Extreme climatic events, environmental degradation and socio-economic inequalities exacerbate the risk of infectious disease epidemics. We lack the evidence-base to understand and predict the impacts of extreme events and landscape changes on disease risk, leaving communities in climate change hotspots vulnerable to increasing health threats. This is in part due to a lack of ‘ground truth’ data describing environmental change in remote and under-resourced areas, as well as a lack of trained research software engineers and data scientists. HARMONIZE will convene a transdisciplinary community of stakeholders, software engineers and data scientists to develop cost-effective and reproducible digital infrastructure for stakeholders in climate change hotspots, including cities, small islands, highlands and the Amazon rainforest. We will strategically undertake one-off longitudinal ground truth data collection using drone technology and low-cost weather sensors, to improve classification algorithms and downscaling of coarser-resolution environmental datasets (e.g., satellite images, climate reanalysis and forecasts). We will then harmonize this post-processed data with socio- economic and health data in an automated workflow packaged for users in bespoke hotspot-specific toolkits. These sustainable tools will facilitate generation of actionable knowledge to inform local risk mapping and build robust early warning and response systems to build resilience in low-resource settings.

Many problems of paramount importance in diverse fields can be described by matrices and their direct solution involves the computation of some function over these matrices. Examples are: the solution of integer or fractional partial differential equations; analysis of complex networks; etc. For many real instances , classic algorithms cannot be employed as they do not scale well to solve these problems. Specifically, intrinsic data-dependencies make them not amenable for large-scale parallelization. Hence, in practice people resort to alternative methods to arrive at some solution, in most cases an approximate solution. Often, it is even hard to gauge the quality of this approximation. We aim to make fundamental contributions to a method for computing matrix functions based on their Neumann series, consisting of weighted sums of powers of the matrix [7]. Moreover, we use a Monte Carlo algorithm where a matrix to the power k is computed using random walks of length k over the matrix [8]. Monte Carlo methods have the preeminent facet of being embarrassingly parallel, achieving a high degree of efficiency on a parallel system. In our method, each walk is independent from one another and can be computed completely in parallel. This class of algorithms positions itself as the most appropriate to exploit the computational power of new exascale machines to the fullest. A second, crucially important feature of this Monte Carlo method, that sets it apart from the classic algorithms, is that it allows calculating individual entries of the result, avoiding the computation of the entire matrix. While real-world matrices are typically sparse, the result matrix will, in general, be dense. Thus, for large problems the representation of the output matrix in memory is in itself unfeasible, rendering its computation using classic methods impractical. With our approach we can, for instance, compute the trajectory of a single state variable in a dynamic system (eg, the voltage of a node in an electrical network) or the communicability of a single node in a network. However, there are two major limitations of Monte Carlo methods. First, the handling of very large instances requires the splitting of the matrix across machines which implies that the walks may need to be continued on another machine. This may entail a significant communication overhead that needs to be aggressively mitigated. The second is that these methods suffer from a very slow convergence rate, hence possibly implying a large number of walks for the desired result precision. In this research project we propose to address these two problems, thus bringing Monte Carlo methods to the forefront as enablers of solutions for up to now intractable problems. We have some initial work on the first of these problems [3], where a distributed implementation of the method was developed that is able to hide communication costs by aggregating walks in asynchronous messages. Still, a more effective solution should be possible, namely by taking advantage of more recent tool features [5,35], such as tasking with data-dependencies and support for tasks in accelerators. We also have recent work that addresses a specific instance of the second problem. We have recently made important contributions [1,2] on adapting multilevel techniques to improve the convergence rate of the Monte Carlo when computing the exponential of a matrix. In this proposal, we aim to study whether this approach can be further optimized and extended to other matrix functions, and we plan to explore other variance reduction techniques to improve the overall convergence rate of the method. In yet another research avenue, we have found that using this Monte Carlo approach with some modifications we are able to efficiently solve systems with derivatives of fractional order. Fractional calculus has a wide range of practical applications, eg [27], and has not yet seen a more widely spread adoption due to its intrinsic computational difficulty. Our expectation is that our idea can have a broad impact and be a significant contribution from this project, and we plan to apply it to both Magnetic Resonance Imaging (MRI)[4] and the modeling of chemical reactions [34]. We propose to demonstrate the usefulness of our method by using it to solve two other highly relevant and current world problems. One is the analysis of complex networks, namely to compute the communicability of a network and the centrality of a node. The first is related to the exponential of the adjacency matrix of the network and the latter to the inverse of that matrix. The other area in which an effective method of determining the exponential of a matrix can be a game changer is in the analysis of dynamic systems. Solving the matrix exponential of the system may allow the direct computation of the system state at a given time, and hence that of any system output at any time in the future, instead of having to simulate the system over time as is done today. Diversos problemas de áreas distintas e da mais elevada relevância podem ser descritos por matrizes, e a sua solução de forma direta envolve o cálculo de alguma função sobre essas matrizes. Exemplos são: solução de equações diferenciais parciais de ordem inteira ou fracionária; análise de redes complexas; etc. Para a maioria das instâncias de interesse, os algoritmos clássicos não podem ser utilizados na sua resolução, pois as dependências de dados intrínsecas aos métodos tornam impossível a sua paralelização em larga escala. Assim, na prática, recorre-se a métodos alternativos para chegar a alguma solução aproximada, cuja precisão é por vezes difícil de estimar. Nesta proposta propomo-nos fazer contribuições fundamentais para um método de cálculo de funções sobre matrizes utilizando as respetivas séries de Neumann, em que o resultado da função é obtido pela soma ponderada de potências da matriz [7]. Neste método a potência k de uma matriz é calculada usando um método de Monte Carlo, que consiste na geração de caminhos aleatórios de comprimento k sobre a matriz [8]. Os algoritmos baseados em Monte Carlo têm a característica de serem trivialmente paralelizáveis, alcançando um nível de eficiência muito elevado em sistemas paralelos. No nosso caso, todos os caminhos são independentes, podendo ser calculados totalmente em paralelo. Este tipo de algoritmos pode ser considerado como dos mais apropriados para explorar ao máximo o poder computacional dos futuros supercomputadores a nível exa-escala. Uma segunda característica de importância crucial dos métodos baseados em Monte Carlo, que os distingue dos algoritmos clássicos, é permitirem o cálculo de valores individuais do resultado, evitando assim o cálculo da matriz completa. Embora as matrizes de interesse sejam tipicamente esparsas, a matriz resultado será, em geral, densa. Assim, para problemas reais, a representação em memória da matriz de saída é só por si inviável, tornando o seu cálculo por meio de métodos clássicos impraticável. Com a abordagem de Monte Carlo é possível, por exemplo, calcular a trajetória de uma única variável de estado num sistema dinâmico, ou a comunicabilidade de um nó individual numa rede. No entanto, os métodos de Monte Carlo apresentam duas limitações importantes. A primeira é que o tratamento de instâncias muito grandes requer a repartição da matriz entre nós computacionais, o que pode implicar que os caminhos tenham que ser continuados noutro computador, com um potencial custo de comunicação significativo que é preciso mitigar. A segunda é que estes métodos sofrem de um ritmo de convergência lento, o que pode exigir um grande número de caminhos para uma dada precisão de resultado. Neste projeto propomos atacar estes dois problemas, procurando trazer os métodos de Monte Carlo para a linha da frente em termos facilitadores de soluções para problemas até agora intratáveis. Temos já algum trabalho inicial sobre o primeiro desses problemas [3], onde foi desenvolvida uma implementação distribuída do método capaz de esconder custos de comunicação pela agregação de caminhos em mensagens assíncronas. Ainda assim uma solução mais eficiente é possível nomeadamente tirando partido de funcionalidades mais recentes das ferramentas [5,35]. Recentemente, fizemos contribuições importantes [1,2] na adaptação de técnicas multinível para melhorar o ritmo de convergência de Monte Carlo ao calcular a exponencial de uma matriz. Na presente proposta, pretendemos estendê-la para outras funções, e planeamos explorar outras técnicas de redução da variância. Noutra via de investigação, descobrimos que, usando uma abordagem similar, somos capazes de resolver sistemas com derivadas de ordem fracionária com eficiência. O cálculo fracionário tem uma ampla gama de aplicações práticas, por exemplo [27], não tendo ainda uma adoção mais geral devido à sua intrínseca dificuldade computacional. A expectativa é que a nossa ideia possa ter um amplo impacto e ser uma contribuição significativa deste projeto, usando como demonstradores a área da Ressonância Magnética (RM) [4] e nas reações químicas [34]. Propomos demonstrar a utilidade do método aplicando-o na resolução de dois outros problemas relevantes: a análise de redes complexas, nomeadamente no cálculo da comunicabilidade de uma rede; e da centralidade de um nó. O primeiro está relacionado com a exponencial da matriz de adjacências da rede e o segundo com a inversa dessa matriz. Um outro problema em que um método eficaz de determinar a exponencial pode ser revolucionário é na análise de sistemas dinâmicos. A exponencial da matriz que modela o sistema pode permitir o cálculo direto do estado do sistema em instantes específicos e portanto, de qualquer variável do sistema em qualquer momento no futuro, em vez de se simular o sistema ao longo do tempo como realizado hoje.

The Atlantic Meridional Overturning Circulation (AMOC) is a crucial component of the climate system due to its role in heat and salt transports, as well as its role in transporting and storing carbon. Variability in the strength of AMOC has been linked to important climate impacts, for instance, the number of Atlantic Hurricanes, anomalous Sahel precipitation, and European weather. Therefore, improved predictions of the AMOC would have important societal benefits. Despite its importance, the predictability of the AMOC remains relatively unexplored on timescales from one season to 10 years ahead, and many uncertainties persist in our understanding of AMOC variability. For example, we are unsure of the relative importance of different processes in driving AMOC variability on different timescales and latitudes, nor how predictable they are in state-of-the-art forecasting systems. Recent studies have provided considerable evidence that the atmospheric circulation in the North Atlantic is much more predictable than previously thought on these timescales. However, the predicted signals are far too small (the so-called signal-to-noise paradox) and predictions need to be calibrated to provide credible forecasts of society relevant variables, such as surface temperature. Given that atmospheric circulation is a key driver of AMOC, then it follows that AMOC predictions on these timescales may also suffer from similar signal-to-noise issues. Furthermore, predictions of AMOC, and its climate impact, could be improved by extending the published statistical calibrations to the ocean circulation. ALPACA will utilise AMOC observations (RAPID and OSNAP) and observation-based AMOC reconstructions to assess the quality of current AMOC forecasts in state-of-the-art seasonal and decadal prediction systems. Furthermore, we will evaluate the processes that contribute to skill and assess their consistency across models. We will also use new simulations to better understand the relative roles of different processes in driving observed variability on different timescales, and we will leverage new large ensemble simulations to quantify the role of external forcing in driving AMOC variability and change. Finally, by exploiting this new understanding, we will determine whether seasonal-to-decadal predictions of AMOC and its climate impacts can be improved through physically-consistent statistical calibrations that reduce the signal-to-noise errors in predictions. ALPACA is a collaboration between the National Centre for Atmospheric Science at the University of Reading, The National Oceanography Centre Southampton, The University of Exeter, and the Met Office Hadley Centre from the U.K., and The National Center for Atmospheric Research and the University of Miami, from the U.S, and the Barcelona Supercomputing Center from Spain.

Decarbonising the transport sector is a top priority worldwide. The difficult-to-decarbonise transport applications (including mainly shipping, road freight and aviation) emit more than 50% CO2 of the entire transport sector. Among efforts on developing low-emission fuels, liquid synthetic fuels that can massively reduce pollutant emissions are drawing increasing attention, as they can be integrated into the current transportation system using existing infrastructure and combusted in existing engines (such as diesel engines for optimal fuel economy) with minor adjustments as drop-in fuels. Liquid synthetic fuels such as oxymethylene ethers (OMEx, which possess liquid properties similar to diesel when x=3-5) can be produced from a range of waste feedstocks and biomass, thereby avoiding new fossil carbon from entering the supply chain. OMEx can also be produced as an electrofuel (or e-fuel), thereby used as a sustainable energy carrier. However, due to the lack of complete knowledge of the physicochemical properties associated with the fuel composition variability, i.e. variation in the oligomer length (the x value of OMEx) and the composition variation of OMEx-diesel blends in real engine environment, there are challenges in utilising OMEx in practical engines, mainly in engine and its operation adjustments for optimal performance and minimal pollutant emissions. To address the technical issues of OMEx utilisation, accurate information on physicochemical properties and pollutant emissions of the synthetic fuels over the engine operational ranges is mandatory, but this is not readily available. This project is intended to obtain a thorough understanding on liquid synthetic fuel utilisation. The project will address the fundamental challenges in utilising renewable synthetic fuels, in particular OMEx and the associated OMEx-diesel fuel blends. The study will follow a combined modelling / simulation - experimentation approach, predicting the physicochemical properties including emission characteristics of the alternative fuels using molecular dynamics simulations, tailor-made experimentation for first-hand information on fuel utilisation, and establishing a database / mapping to guide the synthetic fuel utilisation in real engines over a wide range of conditions using machine learning.

The Subpolar North Atlantic (SNA), which is the region of the Atlantic Ocean between 45-65N latitude, is a highly variable region. Surface temperatures and surface salinity here have varied on a range of time-scales, but the changes are dominated by large and slow changes on decadal or longer timescales. This decadal timescale variability appears to form a key component of a larger climate mode, the Atlantic Multidecadal Variability, which has been linked to a broad range of important climate impacts, including rainfall in the North African and south Asian monsoons, floods and droughts over Europe and North America, and the number of hurricanes. The SNA is also one of the most predictable places on Earth at decadal timescales, which suggests the potential for improved predictions of regional climate and high-impact weather years ahead. However, the origins of this variability in the SNA, and the processes controlling its impacts, are far from fully understood. There is significant evidence to suggest that anomalous heat loss from the subpolar North Atlantic Ocean to the atmosphere can instigate a cascade of changes across the North Atlantic basin in both the ocean and atmosphere. For example, changes in the SNA can change the strength of the ocean circulation to the south, affect the northward transport of heat and freshwater in the North Atlantic, and subsequently affect the upper ocean temperatures and salinity across the whole North Atlantic basin, and into the Arctic. Changes in the subpolar North Atlantic surface temperature are also thought to affect the atmospheric circulation - i.e. wind patterns - in both summer and winter. However, observational records are very short, and so there are significant problems with understanding causality, and considerable uncertainty about how well many of the important processes are represented in current climate models. WISHBONE will make use of new advanced climate simulations and forecast systems to make progress in understanding the impact of the subpolar North Atlantic on the wider North Atlantic basin. It will also test specific hypotheses related to understanding the specific role of heat loss over the subpolar North Atlantic in driving changes throughout the basin including the role of surface anomalies in driving wind patterns. WISHBONE is a collaboration between the National Centre for Atmospheric Science at the University of Reading, The National Oceanography Centre Southampton, The University of Oxford, and The University of Southampton from the U.K., and The National Center for Atmospheric Research, from the U.S.