictions for high-impact SWIO cyclones. Building on this, it will then perform regional kilometre-scale simulations driven with state-of-the-art sub-seasonal forecasts. The project will evaluate how and why this modelling approach improves predictions of key cyclone characteristics which remain challenging to forecast: genesis, track, intensification and associated rainfall. In collaboration with regional institutions – Meteo-Madagascar, East African Climate Prediction and Applications Centre (ICPAC) – we will contribute to international efforts to improve sub-seasonal tropical cyclone advisories. The student will engage with relevant international University of Reading projects (REPRESA and ACACIA), and support early warnings issued to vulnerable communities, aligning with global initiatives such as the WMOs Sub-seasonal to seasonal Applications for Agriculture and Environment (SAGE). This project aims to advance both scientific understanding and useful application of sub-seasonal predictions to mitigate the impacts of cyclones.
Improving subseasonal-to-seasonal forecasts in the tropics using AI
How can we use AIWPs to improve S2S ensemble forecasts in the tropics?
Subseasonal-to-seasonal forecasts (S2S; lead times of 15–60 days) are important in tropical countries for decision-making in agricultural, health, energy, and water security sectors. Current operational S2S forecasts are built using physics-based numerical weather prediction models (NWPs), and usually comprise either a single-model or multi-model ensemble that generates probabilistic forecasts. However, this approach leads to variable and often poor skill over many tropical regions.
AI-based weather prediction models (AIWPs) use a range of state-of-the-art machine learning models, trained on vast amounts of NWP-processed historical weather data, known as reanalyses, learning relationships without explicitly solving the underlying physical equations. As a result, AIWPs can produce weather forecasts with comparable skill to NWPs in large-scale fields but at a fraction of the cost. Some forecasting centres have thus started using them as a complementary addition to their existing NWP forecasts.
In this project, you will investigate the strengths and weaknesses of three types of AIWP at S2S lead times compared to traditional NWP based models: standard AIWPs (such as FourCastNet, trained to forecast up to ten days ahead), S2S AIWPs (such as FuXi-S2S, trained to forecast at S2S lead times), and hybrid models (such as NeuralGCM, which has an NWP at the core, but uses AI for parameterisations). You will then investigate how these are best integrated into existing ensemble prediction frameworks. This may take the form of a skill-weighted average, or an adaptive weighting based on regime-dependent skill.
Dynamics of damaging surface winds associated with European windstorms
The British Isles are located at the end of the North Atlantic stormtrack and are frequently battered by windstorms. The goal of this project is to understand and model the atmosphere and hazards at the scales where impacts of winds are felt, from hundreds of kilometres down to tens of metres. The new generation of km-scale coupled environmental prediction models will be confronted with new measurements that have matching high-resolution spatial coverage.
The international NAWDIC field campaign presents a major opportunity to deliver the high-resolution observations needed. NAWDIC will take place in Jan.-Feb. 2026, with aircraft being deployed from Ireland and ground observations spanning Ireland, the UK and France. A key focus of NAWDIC is the dynamics of descending dry intrusions within cyclones, including windstorms, and their interaction with the surface. The student will have the opportunity to participate as part of the mission science team.
The question is how momentum is transferred downwards to the surface. This aspect is not well understood or modelled, with consequences for forecasting high-impact wind events. In this project the student will take advantage of new Met Office and ECMWF model configurations to investigate the processes linking windstorm dynamics with severe surface impacts, using the campaign observations to evaluate the most realistic configurations. There will be scope, depending on the direction the student takes, to compare models with different architectures (e.g. testing LFRic, the new Met Office model) and to investigate the extent to which AI models can predict the windstorm structure associated with the impacts.
Exploring dynamical causes of signal-to-noise errors in subseasonal forecasts in the extratropics
There are substantial predictability gaps from weeks 3-4 of sub-seasonal forecasts, after predictability arising from atmospheric initial conditions has substantially reduced. However, recent studies have argued that the S2S models show more predictability in their ensemble mean signal than may be first evident from a scan of the ensemble, because the large-scale atmospheric circulation signal-to-noise rations seem to be too low (Garfinkel et al., 2024). These results are reminiscent of the signal-to-noise errors shown more robustly for seasonal forecasts of the winter season (e.g. Scaife & Smith, 2018).
In recent work, specific causes of signal-to-noise errors in wintertime seasonal forecasts of the North Atlantic region have been linked to pervasive upstream biases in the large-scale jet stream and blocking frequency/intensity (O’Reilly, in revision), which are particularly evident in their influence on predictability in the early winter, when the ENSO teleconnection is strongest (O’Reilly et al., 2024). This project will explore how these the mechanisms are related to predictability on subseasonal scales (potential “windows of opportunity”), including how the influence of lead-time dependent biases – specifically large-scale jet, blocking and SST anomalies – that emerge in the models are influencing predictability. The work will focus on understanding how the dynamics of regional jet shifts (and related blocking anomalies) depend on underlying limitations in the model dynamics, with implications for improving associated subseasonal predictions.
The project will involve analysing reforecasts from operational S2S forecasting systems and performing targeted sensitivity experiments with complex models; we will also test specific dynamical mechanisms that emerge in idealised models.
Multi-Scale Parameterization for the Convection Grey Zone
The project will assess the prospects for a multi-scale parameterization suitable for use in the convective grey zone: what benefits could that provide, and what properties would the scheme need to have to realize these benefits? The concept is feasible only because of the functionality of LFRic, which can solve the model dynamics and (parts of) the physics on different grids. Thus, we will pioneer use of LFRic within academia and tackle scientific problems that were not previously accessible.
It is characteristic of a grey zone that many studies show clear benefits at o(1km) scale when switching off the convective parameterization, while many others point to serious drawbacks. Our underpinning assumption is that parameterizations do provide valuable information if applied on a scale where their assumptions are valid. We envisage running the parameterization at o(10km), with information derived from it leading to additional tendencies imposed on the dynamics grid of o(1km). We call this a multi-scale parameterization because radiation, microphysics etc continue to operate at o(1km).
We will use the new CoMorph parameterization in online diagnostic mode within sub-km simulations. We will compare its predictions of convective properties with those produced by the resolved convection, and assess dependencies on the space-time scales of the inputs provided to CoMorph. A machine-learning element might be useful to capture the relationships. From these findings, concepts for multi-scale parameterizations will be devised and tested in o(1km) simulations, establishing what structural characteristics are necessary for a useful multi-scale representation that better matches sub-km results.
Predicting harmful algal toxins in the ocean through satellite data assimilation into a novel biogeochemical model
This project will develop a novel modelling framework to address the challenges of understanding and predicting harmful algal blooms (HABs) and the toxins they release in the marine ecosystems. HABs disrupt vital ecosystem services, leading to substantial revenue losses for fisheries and aquaculture productions, and pose serious human health risks through the consumption of contaminated seafood. This development is needed to better understand the unpredictable behaviour of HAB forming algal species in response to environmental changes and to simulate algal toxins at regional and global scales. The project will leverage two parallel research strands currently being developed by the supervisors from UoR and MO: (1) a novel biogeochemical model explicitly representing toxin-producing HAB species in the ocean and capable of simulating their harmful dynamics, building on the framework of an intermediate complexity ocean biogeochemical model; and (2) advanced spatio-temporal maps of biological carbon held in marine phytoplankton derived from Earth Observation of ocean colour. By integrating these strands, the student will assimilate advanced EO products into the new biogeochemical model to predict the dynamics of HABs and associated algal toxins in specific regions of the global ocean. The assimilated model output will map the high-risk regions for HABs and algal toxin concentrations, which can help management of the fisheries and aquaculture production in the affected regions.
Modelling Tree Species Distribution from Space to Combat Climate Change
Changes in tree species and forest cover have indirect biophysical effects on the climate, e.g., by impacting low-level convective cloud formation1, which may have ramifications for the planetary hydrological cycle. Therefore, accurate monitoring of forest composition is essential for understanding these biophysical effects for devising informed strategies. Several studies exist which make use of remote sensing and EO data to systematically quantify forest cover changes2,3, but we have little information on potential changes in tree species composition. Knowing this is important to predict future climate change effects and track forest exploitation where the most valuable species are removed or replanted preferentially. Most studies in this context focus on investigating the impact of climate change on a few commercially important tree species and often rely on forest inventory data of individual stands to extract ground truth4,5. Moreover, the spatial resolution of existing datasets is also coarse, e.g., the current global state-of-the-art dataset “EU-Trees4F”6 with the highest number of 67 tree species in Europe, provides the tree distribution maps at 10 km spatial resolution. Within this project’s scope, we intend to overcome these shortcomings and aim at enabling large-scale tree species classification maps and deriving their future projections. Later, we will fuse these maps and associated projections with the available climate data, e.g., the cloud fractional cover obtained from a climatology of satellite observations7, to yield crucial insights about key geoinformation, such as changes in tree species, to allow a more holistic understanding of the subtlety of global changes in response to climate change.
1 Duveiller, G., et al., 2021. Revealing the widespread potential of forests to increase low level cloud cover,
Nature Communications, 12, 4337.
2 Wernick, I.K., et al., 2021. Quantifying forest change in the European Union. Nature, 592, E13–E14.
3 Decuyper, M., et. al., 2022. Continuous monitoring of forest change dynamics with satellite time series,
Remote Sensing of Environment, 269, 112829.
4 Axelsson A., et al., 2021, Tree species classification using Sentinel-2 imagery and Bayesian inference,
International Journal of Applied Earth Observation and Geoinformation, vol. 100.
5 Hemmerling, J., et al., 2021. Mapping temperate forest tree species using dense Sentinel-2 time series,
Remote Sensing of Environment, vol. 267, 112743.
6 Mauri, A., et al., 2022. EU-Trees4F, a dataset on the future distribution of European tree species, Scientific Data, 9, 37.
7 Stengel, M., et al., 2017. Cloud property datasets retrieved from AVHRR, MODIS, AATSR and MERIS in the framework of the Cloud_cci project, Earth System Science Data, 9, pp. 881–904.
A Machine Learning Framework for Multiscale Process Based Prediction in Earth System Models
Earth System Models (ESMs) are important tools that help us predict future climate events particular for impacts studies. However, ESMs are often coarse scaled due to computational constraints and do not capture subgrid scale processes essential for predicting regional impacts. Effective methods to downscale from coarser resolutions can allow us to run larger ensembles at relatively lower costs to obtain high resolution data.
In this project, we propose a machine learning (ML) framework to first statistically downscale key climate drivers at lower resolution (global 10km and global 5km) and then to evaluate the downscaled projections so we can better understand which processes require higher resolutions (1 km or finer) for improved regional impacts assessments. We propose to develop a temporally coupled Convolutional Neural Network based architecture for multivariate downscaling to answer the following key research questions:
1) How fit for purpose is the downscaled data prediction and how does its skill compare against higher resolution models? We will use observations to evaluate how well the system captures characteristic features of statistically sparse and small-scale convection systems such as rainfall intensity, duration and spread.
2) Can we gain an understanding of what subgrid scale process effects are effectively captured by the ML architecture? We propose to use an interpretable ML based approach comprising of feature importance ranking methods and emulators to investigate causality relationships. Such an approach will build trust in the downscaling and can help identify the underpinning physical processes
Simulating the convective environment for driving East African storms at the kilometre scale
The monsoon affects large parts of East Africa during two rainy seasons as the intertropical convergence zone passes by. While it is less studied than its West African counterpart, it is no less important for the local population, both for water supply and in damages from extreme events. The Lake Victoria region, for example, is a hive of economic activity, especially fishing, but suffers many intense convective storms. We already know that convective-scale models can better simulate the diurnal cycle timing and intensity of rainfall in East Africa, and the local circulations involved in the initiation of storms. This PhD will seek to understand the scale interactions between large-scale drivers such as equatorial waves, the Madden-Julian Oscillation or Indian Ocean dipole and the convective environment over East Africa in which storms develop. We will aim to answer the question, “Is the kilometre scale necessary to predict extremes in East African rainfall?”. The student will use tools such as the new ensemble of pan-African convection-permitting model experiments to assess how large-scale drivers alter the convective environment and test the improvement as resolution is increased to the kilometre scale. Working with the NCAS National Capability International Programme, the student will track mesoscale convective systems and make comparisons with observed systems, before assessing how the mechanisms behind scale interactions work, testing aspects of the convective environment such as windshear, temperature structure and surface conditions. This PhD will contribute towards better forecasting of hazardous weather events.
Using Machine Learning to obtain novel parameter sets for innovative theories of soil water- and heat transfer
Accurately predicting the weather depends on a thorough understanding of how the land surface and atmosphere interact. These interactions occur through the Earth’s “skin-layer”—the topsoil and vegetation—that exchanges energy, water vapour, and momentum with the atmosphere. Soil properties play a significant role in these exchanges, yet current weather prediction models rely on inadequate equations and parameter sets that need urgent improvement.
You will collaborate with leading scientists to test a novel unified theory of soil heat and water flow. Your PhD research will focus on developing innovative methods to create global, high-resolution soil parameter sets using unconventional data sources. These sources may include engineering data (such as soil plasticity and mineral composition from construction projects world-wide), vegetation data and soil information obtained from remote sensing, and even images from social media. Specifically, you might use machine learning algorithms to analyse satellite imagery and derive soil properties by interpreting patterns and electromagnetic properties of vegetation or develop computer vision models to process photos of soil posted on social media, extracting information about colour and cracking patterns that contain information on the soils’ composition and properties (e.g. rainwater infiltration rates). By training these data-based models on diverse datasets, you will be able to predict soil parameters that are critical for improving weather prediction models.
We are seeking a student with strong data analysis and mathematical skills who is excited to apply novel machine learning techniques to earth sciences and contribute to groundbreaking research that enhances numerical weather predictions.
Understanding (and predicting) tropical cyclone genesis and evolution in the Indian Ocean at coupled km-scale. Project UPTICK
Future projected changes in tropical cyclones (TCs) are uncertain due to our limited understanding of the processes that drive TC precursors, genesis and intensification and how these processes interact with the large-scale environment. Existing literature suggests that, at least in some regions, TC precursor systems evolve from organised convective systems, which are typically unresolved or poorly represented in climate model simulations. TC intensification rates are also strongly influenced by model resolution, but as demonstrated by hurricane Otis, limitations in current modelling predictions can lead to poor forecasts and devastating impacts.
In this project we will use the Met Office’s new regional, coupled km-scale modelling capability to answer the research question: What are the processes that drive TC genesis and intensification in the Indian Ocean, how might these change in future and what are the implications for TCs landfalling Africa? Despite their devasting impacts, little is known about how TCs will change in the region. The CP4A (Convection-Permitting 4km-model for Africa) domain will be extended across the Indian Ocean basin to encompass TC genesis regions. With km-scale resolution and atmosphere-ocean coupling, the model should be capable of representing the key processes for TC evolution, and hence enable new insights and understanding. Simulations using present-day and future climate conditions will be used to test how these processes and interactions may change in future. These new simulations will complement atmosphere-only CP4A simulations to identify the impacts of coupling on TC evolution, including potential re-intensification in the Mozambique channel, evidenced recently by Cyclone Freddy.
Assessing the role of resolution in chemistry-climate model performance and the implications for predictions of future climate and air quality
Earth System models (ESMs) are central to the quantification of historical climate change and prediction of future climate change and air quality under different societal development scenarios. Thus, they inform policies regarding the level of action, e.g. pollutant emission reductions, required to avoid dangerous change. However, while useful, ESMs suffer from biases when compared to present day observations, raising concerns regarding their ability to simulate accurately past or future climate. In turn, this leads to uncertainty in our understanding of how humans have so far altered atmospheric composition and climate, and mitigation policies necessary in the future.
One of the potential sources of bias is the relatively low resolution (grid cells of ~104 km2 at the equator) at which ESMs typically operate. This poses challenges for modelling processes which vary over smaller length scales, for example atmospheric chemistry and aerosol production in urban environments, which are critical for air quality and climate predictions.
We propose to run parallel simulations in the state-of-the-art United Kingdom Earth System Model (UKESM) at resolutions from the standard coarse approach down to the kilometre scale. This will allow us to answer two key questions. First, to what extent are biases in “standard” UKESM improved by increasing model resolution? Second, what are the implications for future predictions of climate and air quality? This will provide valuable context for interpreting the predictions of future climate change and air quality from recent model intercomparison studies and provide a more robust evidence base for policymakers.
Simultaneous evaluation of clouds and radiation in high-resolution models using EarthCARE
Clouds are a major source of uncertainty in both weather and climate prediction via their interaction with solar and thermal-infrared radiation. Launched in May 2024, the EarthCARE satellite offers the most detailed ever estimates of the properties of clouds and precipitation, including how much water they contain, mean particle size, whether they are composed of liquid or ice (or a mixture of both) and for the first time even the density and fall speed of snowflakes. Simultaneously, EarthCARE measures the reflected sunlight and emitted thermal radiation emerging from cloud top, which offers the exciting new opportunity to evaluate the properties of clouds in weather and climate models at the same time as their impact on radiation. This way radiation errors can be traced more directly to their cause than has been possible in the past. In this project, the student will use EarthCARE to evaluate the ECMWF model used for global weather forecasts and the Met Office model used for both high-resolution weather forecasts and global climate projections. When problems are identified in the way clouds are represented in the models, it will be possible to modify them and perform new simulations to test not only whether they agree better with EarthCARE measurements, but also the impact on the accuracy of weather forecasts more generally. This project is ideal for a student wishing to gain a wide range of skills spanning observations, modelling and theory.
When Two Modes Collide: The role of the Madden-Julian Oscillation and El Nino-Southern Oscillation in sub-seasonal predictability in the Maritime Continent
Maritime Continent (MC) region is highly vulnerable to high impact weather associated with heavy precipitation. Improved predictions on sub-seasonal timescales (2-4 weeks) could provide early warnings to allow action to mitigate the impacts of these events.
The MC sits at the centre of action of the Madden-Julian Oscillation (MJO, the dominant mode of Tropical intra-seasonal variability), and the El Niño-Southern Oscillation (ENSO, the dominant mode of tropical interannual variability). These large-scale modes modulate the precipitation over the MC by modulating the local weather systems (e.g. equatorial waves and land-sea breezes) that drive the precipitation. Much is known about how these modes act individually to modulate the weather but these modes don’t act in isolation and little is known about the combined effects of these modes on the weather of the MC.
This project will use a combination of observations, multi-decadal reforecasts from the ECMWF, and convection permitting simulations to:
• Understand how the MJO and ENSO and other modes of interannual variability combine to modulate the weather over the MC
• Understand errors in how these processes are represented in the ECMWF forecast system
• Develop regime-dependent methods using traditional and/or Machine Learning approaches to bias correct sub-seasonal forecasts.
This project is timely for two particular reasons.
• We now have access to multi-decadal reforecasts with many realisations in a model that simulates a reasonable MJO, and cover a long enough period to sample many ENSO events.
• Cutting-edge global convection permitting models allow us to investigate interactions between convection and much larger circulations.