2.1. Ontologies

An ontology is a collection of classes, object properties and data properties expressing facts about and a semantic model of a domain of interest, for example, the built environment. An object property links an instance of a class (the domain of the property) to an instance of a class (the range of the property).Footnote ¹ Object properties may be structured hierarchically. A data property links an instance of a class (the domain of the property) to a data element (the range of the property).Footnote ² Similar to object properties, data properties may be structured hierarchically. A class acquires meaning via relationships to other classes.

An ontology is also known as a TBox (Terminological Component), where a class hierarchy is formed and associative relations are established through user-defined object properties, allowing cross-domain linking and modeling. Data properties allow the representation of different types of data. An ABox (Assertional Component) is an instantiation of a TBox representing entities, together with data and metadata about the entities. Inconsistencies can be determined at the level of data and relationships between the entities in an ABox via the assertion of data types, logical constraints, and cardinality (i.e., how many of something are allowed) in the TBox.

2.2. Linked data

Linked Data (Berners-Lee, Reference Berners-Lee2006; Bizer et al., Reference Bizer, Heath, Berners-Lee and Sheth2011) is Semantic Web data that is linked to other Semantic Web data to facilitate exploration and discovery (as opposed to semantic, but otherwise isolated data). Linked Data uses resource description framework (RDF) (Klyne and Carroll, Reference Klyne and Carroll2004) as the standard language to store data in the form of subject, predicate, and object triples. Related entities across datasets are connected via IRIs which are meant to be resolvable and provide meaningful information when looked up. Linked Data use ontologies to create and link entities and enrich the entities with data and metadata. Linked open data (LOD) refers to Linked Data that is published with an open license, such that it offers the possibility of finding and using related data, for example to resolve issues with naming and consistency.

2.3. Knowledge graphs

A knowledge graph is a network of data expressed as a directed graph, where the nodes of the graph are concepts or their instances (data items) and the edges of the graph are links between related concepts or instances. Knowledge graphs are often built using the principles of Linked Data. They provide a powerful means to host, query and traverse data, and to find and retrieve related information.

2.4. Data storage

Triple stores and quad stores provide solutions for the storage and management of RDF data. As the names suggest, triple stores use a 3-column (subject, predicate, and object) statement format, whereas quad stores use a 4-column (subject, predicate, object, and context) statement format. The context element can provide additional information about the data, for example a time coordinate, a reference or a name. When the context contains a name (which could be codified by an IRI) the statement is called a named graph. Many statements can belong to the same named graph, forming a subunit within the store and allowing the possibility of querying only the named graph. Most stores are able to accommodate both triples and quads and, for historic reasons, are often simply referred to as “triple stores.” A number of storage solutions exist, such as RDF4J (Eclipse Foundation, 2020) and Fuseki (Apache Software Foundation, 2020a) at the lab scale, and Blazegraph (2020) and Virtuoso (OpenLink Software, 2019) at the enterprise scale. In addition to triple stores, ontology-based data access (OBDA) enables ontological queries to be performed over relational and NoSQL databases (Botoeva et al., Reference Botoeva, Calvanese, Cogrel, Rezk, Xiao, Lenzerini and Peñaloza2016). Triple stores and OBDA provide endpoints, known as SPARQL endpoints, that are identified by IRIs and that can be used to query and update data. Authentication and authorization solutions to control security and access exist and are discussed elsewhere (Balfanz et al., Reference Balfanz, Czeski, Hodges, Jones, Jones, Kumar, Lindemann and Lundberg2019).

2.5. Queries and updates

SPARQL (Aranda et al., Reference Aranda, Corby, Das, Feigenbaum, Gearon, Glimm, Harris, Hawke, Herman, Humfrey, Michaelis, Ogbuji, Perry, Passant, Polleres, Prud’hommeaux, Seaborne and Williams2013) is a semantic-aware standard language to query and/or update data via a SPARQL endpoint, with or without activating an inference engine. Queries can be performed over named graphs, over an entire endpoint or (using a federated query) over multiple endpoints. Federated queries offer the possibility to retrieve data from and reason over multiple data sources at the same time, including the possibility of combining data from public and private data sources owned by any number of different parties. GeoSPARQL (OGC, 2012) is a standard defined by the Open Geospatial Consortium (OGC) as a domain-specific extension of SPARQL, allowing spatial reasoning and computations on data to reply to geospatial queries.

Each triple or quad store provides its own application programming interface (API) to support SPARQL query and update operations. However, each API is only compatible with its native store. Jena-JDBC (Apache Software Foundation, 2020b) provides a highly scalable store-agnostic API that can be used to query and update virtually any store.

2.6. Inference and rule engines

Inference engines (e.g., HermiT, Data and Knowledge Group, 2019) can operate on data expressed using RDF to reason about existing knowledge and infer new knowledge (Allemang and Hendler, Reference Allemang and Hendler2011), for example using the domain and range of object and data properties. Rule engines (e.g., Drools; Red Hat, Inc, 2020) can generate new knowledge through the codification and analysis of “if-then-else” conditions to calculate new values for data properties.

2.7. Examples

Google, Microsoft, IBM, Facebook, and eBay recently published a paper (Noy et al., Reference Noy, Gao, Jain, Narayanan, Patterson and Taylor2019) to explain how they use knowledge graphs. Google and Microsoft both use knowledge graphs to power their search engines. IBM uses one to support the IBM Watson question and answer engine. Facebook uses one to help understand what people care about. eBay is building one to connect what a seller is offering with what a buyer wants.

The article provides a number of insights and identifies opportunities and challenges, particularly relating to the scalability and query-response time of triple stores. Google built its knowledge graph using a set of low-level structures and associated reasoning mechanisms that are replicated at increasing levels of abstraction. This gives advantages in terms of scalability because new abstractions are built on tried and tested low-level units, and because it makes the design easier to understand. Facebook addresses the challenge of conflicting information by either dropping information if confidence in it was deemed to be too low, or by retaining the conflicting information and the associated provenance data. In this manner, their knowledge graph is able to explain why certain assertions are made, even if it cannot identify which assertion is objectively “correct.” IBM Watson uses polymorphic stores to offer the ability to support multiple indices, database structures, in-memory and graph stores. Data are split (often redundantly) into one or more of these stores, allowing each store to address specific requirements and workloads.

3.1. Structure

The World Avatar (Eibeck et al., Reference Eibeck, Lim and Kraft2019; CARES, 2020d) represents information in a dynamic knowledge graph using technologies from the Semantic Web stack. Unlike a traditional database, the World Avatar contains an ecosystem of autonomous computational agents that continuously update it.

The main components of the dynamic knowledge graph are illustrated in Figure 1. The dynamic knowledge graph contains concepts and instances that are linked to form a directed graph. The agents are described ontologically (Zhou et al., Reference Zhou, Eibeck, Lim, Krdzavac and Kraft2019) as part of the knowledge graph, and are able to perform actions on both concepts and instances. This confers versatility because it combines a design that enables agents to update and restructure the knowledge graph, with a structure that allows agents to discover and compose other agents simply by reading from and writing to the knowledge graph.

Figure 1. The main components of the World Avatar dynamic knowledge graph.

The links between the contents of the dynamic knowledge graph take the form of Internationalized Resource Identifiers (IRIs), which can be thought of as generalized web-addresses (Ishida, Reference Ishida2008). This ensures an unambiguous representation and enables the efficient traversal of the graph, both in the sense of traversing related content and in the sense of traversing whatever hardware is used to host the content.

3.2. Agents

At any given time, instances of the computational agents described in the knowledge graph may be activated to perform tasks. In this context, “agent” refers to software that uses Semantic Web technologies to operate on the knowledge graph to fulfill specific objectives. Agents may be atomic agents (i.e., not composed of other agents), or composite agents composed of multiple atomic agents operating in a coordinated way.

Agents can perform a variety of activities and can operate on both the concepts and instances in the knowledge graph (Eibeck et al., Reference Eibeck, Lim and Kraft2019):

• • Input/output. Agents can update instances with data from the real world, for example, data from sensors or smart infrastructure. Agents can retrieve data from the knowledge graph and send signals back to real world, for example, to visualize data or to control an actuator.

• • Update. Agents can query the knowledge graph, calculate derived quantities, and update instances with the new data. An example could be querying sources of emissions and the prevailing wind conditions to update estimates of air quality by simulating the dispersion of the emissions (Eibeck et al., Reference Eibeck, Lim and Kraft2019; Zhou et al., Reference Zhou, Eibeck, Lim, Krdzavac and Kraft2019).

• • Restructure. Agents can restructure the knowledge graph by:

  • – Adding instances (i.e., items in ABoxes), for example, to explore the optimal placement of new infrastructure (Devanand et al., Reference Devanand, Kraft and Karimi2019; Eibeck et al., Reference Eibeck, Chadzynski, Lim, Aditya, Ong, Devanand, Karmakar, Mosbach, Lau, Karimi, Foo and Kraft2020).

  • – Adding concepts (i.e., items in TBoxes) to improve the description of data, for example, to add classes relating to dispatchable and nondispatchable resources to an ontology describing energy sources.

  • – Adding relationships between instances or between concepts, for example, using ontology matching to support the creation of relationships between equivalent concepts.

• • Create new agents. Agents can provide services to facilitate agent discovery and to create composite agents to coordinate the activities of agents to perform complex tasks, for example, the estimation of air quality (Zhou et al., Reference Zhou, Eibeck, Lim, Krdzavac and Kraft2019). The possibility of introducing an agent marketplace and quality of service (QoS) measures for agents has been demonstrated using a blockchain-based smart contract system (Zhou et al., Reference Zhou, Lim and Kraft2020a).

The calculations performed by agents can take any form, including physics-based models with a theoretical structure, gray box models that combine some theoretical structure with data-driven components, and pure data-driven models. See Yu et al. (Reference Yu, Seslija, Brownbridge, Mosbach, Kraft, Parsi, Davis, Page and Bhave2020) for examples of these approaches in the context of emissions modeling. The availability of semantically structured machine-queryable data makes the World Avatar well-suited to be coupled to machine learning applications, where it is well-known that the time spent on data preparation and curation typically adds a significant cost to machine learning.

The potential of the World Avatar arises from the generic functionalities offered by agents that integrate ontologies and create new agents. Both would serve to raise the level of semantic interoperability and allow for the automation of processes to include new data and capabilities, and maintain interoperability as the quantity and complexity of the data and agents increases.

3.3. Ontological coverage

The current ontological coverage of the World Avatar includes 3D city data (OntoCityGML, Eibeck et al., Reference Eibeck, Lim and Kraft2019), weather (Weather Ontology, Institute of Computer Engineering at Technical University of Vienna, 2014), process engineering (OntoCAPE, Marquardt et al., Reference Marquardt, Morbach, Wiesner and Yang2010), eco-industrial parks (OntoEIP, Zhang et al., Reference Zhang, Romagnoli, Zhou and Kraft2017; Zhou et al., Reference Zhou, Pan, Sikorski, Garud, Aditya, Kleinelanghorst, Karimi and Kraft2017; Zhou et al., Reference Zhou, Zhang, Karimi and Kraft2018), and electrical power systems (OntoPowerSys, Devanand et al., Reference Devanand, Karmakar, Krdzavac, Rigo-Mariani, Foo, Karimi and Kraft2020). It includes ontologies for quantum chemistry (OntoCompChem, Krdzavac et al., Reference Krdzavac, Mosbach, Nurkowski, Buerger, Akroyd, Martin, Menon and Kraft2019), chemical species (OntoSpecies, Farazi et al., Reference Farazi, Krdzavac, Akroyd, Mosbach, Menon, Nurkowski and Kraft2020b), chemical kinetic reaction mechanisms (OntoKin, Farazi et al., Reference Farazi, Akroyd, Mosbach, Buerger, Nurkowski, Salamanca and Kraft2020a), and combustion experiments (OntoChemExp, Bai et al., Reference Bai, Geeson, Farazi, Mosbach, Akroyd, Bringley and Kraft2021). Selected subgraphs of the LOD Cloud (lod-cloud.net, 2020), in particular DBpedia (Lehmann et al., Reference Lehmann, Isele, Jakob, Jentzsch, Kontokostas, Mendes, Hellmann, Morsey, van Kleef, Auer and Bizer2015; DBpedia, 2020) are also connected to the knowledge graph.

3.4. Use cases

This section presents use cases that demonstrate different aspects of the World Avatar. The use cases are from the Cities Knowledge Graph, PIPS, and J-Park Simulator projects so are mainly associated with Singapore. However, the application of the World Avatar is not restricted to these contexts, as we shall discuss in Section 4.

3.4.1. Urban planning

The Cities Knowledge Graph project (CARES, 2020b) will integrate urban design and planning tools developed by the Future Cities Laboratory at the Singapore-ETH Centre with the World Avatar. The tools include the Collaborative Interactive Visualization and Analysis Laboratory design informatics platform (Future Cities Laboratory, 2020b), the Multi-Agent Transport Simulation Singapore model (Future Cities Laboratory, 2020c), the City Energy Analyst toolbox (Future Cities Laboratory, 2020a), and the ESRI CityEngine (Environmental Systems Research Institute, 2020) 3D modeling software. The concept is shown in Figure 2.

Figure 2. The Cities Knowledge Graph project will develop a pilot for a comprehensive knowledge management platform that provides interoperability between different types of city-relevant data to improve the precision of planning instruments and bridge the gap between planning use cases and knowledge domains. Reproduced from https://www.cares.cam.ac.uk/research/cities/ with permission.

The resulting system will enable interoperability between building data, transport flows, underground infrastructure, humidity, and temperature data to improve the design and planning of cities by automating aspects of data processing, and by integrating concepts and targets from different planning arenas. Work to date has demonstrated the capability to incorporate live building management system (BMS) data into the dynamic knowledge graph (CARES, 2020e) and has integrated the World Avatar with OntoCityGML (Centre Universitaire d’Informatique at University of Geneva, 2012), an ontology that describes 3D models of cities and landscapes based on the CityGML 2.0 standard (Gröger et al., Reference Gröger, Kolbe, Nagel and Häfele2012). Tools have been developed to convert data from CityGML to OntoCityGML format. Work to find the best way to host this data within the dynamic knowledge graph is ongoing.

3.4.2. Intelligent querying and generation of data

The World Avatar contains chemical data, including kinetic reaction mechanisms (Farazi et al., Reference Farazi, Akroyd, Mosbach, Buerger, Nurkowski, Salamanca and Kraft2020a) and computational chemistry (Krdzavac et al., Reference Krdzavac, Mosbach, Nurkowski, Buerger, Akroyd, Martin, Menon and Kraft2019). It is able to use computational chemistry data to calculate thermodynamic quantities required by the reaction mechanisms, and link the resulting quantities to the computational chemistry calculations used to derive them (Farazi et al., Reference Farazi, Krdzavac, Akroyd, Mosbach, Menon, Nurkowski and Kraft2020b). It is able to identify inconsistencies in the data (Farazi et al., Reference Farazi, Salamanca, Mosbach, Akroyd, Eibeck, Aditya, Chadzynski, Pan, Zhou, Zhang, Lim and Kraft2020c) and interact with high performance computing (HPC) facilities in the real world to perform additional computational chemistry calculations to generate the data required to resolve the problems (Mosbach et al., Reference Mosbach, Menon, Farazi, Krdzavac, Zhou, Akroyd and Kraft2020). It includes experimental data (for combustion experiments) and is able to automate the process of calibrating reaction mechanisms versus experiment data (Bai et al., Reference Bai, Geeson, Farazi, Mosbach, Akroyd, Bringley and Kraft2021).

Research in collaboration with CMCL (CMCL Innovations, 2020a) has investigated the use of natural language processing (NLP) to develop an intelligent query interface for the chemistry data available via the World Avatar (Zhou et al., Reference Zhou, Nurkowski, Mosbach, Akroyd and Kraft2020b). Figure 3 shows a screenshot of a demonstration website (Computational Modelling Cambridge Ltd., 2020). It has been shown that the use of intent recognition and topic modeling, and the structure of the domain ontologies has a strong impact on the accuracy of the queries, with shallow ontologies performing better than deep ontologies. This poses a number of open challenges because scientific research typically leads to hierarchical classification systems that do not lend themselves to shallow ontologies.

Figure 3. Screenshot of Marie website (with enlarged text).

The availability of intelligent query interfaces has the potential to improve the discovery and interoperability of data. Work is ongoing to use such interfaces in conjunction with the World Avatar to develop quantitative structure activity/property relationships (QSAR/QSPR) that allow the prediction of the properties of a chemical from its structure, for example, its toxicity (Organisation for Economic Co-operation and Development, 2020) or its photovoltaic performance (Eibeck et al., Reference Eibeck, Nurkowski, Menon, Bai, Zhou, Mosbach, Akroyd and Kraft2021, and references therein). This is the focus of the PIPS (CARES, 2020c) project. PIPS will develop the capability of the World Avatar not only to automate calculations (Mosbach et al., Reference Mosbach, Menon, Farazi, Krdzavac, Zhou, Akroyd and Kraft2020), but to automate experiments by allowing agents to control a laboratory robot. This capability will close the loop between analysis and experiment to automate the generation of reaction and synthesis data, and establish an ontology framework for AI reasoning agents to generate chemical knowledge from data.

3.4.3. Cross-domain calculation of air quality

In recent years there has been increasing concern over the effect of emissions from combustion on air quality (Department for Environment Food and Rural Affairs [United Kingdom], 2019; National Climate Change Secretariat [Singapore], 2020), in particular, the emission of NO _x (Department of Transport [United Kingdom], 2016, 2019; Air Quality Expert Group, 2017; Environmental Protection Agency [United States], 2020) and particulate matter (PM) (Environmental Committee, London Assembly, 2015), and the effect of black carbon (BC) emissions on the climate (Bond et al., Reference Bond, Doherty, Fahey, Forster, Berntsen, DeAngelo, Flanner, Ghan, Kärcher, Koch, Kinne, Kondo, Quinn, Sarofim, Schultz, Schulz, Venkataraman, Zhang, Zhang, Bellouin, Guttikunda, Hopke, Jacobson, Kaiser, Klimont, Lohmann, Schwarz, Shindell, Storelvmo, Warren and Zender2013).

Combustion accounts for approximately 99% of Singapore’s CO₂ emissions (National Environment Agency, 2018, pp. 134–137). In addition to being one of the most densely populated countries (The World Bank Group, 2018), it hosts one of the world’s busiest ports with arrivals exceeding 138,000 vesselsFootnote ³ and a throughput of more than 37 million twenty-foot equivalent units in 2019 (Maritime and Port Authority of Singapore, 2020). There naturally arises the question of how emissions from hard-to-abate sectors like shipping influence factors such as air quality in different regions of Singapore. The interoperability of the World Avatar provides a method to answer these questions. Figure 4 illustrates how this was achieved. See references (Eibeck et al., Reference Eibeck, Lim and Kraft2019; Farazi et al., Reference Farazi, Salamanca, Mosbach, Akroyd, Eibeck, Aditya, Chadzynski, Pan, Zhou, Zhang, Lim and Kraft2020c) for more details of the implementation and results.

Figure 4. Real-time cross-domain estimation of the contribution of emissions from shipping to air quality in Singapore. Agents operating on the World Avatar update the knowledge graph with real-time information about the weather and about ships in the vicinity of Singapore. An emissions agent is able to use the information about the ships to estimate the emissions of unburned hydrocarbons, CO, NO₂, NO _x , O₃, SO₂, PM_2.5, and PM₁₀ from each ship. An atmospheric dispersion agent is able to use the information about the weather, the emissions from each ship and the built environment in Singapore to simulate the dispersion of the emissions. Virtual sensor agents report the resulting air quality estimates at different locations. Adapted from Farazi et al. (Reference Farazi, Salamanca, Mosbach, Akroyd, Eibeck, Aditya, Chadzynski, Pan, Zhou, Zhang, Lim and Kraft2020c).

The choice of which agents to use in the calculation illustrated in Figure 4 is controlled by a composition agent (Zhou et al., Reference Zhou, Eibeck, Lim, Krdzavac and Kraft2019). The emissions can be calculated either by agents that use surrogate models developed using the SRM Engine Suite (CMCL Innovations, 2020c) and Model Development Suite (CMCL Innovations, 2020b), or by agents that directly use the SRM Engine Suite. The atmospheric dispersion can be simulated either by agents that use the Atmospheric Dispersion Modeling System (ADMS) (Cambridge Environmental Research Consultants, 2020) or by agents that use the EPISODE 3D CFD code (Karl et al., Reference Karl, Walker, Solberg and Ramacher2019). The agents wrap around existing software, including commercial software, with the possibility of an agent marketplace (Zhou et al., Reference Zhou, Lim and Kraft2020a) offering an alternative to traditional licensing models.

Whilst the emissions from ships can be described using data-driven models, trained using measurement data, for example, this would preclude the ability to assess the impact of different fuels. The chemical data (species properties, reactions, and reaction rates) that are required to assess the impact of different fuels can be described in the knowledge graph and any unknown quantities can, in principle, be calculated (although the calculation of reaction rates remains challenging). The ability of the World Avatar to enable interoperability between the different components of such calculations has been demonstrated in terms of calculating thermodynamic data for chemical species from computational chemistry data (Mosbach et al., Reference Mosbach, Menon, Farazi, Krdzavac, Zhou, Akroyd and Kraft2020), linking the resulting species data with reaction mechanisms (Farazi et al., Reference Farazi, Krdzavac, Akroyd, Mosbach, Menon, Nurkowski and Kraft2020b), and performing parameter estimation to improve the agreement with experimental data (Bai et al., Reference Bai, Geeson, Farazi, Mosbach, Akroyd, Bringley and Kraft2021). This is depicted on the right-hand side of Figure 4.

This use case illustrates how the World Avatar may be used to calculate the consequences of current activities in the base world. It demonstrates what can be achieved in terms of interoperability, both between models and data from different domains (weather, the built environment, ships, chemistry, emissions, and atmospheric dispersion) and covering different length scales (from the atomic length scales of the computational chemistry calculations to the kilometer length scales of the atmospheric dispersion simulations). Of course, this is unnecessarily elaborate if we are only interested in Singapore. The point, however, is that the World Avatar allows such calculations to be performed for any location.

3.4.4. Parallel worlds for scenario analysis

The parallel worlds capability of the World Avatar supports decision making in a complex environment by allowing the exploration of different scenarios and their associated outcomes. The parallel worlds exploit the structure of the dynamic knowledge graph to use scenario agents to group new instances of any entities that are modified in a scenario and store them in a scenario-specific part of the knowledge graph. Access, queries and updates to the scenario-specific instances are mediated via a scenario agent. The scenario-specific information in the parallel world can be thought of as overlaying the base world, so unchanged entities remain connected to the base world and any changes to the base world are reflected in the scenario. Conversely, any changes in the parallel world remain isolated within the scenario-specific part of the knowledge graph and so do not interfere with the base world. Based on an idea very similar to version-control systems that are widely used among software developers, parallel world containers store the differences to the base world in named graphs where the scenario provides the context. Technical details can be found in Eibeck et al. (Reference Eibeck, Chadzynski, Lim, Aditya, Ong, Devanand, Karmakar, Mosbach, Lau, Karimi, Foo and Kraft2020).

Figure 5 illustrates the parallel world concept. In this example, the parallel world is used to explore the effect of levying a carbon tax on the emissions from power generation processes to motivate a transition from fossil fuels to clean energy technologies. The following questions are addressed:

• • What value of carbon tax is required to make the transition to small modular nuclear reactors (SMRs) profitable for a given set of conditions (e.g., project life span, depreciation rate, electrical load profiles, and generator characteristics)?

• • Which plant(s) should be replaced and where should the new SMR(s) be located?

Figure 5. Parallel world concept for what-if scenario analysis. The panels at the top illustrate the electrical network from the real world (left) and an optimized network that is subject to a carbon tax (right). The network from the real world is described in the base world. The modifications to the network in the parallel world are described in a scenario-specific part of the knowledge graph. The pink triangles denote natural gas generators that are present in both the base world and the parallel world. The blue square denotes an oil generator that is only present in the base world. The radiation symbol denotes a small modular nuclear reactor that is only present in the parallel world. Adapted from Eibeck et al. (Reference Eibeck, Chadzynski, Lim, Aditya, Ong, Devanand, Karmakar, Mosbach, Lau, Karimi, Foo and Kraft2020).

The parallel world shows that oil generators are replaced with SMRs as the carbon tax is increased from S$5 to S$170 per tonne. The types of generator present and the corresponding estimates of the CO₂ emissions are automatically updated in the parallel world to reflect these changes, and an optimal power flow (OPF) agent is invoked to minimize the overall operating cost of the scenario in the parallel world. See references (Devanand et al., Reference Devanand, Kraft and Karimi2019; Eibeck et al., Reference Eibeck, Chadzynski, Lim, Aditya, Ong, Devanand, Karmakar, Mosbach, Lau, Karimi, Foo and Kraft2020) for more details of the implementation and results.

This use case illustrates how digital twinning and what-if scenario analysis using a dynamic knowledge graph can support decision makers to understand the effect of different design choices and policy instruments. In this example, the problem addressed by the parallel world is simple enough to be formulated as a classical optimization problem with a well-defined objective function. However, many scenarios will be too complex for this to be the case. How to address this is discussed in the context of the knowledge-graph-based digital twin of the UK in Section 4.

4.1. Base world

This section describes the progress to date and the immediate next steps for the creation of the base world in a knowledge-graph-based digital twin of the UK. The immediate motivation for doing this is to help address questions relating to how best to help the UK achieve net zero carbon emissions. The energy sector is the largest contributor to emissions in the UK and was responsible for 82% of total emissions in 2018 (Brown et al., Reference Brown, Cardenas, Choudrie, Jones, Karagianni, MacCarthy, Passant, Richmond, Smith, Thistlethwaite, Thomson, Turtle, Bradley, Broomfield, Buys, Clilverd, Gibbs, Gilhespy, Glendining, Gluckman, Henshall, Hobson, Lambert, Malcolm, Manning, Matthews, May, Milne, Misra, Misselbrook, Murrells, Pang, Pearson, Raoult, Richardson, Stewart, Walker, Watterson, Webb and Zhang2020, p. 40). Work to date has therefore focused on data relating to the electric power system, the gas grid, the potential for photovoltaic and wind power, biomass and the built environment, all of which are critical to the future of the energy landscape (Committee on Climate Change, 2018). Beyond carbon emission reduction, potential applications are of course numerous and include, but are not limited to, public health, urban and infrastructure planning, and achieving a circular economy.

Figure 6 shows the geospatial configuration of the National Grid Electricity Transmission and National Grid Gas Transmission systems (National Grid, 2020a, 2020b). Work to extend the ontological coverage of the World Avatar to include this data is ongoing. In addition to the configuration of the networks, data for all the regional generators over 30 MW in the UK (Department for Business, Energy, and Industrial Strategy, 2020) have been added to the World Avatar and used in a preliminary analysis of how a carbon tax on existing generation infrastructure would effect the composition of power generation (Atherton et al., Reference Atherton, Xie, Aditya, Zhou, Karmakar, Akroyd, Mosbach, Lim and Kraft2020). It was shown that the carbon tax motivated a transition from coal to combined cycle gas turbines (CCGT), with nuclear and renewable sources being fully utilized. When regional capacities, demands and (found to be complementary) transmission losses were considered, this transition was found to occur first in London and the South East, where northern coal imports were displaced by local CCGT generation. Figures 7 and 8 show the photovoltaic power potential and wind speed and direction data in the UK. The photovoltaic data are published by SolarGIS (2020) on behalf of the World Bank as part of the Global Solar Atlas (2020). The wind data are part of the ERA5 data set published by the EU Climate Data Service (Hersbach et al., Reference Hersbach, Bell, Berrisford, Biavati, Horányi, Muñoz Sabater, Nicolas, Peubey, Radu, Rozum, Schepers, Simmons, Soci, Dee and Thépaut2018). Work to extend the ontological coverage of the World Avatar to include this type of data is an immediate priority.

Figure 6. Geospatial configuration of the National Grid Electricity Transmission (red) and National Grid Gas Transmission (blue) systems. The markers (blue) show the locations of the intake terminals for the National Grid Gas Transmission system. The inset shows real-time data for the gas flow into the network from the Easington North Sea gas terminal. Data obtained from the National Grid (2020a, 2020b).

Figure 7. Map showing the average annual photovoltaic power potential in the UK from 1994 to 2018. The inset shows the seasonal variation in the average daily photovoltaic potential and temperature in the vicinity of Cambridge, UK. Contains data licensed by The World Bank under the Creative Commons Attribution license (CC BY 4.0) with the mandatory and binding addition presented in Global Solar Atlas terms (https://globalsolaratlas.info/support/terms-of-use).

Figure 8. Map showing wind data for the UK at selected points in time. The data show the horizontal component of the velocity 100 m above sea level. Left: November 19, 2020 12:00. Right: February 1, 2020 15:00.Contains modified Copernicus Climate Change Service information [2020]. The data are available under an open license from Copernicus Products.

Figures 9 and 10 show an example of land use data in the vicinity of Cambridge, UK and data about the built environment in the vicinity of the Department of Chemical Engineering and Biotechnology at the University of Cambridge. The land use data are available via the Crop Map of England (Rural Payments Agency, 2019), which provides geospatial data about the types of biomass grown throughout the whole of England. The data about the rivers and streams in Figure 9 are available using the ordnance survey (OS) Features API (Ordnance Survey, 2020a). The base map and building data in Figure 10 are available using the OS Vector Tile API (Ordnance Survey, 2020c). Every building with a postal address in the UK has an associated Unique Property Reference Number (UPRN), likewise every road has a Unique Street Reference Number (USRN). These identifiers and items linked to them are available using the OS Linked Identifiers API (Ordnance Survey, 2020b). Work to extend the World Avatar to include the biomass data is ongoing. Work to include information about the built environment using OntoCityGML (Centre Universitaire d’Informatique at University of Geneva, 2012) is underway as part of the Cities Knowledge Graph project (see Section 3.4.1).

Figure 9. Land use in the vicinity of Cambridge, UK. Contains data from the Crop Map of England (CROME) 2019 (Rural Payments Agency, 2019) licensed under an Open Government License.

Figure 10. Ordnance Survey building data in the vicinity of the Department of Chemical Engineering and Biotechnology (UPRN:10090969505) and the Centre for Digital Built Britain (UPRN:10090627569) in Cambridge, UK. Contains public sector information licensed under the Open Government License v3.0.

Figure 11 shows the built environment in the vicinity of Manchester Piccadilly Railway Station overlaid with data about the solar radiation incident on the roof of each building. The data are published as part of a virtual city map of Manchester (Virtual City Systems, 2020). This illustrates just one way in which the types of data in Figures 6–10 can be combined. Making such data available as part of a knowledge-graph-based digital twin of the UK presents the opportunity to include them in parallel world scenario analyses, for example, to investigate the building-by-building potential of installing solar panels and battery storage, either locally within each building, at a street or district level.

Figure 11. Modified screenshots of city data in the vicinity of Manchester Piccadilly Railway Station (Virtual City Systems, 2020). Contains public sector information licensed under the Open Government License v3.0.

4.2. Goal alignment for parallel worlds

The parallel worlds capability of the World Avatar offers a powerful tool to explore the design space of complex problems. However, the problems of interest, for example, how best to develop the UK infrastructure, rapidly become too complex to be treated in the manner of classical optimization problems. This is the case both in the sense that the problems are not amenable to defining a straightforward objective function and in the sense that the design space is too complex to be explored without some additional insight to suggest propitious scenarios. In an ideal world, computational agents would suggest scenarios and automate the exploration of the design space. In order to attempt this, it will be necessary to equip the agents with an understanding of our values and what we, as a society, want to achieve. In other words, we need to equip the agents with goals.

There are many ways to define goals. The sustainable development goals (SDGs) (United Nations, 2015) defined by the UN offer one possible starting point. There are 17 SDGs, which are further specified in terms of 169 targets, each of which is associated with a set of proposed indicators. Each level in this hierarchy is increasingly specific, so whilst the goals are abstract, the indicators are specific and measurable. Table 1 shows an example for SDG 9.

Table 1. Sustainable Development Goal 9, Target 9.4 and Indicator 9.4.1.

A sustainable development goals interface ontology (SDGIO) already exists (Sustainable Development Goals Interface Ontology [SDGIO], 2015). Work is underway to link the World Avatar to this ontology and to create agents to evaluate selected SDG Indicators. The current focus is on Indicator 9.4.1. Data for all the UK power stations (and in fact power stations worldwide) are available via the World Avatar, and agents are being developed to estimate the emissions intensity of power generation in the UK and worldwide. Such agents can be written by placing thin, standardized wrappers around conventional pieces of software, irrespective of whether they are in-house, third-party, or expensive-to-evaluate (see, e.g., Mosbach et al., Reference Mosbach, Menon, Farazi, Krdzavac, Zhou, Akroyd and Kraft2020). In the first instance, the emissions intensity of power generation will be calculated as CO₂ emissions per unit electrical power produced, as opposed to per unit value added. The rationale for this is that the power produced is a measurable physical quantity, whereas value added is less straightforward to define. The ability to evaluate these indicators and learn how actions investigated in parallel worlds translate to progress toward things like the SDGs provides a possible route for how to develop the insight required to suggest new scenarios.

Another possible route forward could be to calculate goals based on environmental, social, and governance (ESG) metrics, either for individual companies or for specific industrial sectors. Such metrics are projected to be incorporated into future financial regulations and will be strongly linked to the risk and growth profiles of the companies involved. The use of the dynamic knowledge graph could help address the issues of ambiguity and consistency that currently affects such metrics because all metrics would be calculated on the basis of the same data set. For example, environmental factors such as emissions, waste management, water and energy usage could be assessed based on real-world data from the knowledge graph. Further, the parallel worlds capability would allow industries to explore the impact of different strategies, for example investment in different energy technologies, and use this to guide their corporate social responsibility (CSR) strategy.

Inevitably, there will arise scenarios where goals conflict with each other (Fuso Nerini et al., Reference Fuso Nerini, Tomei, To, Bisaga, Parikh, Black, Borrion, Spataru, Castán Broto, Anandarajah, Milligan and Mulugetta2018). For example, building more roads may enable better access to schools and therefore better education, so lifting people out of poverty and reducing hunger. However, this would also lead to more road use which could (depending on the type of transport) lead to more pollution. It is therefore necessary to consider goal alignment, that is to say, how to ensure that the scenarios considered by the World Avatar support the goals of humanity. This could potentially be achieved by using the goals not only to inform, but also constrain the suggestion of scenarios. The questions of how to suggest scenarios and achieve goal alignment remain important topics of research.