Information Circularity Assistance based on extreme data

2.1 Key questions and indicators for decision-making

A wide range of circularity indicators are discussed in the literature [16], [17] and examined for their informative value in individual case studies [18], from enterprise down to product-related indicators referred to as nano-level [17]. Key questions to be answered include:

1. What proportion of recycled material should the production system be designed for to be able to realize defined product characteristics?

2. How shall product features be designed to withstand the load spectrum with the available material mix?

3. How can it be ensured that a product can still be operated in an eco-efficient manner in 20 years?

A German study on the reuse of electric/electronic components in washing machines indicates that a differentiated analysis of indicators is necessary for a meaningful sustainability assessment [19]. Furthermore, the approaches of Dynamic Material Flow Analysis (dMFA) and Life Cycle Assessment (LCA) can be combined (cf. approach by [20]): For a given material mix, it is necessary to determine how existing processes and equipment must be calibrated to be able to manufacture? What bandwidth must be provided for in the production system? Which process parameters must be adjustable, and in what range? One challenge in this regard is the unpredictable product quality of recycled products [21]. This illustrates the challenges in the very early phases of product creation concerning the product and production system: it has been demonstrated that sufficiently precise forecasting models can only be calculated for a few years [22] or are dependent on a number of assumptions, including the frequency of reuse [19].

A systematic literature analysis based on PRISMA [15] led to 37 results (Scopus 18, Web of Science 19), which were reduced to 7 articles that deal more specifically with the use of quantitative or qualitative data. There are no approaches or statements on how to combine quantitative data to make qualitative long-term statements. An additional analysis of the literature confirms that few approaches refer to support in the early phases based on business models or product concepts but then provide purely qualitative statements. Reviews found in this analysis support that quantitative circularity metrics are currently primarily used in the late phases of development [23] and are inadequate for fully representing the complexity of the Circular Economy [24]. One example is the Circularity Evaluation Tool (CCET) [25]: The circularity potential is evaluated subjectively based on parameters that in turn have been abstracted from design guidelines [26]. The assessment is based on product types and does not take into account any available data on circularity indicators. The analysis of circular trends is initiated as the first phase in the Strategic Planning Decision Framework oriented to Circular Business Models (SPDF-CBM) [27], yet it is not subject to a systematic elaboration. Models for evaluation, which are refined through qualitative statements, reduce the forecasting errors, but assume valid data and only allow statements about events in the near future [28], [29]. Similarly, only positive scenarios are considered. The quality of recyclate is not considered [30], [31]. Based on the conducted literature review, no approach combines available data for quantifiable short-term statements with qualitative assessment schemes for long-term forecasts.

2.2 Scenario-Technique for strategic product planning

The Scenario-Technique serves as a support to systematically use trends in the multi-criteria evaluation of strategic options. Decisions in favor of Circular Economy business models require such considerations with continuous updates requiring agile Scenario-Technique approaches [11]. Algorithms like cross-impact analysis [32], consistency assessment [33], and Intuitive Logics (e.g. [34]) are applied, only partially including quantitative probabilities. Along consistency-based Scenario-Technique, artifacts are generated and correlated:

1. Influencing factors: Identification and selection of relevant factors, e. g., in relation to product type, application domain/market, technologies, and material circularity

2. Influence matrix with the active and passive sum of interactions [35] (e. g., in the system grid [33], possibly with time reference/dynamics [36], or impulsivity [11])

3. Key factors as a subset of all influencing factors (e. g. [37])

4. Projections of “multiple futures” [11], [38]: Access to expert knowledge and distributed data sources from statistical databases with an evaluation of the respective data quality

5. Consistency assessment with numerical scales (see, e. g. [39]) and consistency matrix

6. Raw scenarios, selected according to diversity, stability, and relevance [36], [40]: technical evaluation of projections, which must be explainable

7. Scenarios with consequences that are generated in a comprehensible and credible manner with uncertainty to support decisions in terms of lead strategy and alternative strategies [11].

Algorithms partially supported by Scenario-Technique tools and data-based concepts like the Integrated Scenario-Data Model (ISDM) presented in [12] are used to generate these artifacts, for example, to evaluate indirect influences, to include influence strengths or to take input-output relationships into account [40]. Branch-and-bound [41] or evolutionary algorithms are used in scenario building [42]. Fuzzy approaches exist for scenario building as well as for consistency and influence assessment (like [43]). A considerable effort is caused by consistency assessment [39], [44], [45]. It depends to a large extent on personal experience in the application and is critical with regard to the quality of the scenario results [45], [46]. Here, initial studies show potential in the application of Data Science and Artificial Intelligence in consistency-based scenario technology [47] and in scenario-based foresight in general [48]. Examples are described both in the form of data-based approaches with neural networks and through integration into a higher-level knowledge base [39]. Web mining and text mining are also used in the preparatory steps of the Scenario-Technique to identify and aggregate relevant topics and derive influencing factors and projections [49]. All selection rules used are heuristic, require expert knowledge to use, and have a critical effect on the quality of the resulting scenarios. Within the individual process models, there are breaks between the artifacts [46], [50]. The resulting loss of information makes the interdependencies difficult for the user to understand [46], [50], [51]. Only experts can estimate causalities based on implicit knowledge. In terms of procedure and artifacts (see list above), the Scenario-Technique offers a basis for generating circularity-related long-term statements through forecasts of the consequences of strategic decisions. However, to date, there is no approach for deriving projections from data-based short-term statements and relating these to qualitatively described long-term forecasts in terms of influence and consistency.

2.3 Digital Twin-based analysis of extreme data

In order to derive these projections and relate them to long-term statements, it is necessary to map extreme data that occurs during the product life cycle and to create the corresponding data layers. Digital Twins are often used to map data for an asset (product, production system, etc.) with the aim of providing a digital representation of the asset [52], [53]. The basis of Digital Twins are models and their relationships. In addition to static models, which must be synchronized with reality, dynamic processes of the product must also be captured through active data acquisition and executable models to enrich the static image with dynamic behavior [54]. In addition to these aspects, three further properties must be included according to [54], [55]:

1. The models and their relationships are synchronized with the asset so that the static image of the asset is always replicated.

2. There must be active data acquisition from the asset to the Digital Twin so that the dynamic processes are also captured.

3. The existence of an executable model (i. e., a simulation model) is necessary to expand the static image with the replication of the dynamic behavior of the asset.

The Digital Twin has already been investigated and further developed in preliminary works as a basis for the application of Data Science and Artificial Intelligence [56], [57]. A Digital Twin, as described above, has a data acquisition interface that is responsible for acquiring the data from often heterogeneous sources. A workflow for the acquisition of data from heterogeneous data sources was proposed in [58] and is illustrated in Figure 1. First, a communication must be set up as the basis for data collection. Then, the data must be integrated into an appropriate software environment to perform the following steps. The data is validated, outliers can be detected and the dimensionality of the data can be reduced, depending on the application and the amount of data. Finally, the data is semantically annotated to ensure consistency and the ability to align data from heterogeneous data sources.

Figure 1:

Exemplary workflow for the data acquisition from heterogeneous data sources, following [58].

However, the existing definitions and applications of the Digital Twin mainly target at the production system or the product during design or production [59], [60]. Often, only one phase of the product life cycle is considered. However, a lot of data is generated for a product during its life cycle, e.g., during manufacturing, usage, condition monitoring, or recycling. The Digital Twin shall allow access to data throughout the life cycle, as described in the WiGeP position paper for the Digital Twin [61]. To enable long-term statements, the Digital Twin of a product must be able to represent this product over its whole life cycle. According to [62], a reliable Digital Twin must evolve with the changing characteristics of reality. These changing characteristics are often unknown when the Digital Twin is initially set up and can also include technological leaps (for example, concerning materials, material-process combinations, automated treatment, condition testing for recycled materials, and simulation models). A literature review [63] concluded that the integration of heterogeneous data structures in Digital Twins will be both a great opportunity as well as a major challenge in the future.

2.4 Identification of relevant data science and AI approaches

As previously stated, a variety of data is generated throughout the life cycle of a product, resulting in extreme data. To capture this data and information and gain knowledge from it, different Data Science and Artificial Intelligence methods exist, which may be used. The objective of this section is to describe potential technologies that may be applied in conjunction with Digital Twins to derive meaningful insights from the data for the Scenario-Technique. This chapter aims to provide an overview of technologies and not a comprehensive list of approaches. Further possibilities for enhancing the Digital Twin with Machine Learning algorithms are discussed in several publications, such as [64].

Knowledge Graph components will be used with extreme data and the Digital Twin. To derive insights from the Knowledge Graph, Graph Neural Networks or graph embedding algorithms for data analysis are used according to [65], [66]. By incorporating a Knowledge Graph, relationships within extreme data can be mapped, leading to Linked Data. However, querying a Knowledge Graph becomes complex for big graphs. Therefore, queries based on Natural Language Processing (NLP) can help in reducing the complexity. Large Language Models (LLMs) such as GPT-4 [67] or BERT [68] can understand natural language queries. A domain-specific customization is required to further translate this natural language into precise queries to the Knowledge Graph. This enables efficient exploration and querying. In [69], the application of LLMs for the automation domain and the interaction of these models with graphs and ontologies have already been investigated. This related work shall serve as a basis to enable natural language queries by the users to improve interaction for the extreme Scenario-Technique.

Table 1 gives an overview of the mentioned AI approaches, which are seen to be suitable for using them within the Information Circularity Assistance to allow a consistent information view across the life cycle of a product and allow the Scenario-Technique to use the information.

A Knowledge Graph is constituted by a knowledge base, which can be realized by an ontology [71]. A number of studies have concentrated on the construction of appropriate ontologies for the product life cycle. These offer the potential for the sharing and reuse of collected information in a uniform semantic description, integrating complex expert knowledge, and applying automatic reasoning [72]. In [73], the ontology is divided into several key entities for this purpose, including product, process, material, and properties. For example, the recycling of materials in products can also be modeled by this approach. In [74], an ontology was created with a focus on Additive Manufacturing. Its objective is to cover the entire product life cycle, thereby providing the opportunity to support the design of new products using the available data. In contrast [75], presents an ontology-based End-of-Life model that considers the possible reuse of individual product components. This could, for example, result in the products being offered again as refurbished ones. Table 2 provides a summary of the discussed ontologies. Starting from those ontologies, which already exist for different areas of product circularity, a new ICA base ontology shall be created.

Ontology learning approaches shall be used to further enhance these ontologies with the occurrence of extreme data, which is accessible in unstructured or semi-structured formats. As stated in [76], a variety of ontology learning approaches may be employed depending on the type of data to learn the correlations for the ontology. This may entail either the creation of a new ontology from scratch or the refinement of an existing one. For instance, linguistic approaches that employ natural language processing [77], [78], [79], [80] are frequently utilized for unstructured data to facilitate pre-processing, term/concept extraction, or relation extraction. Statistical methods (for example, contrastive analysis or hierarchical clustering) [81], [82] are frequently employed for semi-structured data to extract terms, concepts, and relations within the data set. Inductive logic programming [83], [84] can be utilized as a third learning method for axiom formation. To integrate the extreme data over the product life cycle, existing ontologies need to be refined during the life cycle, since not all properties and relationships can be modeled at the beginning. Therefore, linguistic and statistical ontology learning methods shall adapt the ontologies according to the occurring extreme data. The introduced ontology learning approaches and the concrete examples are grouped and listed in Table 3.

In general, the impact of circularity (for example, quantities or quality of recyclate) must be addressed in order to enable the implementation of hybrid decision support. To employ Data Science and AI technologies, it is essential to undertake a combined modeling of the data. The data pertaining to Digital Twins and Knowledge Graphs must be integrated and semantically enhanced through the utilization of ontologies. Existing ontologies for specific phases of the product life cycle can be aligned for this purpose and subsequently refined and augmented through ontology learning processes. This design of the extreme data serves to reinforce the Scenario-Technique.

The Digital Twin is founded upon a comprehensive set of data that encompasses the entirety of the product life cycle. This data set includes a variety of indicators that facilitate reuse, repurposing, repair, refurbishment, remanufacturing, and recycling. It may comprise both real data and synthetic data pertaining to the product’s life cycle. The synthetic data cloud is manually generated based on future scenarios that predict based on the analysis and assumption of historical data. As previously discussed, a comprehensive relevance analysis and prediction synthesis is essential, using Data Science and Artificial Intelligence-based algorithms.

Concurrently, the Scenario-Technique is to be developed into an advanced methodology, designated the Extreme Scenario Technique. This evolution incorporates insights derived from the Digital Twin analysis and projection synthesis, which is necessary to ensure the effective delivery of pertinent information. The Extreme Scenario Technique is specifically designed to address decision scenarios related to circular products and production systems. Following the Data-Information-Knowledge-Wisdom (DIKW) pyramid, the wealth of extreme data within the product life cycle is extended to facilitate actions in product planning and requirements engineering for future products, in line with the principles of the V-model.

3.1 Criteria for matching Scenario-Technique demands and Digital Twin offerings

As previously stated, the core of the Scenario Technique entail a procedure to systematically derive scenarios from influencing factors. This method requires the availability of extensive data and expert knowledge, as outlined in Section 2.1. Descriptors are derived from influence factors and projections are identified for each descriptor. The resulting scenarios are the product of bundling projections [11].Throughout the process, analysis and synthesis are performed. For instance, influence factors are checked for interdependencies, and projections are analyzed for consistency. The ISDM [12] was designed to provide comprehensive support for all steps of the process, encompassing both the establishment of connections to pertinent data sources and the management of data within the context of a Scenario-Technique project. These intermediate steps represent the demands, that the Scenario-Technique places on the Digital Twin. For information circularity, the complexity of data demands to perform the Scenario-Technique increases. These demands can be viewed as the interface between the Scenario-Technique and the Digital Twin, as illustrated in Figure 2. The Digital Twin presents further opportunities for enhancement, which shall enrich the Scenario-Technique. Building upon this foundation, five use cases for data science and Artificial Intelligence have been identified, which refer to different Scenario-Technique artifacts identified in Section 2.2. These use cases comprise the columns in Figure 3:

1. Inference of relevant influence factors: The identification of influence factors requires considerable effort to describe future influences on a product in Circular Economy. In the original Scenario-Technique approach, both elaboration and checking for relevance require experts to contribute their knowledge, typically in workshop settings. Previous approaches that aim to reduce efforts subsume the reuse of influence factors from previous projects or generic sets that are relevant across application domains. Even with these approaches, the analysis of relevance in a specific Scenario-Technique project remains a challenge. Linked Data and domain ontologies in combination with a semantic representation of Information Quality seem promising approaches to support such inference.

2. Correlation-based prioritization: The selection of influence factors, or descriptors, constitutes a significant foundation for the resulting scenarios. The prioritization needs to ensure that the data processed throughout Scenario-Technique steps is manageable and that the resulting scenarios are comprehensive for decision makers. At the same time, these scenarios must be representative of the entire domain of interest (like product creation and CE, combined with technologies, industries, markets, etc.). Only indications are possible anyhow due to the future-oriented perspective so correlation-based approaches in AI are applicable.

3. Projection synthesis: Per descriptor, projections need to be described indicating how this descriptor will evolve in the future. As previously stated in Section 2.1, for CE, a combination of qualitative (e.g., “political obstacles on global trade will {increase, remain, decrease}”) and quantitative projections (e.g., “recyclate material will be available in {2, 5, 10} years”) should be employed wherever feasible. he formulation of such projections necessitates a profound understanding of the pertinent domains, which may vary across Scenario-Technique projects. The synthesis of projections can be achieved through the utilization of databases such as Statista, trend radars accessible in natural language, and simulations or simulation results.

4. Reasoning on consistency: Assuming projections are available, consistency analysis of such projections requires an interpretation of the semantic interdependencies across all descriptors. For example, recycled material may be technically available in the near term, but economically affected by the evolution of global trade. Consistency analysis must therefore be based on a multi-domain understanding of all influence dimensions. For many domains, such as materials, technologies, or markets, domain ontologies are available as a promising basis also for consistency reasoning in the Scenario-Technique.

5. Inference of consequences: When bundling projections into scenarios, a final consideration is how useful the resulting scenarios are for decision making. A typical approach is to aim for extreme scenarios, such as “best” and “worst” scenarios, to cover the full range of possible futures. This is a simplification compared to the actual objective of this step: The creation of actionable scenarios, i.e., scenarios that support decisions for or against solution alternatives. Therefore, the understanding the consequences that may be implied by certain projections and their combinations should be considered. Inferences based on correlations in the sense of “what-if” tasks can help in the final selection of scenarios (i.e., projection bundles).

Figure 3:

Steps from the Scenario-Technique and data categories in Digital Twins spanning use cases for extreme data which could be interconnected in the Information Circularity Assistance.

The life cycle of a product generates a huge amount of (extreme) data from different phases, e.g. planning and engineering, realization including production, operation/use, or service (especially maintenance, repair, and overhaul). In the literature there are different groups of phases for the product life cycle, e.g. grouping according to the beginning, middle, and end of life [61] or dividing into design, development, manufacturing, operation, and disposal [85]. The Industrial Digital Twin Association and asset administration shell also discuss the possibility of incorporating product life cycle data. However, there is currently no clear definition or specification.

All of the proposed groupings lead to different categories of data that must be provided by the Digital Twin to achieve proper information circularity. This paper is based on the gPLC model [1], with phases of planning, engineering, realization including production, operation and service delivery, as well as material and information circularity. Despite all possible variants, it is necessary to provide a categorization of the data, including an exemplary description, in order to subsequently identify relevant data science and Artificial Intelligence methods. The categories are named and examples for these data types are given. This is to provide an exemplary view of the expected extreme data that needs to be captured, managed and processed within a Digital Twin, that interacts with the previously described Scenario-Technique. These data categories form the rows in Figure 3:

1. Planning Data of various types, such as data related to project timelines, resource allocation, and production schedules from the integration with planning software: This help optimize plans for sustainability and compliance. Digital Twins can also maintain a digital record of quality assurance processes, compliance documentation, and certifications earned by the product or its components. Such a centralized repository ensures traceability and facilitates audits or regulatory reporting. In addition, specifications for the new product, including sustainability requirements (e.g. percentage or recycled materials) can be used. This includes documents such as the system and component specification or recycling rate.

2. Engineering Data: The Digital Twin can be integrated with various product development tools such as computer-aided design (CAD) software, product life cycle management (PLM) systems, and simulation tools. By connecting to these systems, it can access detailed engineering data such as design specifications, prototypes, simulation results, and design changes throughout the product development process. These models are part of the digital asset management component of the Digital Twin. For example, simulation models can provide insight into the environmental impact of different design choices, manufacturing processes, and supply chain decisions. Digital Twins can also perform virtual life cycle assessments to evaluate the environmental impact of the product from raw material extraction to end-of-life disposal. By integrating LCA methodologies into the Digital Twin framework, factors such as carbon footprint, water use, and ecological footprint can be quantified.

3. Realization/Production Data: of such a system. Examples include supply chain data, manufacturing process data, or quality metrics. Digital Twins can provide real-time visibility into the entire supply chain network, including suppliers, manufacturers, distributors, and retailers. By integrating with enterprise resource planning (ERP) systems, Digital Twins can access data about supplier performance, inventory levels, lead times, and production schedules. This visibility enables organizations to identify bottlenecks, optimize inventory levels, and mitigate supply chain risks. Digital Twins can also help to ensure regulatory compliance throughout the production life cycle by maintaining comprehensive records of production processes, materials used, and environmental impact. By integrating with regulatory databases and compliance management systems, Digital Twins can automate compliance monitoring, generate audit reports, and ensure adherence to regulatory requirements.

4. Operation Data: Gathering data from the operational phase is often difficult, resulting in sparse and therefore extreme data. Digital Twins can monitor product usage in real-time by integrating with embedded sensors and IoT devices. These sensors collect data on operational status, performance metrics, usage patterns, and environmental conditions during use. By analyzing this data, Digital Twins provide insights into how the product is being used, including usage patterns, frequency of use, and operational efficiency. In addition, Digital Twins can capture feedback and customer reviews throughout the product life cycle, including user interactions, service requests, and product reviews. By aggregating and analyzing this feedback, organizations gain valuable insights into customer preferences, pain points, and satisfaction levels. As a result, they can improve product usability, address customer concerns, and increas overall customer satisfaction.

5. Circularity Data: Additional data occurs at the end of a product’s life when the product or parts of it may be reused, remanufactured, refurbished, etc. Circularity data can include life cycle data by collecting data on the environmental impact of the product throughout its life cycle. A Digital Twin can facilitate the final life cycle assessment. This includes data on energy consumption, greenhouse gas emissions, water use, and waste generation. By analyzing this data, organizations can identify opportunities to improve the product’s environmental performance, reduce resource consumption, and minimize waste generation through Circular Economy practices.

In conclusion, a huge amount of (extreme) data is generated during the product life cycle. Examples of occurring data across identified data groups are described, which will be used in the future to enrich the Scenario-Technique. Based on this overview and a concrete product, it is possible to identify opportunities that can already be realized based on existing data and a Digital Twin. However, it also clearly shows the need for future Digital Twin designs to handle this information circularity and not just cover single data groups. Figure 3 gives an overview of the aspects discussed in this chapter with the described data groups that are stored and modeled in the Digital Twin, the occurring extreme data, and the steps of the adapted new Scenario-Technique. The appropriate Data Science and Artificial Intelligence technologies serve as a kind of interface between the fields, which will be discussed in more detail in the following chapter.

As visualized, both fields (Scenario-Technique and the Digital Twin) need to be linked. Inputs from the Digital Twin improve the Scenario-Technique and then create new “data demands” for the Digital Twin, so there is a loop of adjustment between them. From the other side, the Scenario-Technique, with its adapted steps for the information circularity, also creates new possibilities for the Digital Twin within the Circular Economy. Appropriate data science and Artificial Intelligence technologies act as an interface or as a step to enrich the information transferred between the two fields.

3.2 Interconnecting the Digital Twin with extreme data

This section discusses the identified criteria for matching Scenario-Technique demands and Digital Twin offerings from Section 3.1 by introducing suitable Data Science and Artificial Intelligence technologies from the literature (Section 2.4) to enable the interconnection of the two fields (Scenario-Technique and Digital Twins), as visualized in Figure 3. Depending on the identified use cases within the Scenario-Technique (1–5 from Section 3.1) and the data category (A to B from Section 3.1), different Data Science and Artificial Intelligence technologies are suitable. In addition, different technologies may be suitable for a specific use case across different data groups or vice versa for a specific data group across all use cases. This will be discussed below.

Extensive expert knowledge is required to derive influence factors. Therefore, approaches for the reuse of influence factors are developed. For further automation and enrichment, the information from previous projects or the generic sets across application domains should be modeled in a computer-readable format. Linked Data in general (especially graph-based approaches as discussed above) as well as domain ontologies are approaches to formalize the influence factors and to allow the modelling of complex interactions between them. The prioritization of the influence factors is crucial for the subsequent scenario generation. For this step, correlation-based approaches including mathematical measures are a suitable choice. These approaches can be used to identify groups of influence factors correlations and to remove potentially redundant or uncorrelated factors. Experts can be involved in this step, and the correlation-based approaches can support them by enriching the previously built semantic representation. Different types of data may be suitable for projection synthesis. Retrieving relevant data from databases can be one step, where natural language processing can help to find matching database entries for selected descriptors. Trend radars may also be analyzed by natural language processing, or by using LLMs for a semantically enriched understanding or the reports and conversion into an appropriate format to match the information to the descriptors. Another source can be simulation models, where machine learning models can be used to increase the accuracy of the simulation model by adding additional parameters. In addition, machine learning can be used to build surrogate models to reduce processing time, allowing for faster evaluation of the indicators. During the consistency analysis, the semantic dependencies between the descriptors must be interpreted. Available domain ontologies with appropriate reasoning algorithms are promising approaches. As a final step in the Scenario-Technique, consequences must be derived. Based on the (consistent) projection and other defined effects, concrete scenarios are derived to support decision making. Machine Learning models such as decision trees can be a helpful technology to support the final selection of scenarios here.

On the other hand, different data is generated throughout the product life cycle. The data changes during the life cycle, leading to different approaches to analyze them. During the “beginning of life” (BoL), in our grouping the planning and engineering phase, the data often comes in the form of linked data/information. Different materials with different properties are linked to a final product with certain properties. This can be modeled within an information model and enriched with semantics, so semantic approaches may be more appropriate in these phases. The “middle of life” (MoL), in our case the data from production and during operation, often generates numerical data in the form of vectors (e.g. data from the production process, production parameters, environmental data during operation). This type of data is suitable for analysis using machine learning techniques, which are designed to analyze numerical data of different types and formats (e.g. image data, time-series data, text data, and parameter values) [86]. At the “end of life” (EoL), circularity data is generated. This data is a mixture of the two groups described above. Products and their composition may generate information that needs to be considered as linked data, leading to semantic approaches such as ontologies. In addition, tests can be performed on the products that generate data, that can be analyzed by machine learning models (for example, optical inspection which is analyzed by neural networks to determine the status of the product for reuse).

This evaluation leads to an intersection of approaches, as shown in Table 4. The rows are the retrieved data groups and the columns are the steps of the Scenario-Technique. Combining suitable Data Science and Artificial Intelligence technologies for the different groups, the final table is generated. Within the Machine Learning approaches, Knowledge Graph embeddings and Graph Neural Networks as discussed previously are also included.

Based on this initial evaluation and assignment of approaches, some specific cells of the table will be discussed here. In addition, the application of technologies to specific scenarios will be part of the following chapter.

1. Prioritization & planning data: Within the prioritization step, planning data can be used with Linked Data and ontology techniques. More specifically, the built-in semantics allow, for example, relevance rating or basic graph analytics, such as centrality measures, to provide meaningful insights for the prioritization step. In addition, data availability can be checked against the ontology to identify missing data. This information can either be used to collect that data or to reduce the priority.

2. Projection synthesis & production data: Projection synthesis requires data access and calculations, for example to assess data quality and make qualitative statements. This can be done by machine learning models based on production data that can, for example, provide probabilities for product performance metrics depending on the materials used and their quality (for example, percentage of recycled materials used). Large Language Models can help to create qualitative statements from the set of predictions generated by a machine learning model.

3. Consequences & circularity data: In order to draw conclusions, explainability is a necessary part. Therefore, comparisons and analogies are made, which can be enriched with machine learning based predictions based on the given product life cycle data. Semantic dependencies need to be understood, so Linked Data approaches are suitable to identify issues and effects within the provided data.