1. Introduction
Analogy is crucial during conceptual design processes, where designers generate new ideas by adapting knowledge from familiar domains to new contexts (Reference Dahl and MoreauDahl & Moreau, 2002). Design by Analogy (DbA) has been recognized as an effective approach to fostering creativity and cross-domain innovation. Through analogical reasoning, a designer can retrieve existing principles from a known system (the source) and establish connections and transfers between seemingly unrelated domains, enabling novel and relatively unknown design opportunities (the target) (Reference GentnerGentner 1987; Reference Gust, Krumnack, Kühnberger and SchweringGust et al., 2008).
Computational tools have increasingly been used to support early-stage design activities by organising and retrieving large bodies of knowledge. Semantic networks play a critical role as knowledge-representation models that describe how concepts are related within a structured network (Reference SowaSowa, 1987). A semantic network represents ideas as nodes and their semantic relationships as edges, enabling intuitive visualization and associative exploration of knowledge. A knowledge graph can be viewed as a more formalized and computational extension of a semantic network, offering explicit and structured representations of these relationships (Reference Ji, Pan, Cambria, Marttinen and YuJi et al., 2022). These representations are particularly valuable in DbA, as their structured and interconnected nature enables designers to systematically organise and retrieve design knowledge, enrich semantic understanding, and transfer meaningful relations across domains for analogical reasoning. Therefore, semantic networks and knowledge graphs hold significant potential to support the DbA process by transforming information into actionable insights that stimulate creativity during conceptual design (Reference Shi, Chen, Han and ChildsShi et al., 2017). However, despite this potential, there is still no unified computational structure that explicitly connect the cognitive stages of DbA with an implementable semantic network technique.
Prior research of semantic-network-based tools in DbA has often focused on isolated stages, such as retrieval or mapping only, without sufficiently alignment with the overall cognitive process in DbA, which restricts their ability to support DbA systematically. To address these gaps, this study proposes a computational framework for DbA aimed at strengthening semantic-network-based support for conceptual design. The framework makes two key contributions:
-
• This study proposes a systematic computational framework for analogical reasoning by introducing four key stages: encoding, retrieval, mapping, and evaluation. The framework aligns with the cognitive processes of DbA, enabling consistent computational support throughout the DbA process.
-
• This study provides guidance for the implementation and evaluation of enabling techniques within each stage of the framework. This guidance establishes a foundation for developing practical and verifiable computational tools, promoting the systematic application of semantic networks to support creativity and innovation in conceptual design.
Collectively, the framework transforms semantic networks into active, design-support structures, providing a foundation for computational tools that enhance analogical reasoning and advance innovation in DbA.
2. Literature review
2.1. Design by Analogy (DbA)
Analogy enables the transfer of knowledge from a source domain to a target domain (Reference GentnerGentner, 1987). Design-by-Analogy (DbA) applies this mechanism to design by adapting insights from prior solutions to new or unfamiliar contexts (Reference Hey, Linsey, Agogino and WoodHey et al., 2008). This transfer may involve elements, functional relations, or system-level configurations, allowing designers to draw on knowledge beyond their immediate disciplinary fields (Reference Childs, Han, Chen, Jiang, Wang, Park, Yin, Dieckmann and VilanovaChilds et al., 2022; Reference ReisbergReisberg, 2013).
Different cognitive models of DbA have been proposed, commonly involving three or four processes. Models with three processes typically involve retrieval, mapping, and evaluation (Reference Vosniadou, Ortony and VosniadouVosniadou, 1989; Reference Gust, Krumnack, Kühnberger and SchweringGust et al., 2008; Reference Gentner and SmithGentner & Smith, 2012; Reference Gentner, Maravilla, Ball and ThompsonGentner & Maravilla, 2018), while models with four processes comprise encoding, retrieval, mapping, and evaluation (Reference HallHall, 1989; Reference FrenchFrench, 2002; Reference Kokinov and FrenchKokinov & French, 2003; Reference Linsey, Markman and WoodLinsey et al., 2012). Although earlier studies provide varying descriptions of these processes (Reference HallHall, 1989; Reference Kokinov and FrenchKokinov & French, 2003; Reference Markman and GentnerMarkman & Gentner, 1993; Reference ReisbergReisberg, 2013), models with 4 processes offer a clear basis for analogical reasoning. Therefore, models with 4 processes have been adopted here to maintain a consistent analytical framework throughout the review. In the four-process model of DbA, encoding involves organising source information; retrieval identifies potential analogies; mapping provide structured comparison of source and target problems; and evaluation assesses how well transferred insights address the target problem.
2.2. Semantic network
Semantic networks represent concepts and their relationships through graph-based structures (Reference SowaSowa, 1987). They are widely used in neurolinguistics and natural language processing (NLP) to support semantic parsing (Reference Kamath and DasKamath & Das, 2019) and word-sense disambiguation (Reference Hwang, Choi and KimHwang et al., 2011), and are often implemented as graph databases or concept maps (Reference Borge-Holthoefer and ArenasBorge-Holthoefer & Arenas, 2010).
Knowledge graphs represent an advanced form of semantic network, incorporating richer attributes and more explicit relational information (Reference Guo, Zhuang, Qin, Zhu, Xie, Xiong and HeGuo et al., 2020; Reference Hogan, Blomqvist, Cochez, d’Amato, Melo, Gutierrez, Gayo, Kirrane, Neumaier, Polleres, Navigli, Ngomo, Rashid, Rula, Schmelzeisen, Sequeda, Staab and ZimmermannHogan et al., 2021). According to their advantages in DbA, semantic networks can support the organisation, retrieval, and comparison of information across its four stages. Semantic networks expose conceptual and functional associations that could facilitate analogy-making, while knowledge graphs provide more detailed and interconnected representations that could aid systematic reasoning.
2.3. Sematic network for DbA
General-purpose semantic networks such as WordNet (Reference Fellbaum, Poli, Healy and KameasFellbaum, 2010) and ConceptNet (Reference Speer, Chin and HavasiSpeer et al., 2018) are widely used as knowledge bases in computational design tools (Reference Han, Forbes, Shi, Hao and SchaeferHan et al., 2020; Reference Bae, Kwon, Chandrasegaran and MaBae et al., 2020; Reference Georgiev and GeorgievGeorgiev & Georgiev, 2018). However, the support for DbA in these networks remains limited, as they are not specifically constructed for analogical reasoning. This highlights the need for DbA-oriented semantic structures that can better represent the functional and relational information required for analogical reasoning in design.
Research on causal-function networks and recent semantic modelling approaches demonstrates how structured relations can support knowledge discovery and the exploration of technological information. However, existing semantic networks provide only partial support for DbA. Many tools emphasis encoding or retrieval (Reference Kim and KimKim & Kim, 2012; Reference Sarica, Song, Low and LuoSarica et al., 2019; Reference Siddharth, Li and LuoSiddharth et al., 2022; Reference Siddharth and LuoSiddharth & Luo, 2024); mapping is less supported (Reference Sarica and LuoSarica & Luo, 2021; Reference Sarica, Song, Luo and WoodSarica et al., 2021; Reference Siddharth, Li and LuoSiddharth et al., 2022; Reference Cai, Liu, Jing, Zuo, Sun, Childs and ChenCai et al., 2025), and evaluation is rarely addressed (Reference Siddharth and ChakrabartiSiddharth & Chakrabarti, 2018; Reference JohnsonJohnson, 1992). These systems also often lack representations of deeper functional relations that are crucial for identifying cross-domain analogies.
These limitations reveal a two-level gap. At the surface level, current tools cover only selected DbA stages and do not support the overall process. At a deeper level, no existing computational framework consistently links semantic representations to all four DbA stages. As a result, current tools remain centred on retrieval and mapping, with limited support for encoding and especially for evaluation. Therefore, this research aims to provide the representational foundation for computational tools that support the full DbA process.
3. Framework
The key distinction between a cognitive framework and a computational framework lies in their functions and requirements. The cognitive framework explains human mental processes in DbA such as recalling experiences, identifying commonalities, and evaluating analogies (Reference Gentner and SmithGentner & Smith, 2012), whereas a computational framework translates these processes into digital operations through algorithms and interactive systems (Reference Hollan, Hutchins and KirshHollan et al., 2000; Reference Zhang and PatelZhang & Patel, 2006). To enable such translation, clear and concise requirements need to be defined to ensure that computational tools align with the needs of DbA (Reference Goel, Rugaber and VattamGoel et al., 2009; Reference Grace, Maher, Fisher and BradyGrace et al., 2015). To ensure this alignment, a set of requirements was derived from both cognitive studies and analyses of existing digital design tools (Reference Lee, Eastman, Taunk and HoLee et al., 2010; Reference Robertson and RadcliffeRobertson & Radcliffe, 2009; Reference SevaldsonSevaldson, 2005). These requirements serve as design principles that bridge cognitive understanding and computational implementation:
-
• Comprehensiveness: The framework should encompass a wide range of knowledge domains, encourage and support acquiring and integrating knowledge from different disciplines to enrich design thinking and innovation.
-
• Structured Representation: The framework should support transforming rich data and information generated from design tasks into a structured format, ensuring accessibility, organization, and re-usability of information.
-
• Adaptability and Scalability: As design needs evolve and the knowledge base grows, the framework should be able to adapt to new design challenges and expand its knowledge, maintaining the ability to update itself in constantly changing design environment.
-
• User-Friendliness: To ensure the widespread application of the framework, it should be user-friendly, simplifying complex design tasks and providing intuitive navigation and interaction mechanisms.
-
• Validation and Evaluation: The framework should include mechanisms to assess its effectiveness, validating its utility from not only just in theory but also in practice aspect. This means the framework should be capable of improving through case studies, user feedback, and performance metrics.
Cognitive research on DbA identifies four main stages of analogical reasoning: encoding, retrieval, mapping, and evaluation. Based on these stages and the requirements above, we define four computational stages as shown in Figure 1. Each stage corresponds to one stage and is designed to satisfy a subset of the requirements.
-
• Encoding Stage: Structures large and unorganised data into a semantic network, mirroring the cognitive encoding process. It ensures that rich design information is systematically represented and suitable for computational processing.
-
• Retrieval Stage: Enables designers to navigate the semantic network and retrieve information relevant to their design queries using advanced search algorithms, ensuring quick and accurate data access.
-
• Mapping Stage: Corresponds to the cognitive mapping process by establishing connections between source and target domains. It supports analogy formation, enabling insights and principles from known designs to be applied to new contexts.
-
• Evaluation Stage: Aligns with the cognitive transferring process by assessing the quality, relevance, and applicability of retrieved and mapped information through quantitative metrics, case studies, and user validation.
Overview of the proposed four-stage computational framework for design by analogy; it shows how semantic networks are used across encoding, retrieval, mapping, and evaluation
It should be noted that in some cognitive processes, retrieval and mapping are merged into a single step, where identifying relevant information (retrieval) and connecting it to new contexts (mapping) often occur simultaneously. However, in this framework, we choose to separate these two stages to highlight the differences. Retrieval focuses on locating and extracting relevant information accurately and efficiently, emphasises establishing relationships between retrieved knowledge and new design problems, enabling knowledge transformation and transfer. This separation allows for clearer implementation within the computational framework.
The four stages function as a connected process. Each stage produces results that support the next, forming a practical workflow that aligns with the cognitive processes in DbA. In the framework, encoding organises raw information into a semantic network. This structured knowledge then supports retrieval, where relevant information is located for a given design problem. The retrieved results provide the basis for mapping, where relationships between source and target domains are identified and applied. The outcomes of mapping are then examined in the evaluation stage, which checks their relevance and usefulness for the design task. The evaluation results feed back into the earlier stages. They help refine what information should be encoded and how retrieval should be carried out in later use. In this way, the framework forms a cycle in which knowledge is built, searched, connected, and tested, supporting the step-by-step reasoning process in DbA.
The framework also meets the methodological requirements defined earlier. Encoding and retrieval support structured representation and comprehensiveness by organising and accessing large bodies of information. Mapping contributes to user-friendliness by helping designers work with cross-domain relations in an understandable way. Evaluation provides the basis for validation and adaptability by examining performance and guiding improvements. Together, this process shows how the four stages operate as a single system, linking cognitive principles with computational functions in a clear and workable manner.
4. Framework implementation guidance
4.1. Encoding
Encoding is the foundational stage of the framework. It structures and organises diverse design data so that they can be searched, compared and reused in later stages. The data sources used in DbA include design documents, academic publications, patents and encyclopaedic entries (Reference Feldman and SangerFeldman & Sanger, 2006). These sources differ in format and level of specificity, which requires flexible encoding strategies. The overall aim is to extract concepts and relations and represent them in a semantic network or, when relations need explicit definition, a knowledge graph (Reference Ji, Pan, Cambria, Marttinen and YuJi et al., 2022).
The encoding process relies on established natural language processing methods (Reference Tamašauskaitė and GrothTamašauskaitė & Groth, 2023). Text preprocessing ensures consistency in spelling, punctuation and formatting. Tokenisation divides text into linguistic units for analysis. Entity recognition identifies terms relevant to design, such as functions, components or mechanisms, while dependency parsing captures the relationships among them. Knowledge is then converted into structured representations for integration in the semantic network. The reliability of extracted concepts and relations can be assessed through standard quantitative measures and, where necessary, supported by expert-in-the-loop validation to mitigate noise and domain-specific ambiguity. Large language models can support extraction but the network remains central because it determines how encoded information can be retrieved and mapped in subsequent stages. The outcome of encoding is a structured design knowledge base that organises concepts and their relations in a form suitable for computational analogical reasoning.
4.2. Retrieval
Retrieval identifies information within the encoded network that relates to a specific design query. After encoding has produced nodes and relations, retrieval methods navigate this structure to locate relevant concepts or cases. In semantic networks, associations are often inferred or statistical; in knowledge graphs, relations are explicitly defined, enabling more controlled searches guided by functional or categorical constraints.
Several categories of algorithms can support retrieval. Graph traversal methods, such as depth-first and breadth-first search, locate connected concepts around a query node (Reference DijkstraDijkstra, 1959). Subgraph matching identifies structures that resemble the query in either exact or approximate form (Reference Cordella, Foggia, Sansone and VentoCordella et al., 2004). Embedding-based approaches, including methods such as Node2Vec and GraphSAGE, represent nodes as vectors so that similarity can be computed in a continuous space (Reference Grover and LeskovecGrover & Leskovec, 2016; Reference Hamilton, Ying and LeskovecHamilton et al., 2017). These techniques allow the system to return concepts, examples or design precedents that form the candidate set for analogical reasoning.
4.3. Mapping
Mapping links information retrieved from the source domain to the target design problem. It identifies parallels that can support the transfer of principles or mechanisms across domains. In the semantic network, mapping is enabled by analysing relationships and attributes of retrieved concepts and connecting them to elements of the design problem.
Similarities may be structural, attributional or overall (Reference McAdams and WoodMcAdams & Wood, 2002). Structural similarity concerns the arrangement of relationships among elements. Attributional similarity focuses on shared properties or descriptors. Overall similarity integrates both perspectives for complex problems. These types of similarity provide the basis for identifying whether a source concept offers insight that can be adapted to the target context. Mapping can be supported by shortest-path methods or graph-matching techniques when structural alignment is required. In knowledge graphs, where relations are explicit, mapping often relies on following predefined links rather than computing new correspondences. In many cases retrieval and mapping may occur jointly, but here they are distinguished to clarify the transition from locating information to interpreting it for analogical transfer.
4.4. Evaluation
Evaluation examines both the structure of the encoded network and the usefulness of the analogies produced. At the network level, evaluation considers coverage and accuracy by comparing extracted concepts and relations with reference datasets or benchmarks. Standard quantitative metrics such as precision, recall, F1-score, and accuracy can be used to assess extraction performance, and may be complemented by expert-in-the-loop validation to improve reliability, particularly in complex domains.
At the DbA level, evaluation focuses on whether the retrieved and mapped results help designers understand problems, explore alternatives and generate ideas. Rather than being assessed in isolation, the effectiveness of retrieval and mapping is reflected in downstream design performance, and is commonly examined through case studies and design tasks. Creativity measures such as novelty, quality and variety provide additional insight into the value of the analogies produced (Reference AmabileAmabile, 1982; Reference Shah, Vargas-Hernandez and SmithShah et al., 2003; Reference Dean, Hender, Rodgers and SantanenDean et al., 2006). These assessments support refinement of the framework and help ensure that each stage contributes effectively to design-by-analogy.
4.5. Application cases
The framework has been further examined through three application cases: WikiLink, Patent-KG and AskNatureNet. These systems were built at different stages of the research and for different types of design knowledge, yet all of them follow the same four-stage process of encoding, retrieval, mapping, and evaluation.
Methodological process for the WikiLink case study, including encoding articles, hyperlinks, and categories into a semantic network; retrieval of related concepts; mapping through cross-domain connections; and evaluation via network-structure analysis and design case studies
WikiLink, as shown in Figure 2, uses Wikipedia as a general-purpose knowledge source. Its data extraction modules encode articles, hyperlinks, and categories into a semantic network. Retrieval is supported through functions that locate related concepts, while mapping emerges as designers trace meaningful connections across domains. Evaluation includes structural analysis of the network and design case studies that examine whether the retrieved concepts contribute to analogical reasoning.
Methodological process for the Patent-KG case study, including encoding patent elements; retrieval via graph search and similarity matching; mapping to link patents to the target problem; and evaluation based on extraction accuracy and design case studies
Patent-KG, as shown in Figure 3, uses patent texts and metadata to build a knowledge graph of technical entities and relations. Encoding extracts these elements using an unsupervised attention-based method. Retrieval relies on graph search and similarity matching. Mapping occurs as designers explore functional or relational paths that link patents to the target problem. Evaluation includes extraction accuracy and design cases that demonstrate how Patent-KG helps explore alternative solutions.
Methodological process for the AskNatureNet case study, including encoding biological knowledge into a semantic network; retrieval guided by problem statements; mapping biological functions to engineering needs; and evaluation via expert judgement and designer studies
AskNatureNet, as shown in Figure 4, focuses on biological knowledge. Encoding organises biological strategies and functional descriptions into a semantic network. Retrieval is driven by problem statements, identifying relevant biological examples. Mapping aligns biological functions with engineering needs. Evaluation includes expert judgement and experimental studies with designers to assess whether the biological analogies improve idea generation.
Taken together, these three tools show that the four stages of the framework can be realised on different types of data and in different design contexts. WikiLink, Patent-KG and AskNatureNet each implement encoding, retrieval, mapping and evaluation in a concrete way, and each has been evaluated through a combination of quantitative metrics and design studies. This provides practical support for the feasibility of the framework and indicates that it can serve as a basis for future computational tools that aim to support DbA.
5. Discussion
In this study, a comprehensive framework has been presented that has been designed to leverage semantic networks to support the DbA process, translating cognitive reasoning into computationally implementable stages. While prior studies have examined individual aspects of DbA or developed tools addressing isolated stages, this work contributes a unified computational framework that explicitly aligns the established cognitive stages of DbA with corresponding computational functions. The framework ensures methodological clarity and system scalability by clarifying the functional requirements and integrating them across four interconnected stages: encoding, retrieval, mapping, and evaluation. The framework was further verified through three case studies: WikiLink, Patent-KG, and AskNatureNet, each demonstrating the applicability of the proposed framework to different types of design knowledge.
The proposed framework bridges cognitive and computational perspectives to advance design innovation. Rather than introducing new cognitive theories or semantic technologies, the primary contribution lies in systematically integrating existing theories, representations, and techniques into a coherent and implementable workflow for DbA. The framework supports designers in leveraging cognitive strengths while enhancing efficiency through digital execution and evaluation, enabling a more systematic and evidence-based DbA process in design.
By addressing gaps in clarity, systemic integration, and implementability, the framework transforms semantic networks from passive knowledge repositories into robust design-support structures. Beyond its current implementation, this study provides a systematic understanding for the adaptation of a semantic network, and provides a guidance for implementation in DbA, giving a foundation for future extensions, including computational tools for DbA and models that address specific stages of the process, to further enhance the implementability, evaluability, and creative potential of computational design systems.
6. Conclusions
A computational framework for use in design has been presented that operationalises DbA through semantic networks, aligning cognitive reasoning with computational implementation. Guided by five key requirements: comprehensiveness, structured representation, adaptability, user-friendliness, and evaluability, the framework integrates four stages: encoding, retrieval, mapping, and evaluation. These stages collectively enable the encoding of unstructured design information into organised knowledge, the retrieval of relevant data for problem-solving, the mapping of relationships between source and target domains to support analogy formation, and the evaluation of outputs through measurable indicators, establishing a systematic and traceable process for DbA.
By integrating cognitive insights with computational mechanisms, systematic understanding for the adaptation of a semantic network has been developed, and this provides guidance for implementation in DbA. The framework was further verified through three case studies: WikiLink, Patent-KG, and AskNatureNet. Each case implements the proposed framework to address different types of design knowledge.
The study addresses current limitations in defining functional requirements, ensuring technical implementability, and integrating processes across the DbA workflow, providing a practical foundation for developing computational tools that better support designers in exploring, relating, and applying cross-domain knowledge during conceptual design.
Open access
