AI4NetZero

By M. Ali, Research Associate

Self-Learning Digital Twin for Agricultural land is a large project with seven work packages that include different teams creating and consuming machine learning models and curating datasets. These teams need to collaborate with each other in terms of models and datasets. Additionally, the digital twin needs to accommodate evolution of models and new models that are designed and developed by other teams. This gives us a stack of AI models, datasets, physics and mathematical equations and transformers that need to be integrated with the Digital Twin (DT) with minimal effort.

Traditionally, current solutions create monolithic digital twins that are tightly coupled with models and datasets. These monolithic solutions package all the models/datasets together in a single software that is designed to run on a single system. Few problems with this approach are:

• A monolithic digital twin can’t add new models after it has been created
• The core functionality of a monolithic digital twin is fixed and doesn’t allow the user to add new functionality to the existing models
• It could only answer questions for which it has been designed and developed
• It is not performance efficient, as all the models and datasets are run on a single machine that could become a bottleneck for large number of models and their datasets during training and inference

To circumvent these issues, a distributed agent oriented digital twin workflow is proposed that has three main components that are designed to run on different machine making the Digital twin a distributed system:

• The front end (GUI)
• The backend (Business logic)
• The middleware (workflow and API)

The focus of this work is on the middleware that includes the workflow model and the associated application programming interface (API) that enables communication between the front end and the middleware workflow model. A digital twin system workflow model is a directed acyclic graph with well-defined nodes and communication interfaces. Each node in the system represents a computational entity. These nodes communicate using interfaces of two types:

• A data interface to connect any dataset or IoT stream with the system model
• A model interface to specify the inputs and outputs of a computational node i.e. a machine learning model, transformer, or pre-processing unit

Nodes in the workflow model represent machine learning, mathematical or physical models or data sources. These nodes are designed to be active agents that are ready to process the data in their inputs to produce the outputs. Each agent can talk to other agents through a well-defined interface of two types:

• Generic common interface
• Agent specific interface

The generic interface allows agents to have common functionality and to enable the discovery of agent specific functionality using interfaces. The interfaces can be sent and received dynamically on agent request

The workflow models allow the digital twin to be flexible and extensible to new models and datasets. It also enables different components and workflow nodes to be managed and run on multiple machines to improve the performance. Finally, it allows dynamic discovery of the model features and allows creation of domain specific scenarios dynamically and to run and test them to answer multiple questions.

By Dr Georgios Rigas, Principal Investigator

The energy and transport systems that underpin key renewable or zero-emission involve multi-physics systems (i.e., reacting hydrogen with multi-phase flows relevant to aviation), are multi-scale and turbulent, (i.e., the optimisation of wind farms’ power output), and need adaptive solutions (i.e., the intelligent cooperation of road vehicles to minimise aerodynamic resistance). On the one hand, fast solutions to these problems are available through cheap models, but these solutions are not optimal. On the other hand, optimal solutions to these problems can be achieved (in principle) by simulation or experimentation, but the cost and time required are prohibitive.

In this project, we seamlessly combine two disciplines: physics-based modelling, which is generalisable and robust but may require tremendous computational cost, and machine learning, which is adaptive and fast to be evaluated but not easily generalisable and robust. The intersection of the two spawns scientific machine learning, which maximises the strengths and minimises the weaknesses of the two approaches.

The outcome of this is real-time digital twins that enable the static or dynamic optimisation of hydrogen combustion systems, wind-farm layouts and operation, and active road vehicle aerodynamics. Using Bayesian optimisation and Reinforcement Learning, an energy efficiency improvement of 10-30% can be achieved at zero cost using existing infrastructure/hardware and requiring only software modifications. The improved energy efficiency of the proposed approaches has been demonstrated through high-fidelity simulations and wind-tunnel experiments in the National Wind Tunnel Facilities.

By Nawid Keshtmand, Research Associate

The presentation focused on the importance of validating inventory estimates of greenhouse gas emissions using top-down estimates. This is important as there could be misreporting or gaps in the science which could lead to errors in inventory estimates.

The presentation then looked at the process of performing a top-down atmospheric estimate which consisted of various stages, with the most computationally expensive step being the use of the physics-based atmospheric dispersion model. The next part discussed the amount of data obtained using greenhouse gas measurement sites as well as satellite measurements. Due to the computational cost of the top-down emissions estimate pipeline, it is currently unable to process the large amount of data which comes in from satellite measurements. Therefore, to try to solve this issue, we look at reducing the computational cost of the top-down emissions estimate pipeline. We approach this by replacing the physics based atmospheric dispersion model with a machine learning based emulator.

The emulator we use is a graph neural network which consists of an encoder, processor and decoder. The inputs of the GNN is meteorological variables (such as

temperature, pressure, wind vectors, atmospheric boundary layer etc) and the output of the GNN is a footprint. We test the GNN on satellite data from Brazil. We train the neural network on 2 years of data (2014 and 2015) and test using data from 2016. We see that the footprints generated from the GNN are similar to the ground-truth footprints generated from the physics model and we are now using the emulated footprints to infer Brazil’s emissions and comparing it with the estimate obtained when using the ground-truth footprints. My research in particular is examining how we can use prototypes as an additional input in the GNN in order to improve the performance of the GNN. We can choose prototypes in various different ways such as using random prototypes, prototypes generated using k-means as well as prototypes which we hand pick using expert knowledge. We examine the effect of using the prototypes by comparing the performance of the model with the prototype against the situation where there is no prototypes added. We did this by having an oracle scenario where we choose a particular prototype which is closest to the footprint we want to predict (which would not be possible to do in practice). It was seen that by adding the prototype to the GNN, we can reduce the Mean-squared error obtained between the predicted and true footprints.

Key takeaways

We have developed a GNN architecture which is used as an emulator of the atmospheric dispersion model. We can generate footprints which we are using to
infer Brazil’s emissions. We are working to improve the performance of the model by
adding prototypes as an additional input to the GNN model.

Conclusion

Successfully able to use a GNN to emulate an atmospheric dispersion model.

By Dr Frank McQuade, Bloc Digital Visualisation Lead

This project has made strides in developing the visualisation application for Net Ecosystem Exchange (NEE) data. Understanding This summary highlights the project's achievements in building the application, its current capabilities, and its shift towards a user-friendly web interface.

The initial phase focused on creating a workflow for the application. This involved selecting an area of interest and building a Unity-based program. A crucial step was integrating this application with the digital twin data via a cloud-based API. The current application showcased the capabilities by connecting to both the Jules Emulator and Earth Observation models. Users can select crop type and climate model, allowing the application to generate NEE data specific to that region and choice. This empowers researchers and stakeholders to explore NEE data across various scenarios.

Beyond the application, the project established a robust cloud-based server. This server acts as the backbone, managing the application securely and facilitating communication with digital twin data. It also manages connections to various data sources, ensuring access to the latest information.

The project is taking a now transitioning towards a web interface. This next generation user interface offers several advantages including incorporating external data sources, including the successful integration with OpenStreetMap. This opens doors to richer visualisations by overlaying additional geospatial information.

The most significant advantage of the web interface is its user-friendliness. Unlike the current application, the web interface won't require remote installation, making it readily accessible to a wider user community, including stakeholders who may not have the technical expertise for software installations.

In conclusion, the visualisation project has established a powerful proof of concept for sharing NEE data. The current application demonstrates core functionalities and integrates with existing data sources. The cloud-based server provides a reliable foundation for future enhancements. Next steps in the project is the transition towards a user-friendly web interface, to allow broader engagement and further exploration of stakeholder needs. This web interface will be instrumental in creating a more robust and flexible implementation that will allow greater capabilities as the project moves forward.

By Dr Hibist Kassa, Policy Research Fellow

Inter-governmental Panel on Climate Change, Sixth Assessment Report highlights that GHG emissions have continued to increase, with unequal historical and ongoing contributions arising from unsustainable energy use, land use and land-use change, lifestyles and patterns of consumption and production across regions, between and within countries, and among individuals. The report recognises that ecosystem-based adaptation approaches such as urban greening, restoration of wetlands and upstream forest ecosystems have been effective in reducing flood risks and urban heat. Global GHG emissions in 2030 (implied by Nationally Determined Contributions in October 2021) make it likely that warming will exceed 1.5°C during the 21st century and make it harder to limit warming below 2°C. Hansen, Karecha and Sato (2024), conclude that the 12-month mean of global temperature is rising at 1.56°C relative to 1880-1920.

A combination of gaps between projected emissions from implemented policies and those from NDCs, that are already far shorter than those required to curb emissions, as well as, shortfalls in finance flows have undermined meeting climate goals across all sectors and regions. This lack of global leadership is further deepened by the Uturn in delaying or reversing UK commitments to reducing carbon emissions. This amplifies uncertainty in a policy context where there has been a shift from payments per hectare for landscape management, towards one that values farm and animal welfare practices as public goods. Since agriculture is a devolved area, the Climate Change Committee (2023) and National Union for Farmers have made calls for greater policy co-ordination at a UK level. Garvey and Jordan (2023) suggest that in seeking de-Europeanisation via Brexit, UK had become locked into disengagement with the EU and divergence between the devolved nations. Brexit in creating the opportunity of crafting new regulations for the environment, also required new institutions in a period of eroded capacity due to austerity.

This creates conditions of uncertainty for environmental regulation that has bearing on vulnerable ecosystems such as peatlands and animal welfare practices in ruminant production. These are the use-cases for the development of the Digital Twin in this project that will be used by policymakers, tenant farmers, land managers and other stakeholders. The aim is to simulate real world conditions to guide policymaking that encourage farm practices that reduce emissions intensity. While agriculture can play a role in reducing emissions, energy, transport and the built environment remain to be key emissions reductions sections. To this extent, there are concerns of challenges towards achieving absolute emissions reductions. Some of these are a result of the following:

• Carbon offsetting permitting fossil fuel industry, weapons manufacturers and airlines to continue to pollute, therefore not taking any measures to reduce emissions, Scottish Woodlands restricted approach to carbon trading
• Greenwashing techniques superficial changes to product packaging, minor increases in plant-based products, and continued encouraging of meat multibuy offers. Disclosures of emissions information has also been partial, with a focus on reduced emissions from stores and vehicles, when 95% of emissions comes from sales
• Continued overconsumption of meat and dairy, which has an effect on production that relies on feedstock imports that also places pressure on agriculture such as in the Amazon. Innovations in livestock rearing for low carbon beef introduced in the UK need to be examined in terms of impact on emissions and unintended consequences
• Globalised nature of livestock production emphasises the large carbon footprint of supermarkets in UK. UK Soy consumption impacts on Amazon
• Low carbon requirements added to trade agreements impacts less competitive farmers, for example, family farms, as well as importers to the UK (as NUF demands)
• Climate change impacts to farm practices: Calon Cymru Network oral evidence to UK Parliament cites how ‘changing rainfall patterns are undermining the basic assumptions of hill farming…causing more damage from winds and floods.’ For instance, knowing how tree planting is done to avoid emissions increase from environment

The Digital Twin is being developed on data sets from East Anglia. According to the UK Centre for Ecology and Hydrology (UKCEH), East Anglia fens peatland produces vegetables worth 3 billion pounds a year. While peatlands in their natural state capture CO2 through photosynthesis, once drained for agriculture purposes, it becomes a net source of CO2 emissions. Excessive rewetting can also increase methane emissions.

England’s Agricultural Transition Plan aims to ensure 40% of agricultural soil into sustainable management by 2028 and 60% by 2030. It also focuses on productivity and transition to low-carbon farming systems for arable and livestock. This emphasises the creation of new woodland, and its sustainable management, along with peatland restoration, agroforestry approaches on farms and planting of energy crops.

In the Agricultural Transition Plan, policymakers aim ‘to target the right level of ambition, combination and scale, in the right places and in joined-up ways to deliver target outcomes.’ England’s Peatland Action Plan sets out the government’s long term vision for managing, protecting, and restoring peatlands. This includes ending the horticultural use of peat in the amateur sector. It also includes recognition of how Greenhouse Gas Removal feasibility is impacted by decisions on alternative land use as well as technological options for production.

Key takeaways

Innovation and technological measures in agriculture and land use alone are not enough to reduce emissions. The Climate Change Committee recognises that robust policy framework is required for innovation and technology to be effective alongside behaviour change. The report continues to recognise that a UK food strategy that relies on innovation to decarbonise and increase productivity across food chains, without strong policies, puts emission reduction at risk. Identifying how productivity goals as well as emissions reductions targets can be met simultaneously or if there are trade-offs, what exactly those are need to be understood.

In the Making Landscape Decisions (2022) report, it was highlighted that land use decisions, especially restoration, must be place-specific and across all scales. There is a need to identify co-benefits and risks of competing land demands, such as food security, agricultural productivity, nature conservation and housing, as well as attendant social implications.

Conclusions

There is an uneasy compromise between restoring natural peat functioning and existing economic value and livelihoods in the case of intensive commercial agriculture such as in East Anglia. The long-term impacts of new measures for production of low carbon beef and dairy should be weighed against other environmental and health impacts. The Digital Twin aims to create an interactive platform that facilitates decision-making that is responsive to concerns such as this. However, there needs to be reflection on the limits or extent of trustworthiness given it aims to simulate real world conditions. Any shortfalls may unravel presumed emissions reductions from changes to farm practices and livestock rearing.