Graphical Abstract Figure

The expanded C-K method demonstrating the roles of both LLMs and human designers

Graphical Abstract Figure

The expanded C-K method demonstrating the roles of both LLMs and human designers

Close modal

Abstract

To obtain innovative concepts in the design, designers often need to retrieve and use interdisciplinary knowledge. Concept–knowledge (C–K) theory emphasizes the role of knowledge and introduces the knowledge (K) space and concept (C) space, employing operators to transform the contents between these spaces. Some studies, based on this theory, have successfully provided designers with different forms of knowledge to stimulate concept generation. However, the amount of knowledge provided in these studies is limited, and they fail to offer convenient methods for knowledge retrieval and reasoning, making it challenging to meet the needs of conceptual design across different fields. This paper proposes an enhanced C–K method leveraging large language models (LLMs) to help designers retrieve knowledge and uncover potentially new concepts. Our method redefines the C space and K space within the context of LLMs, dividing the properties of concept into function, appearance, and technology, and requiring the knowledge to correspond to these properties, thereby facilitating a structured connection between concepts and knowledge. Based on this definition, we achieved flexible knowledge retrieval and concept ideation leveraging LLMs. We also conducted a case study on wearable devices to validate our method. The results showed that our method helped designers to retrieve professional knowledge and inspired them to create feasible and innovative concepts.

1 Introduction

Conceptual design is an early stage of product development, aimed at providing solutions and frameworks for subsequent product realization. During this process, designers are required to propose multiple concepts to solve a given problem. The generation of feasible and innovative concepts has been a concern for many design scholars and practitioners [1], ensuring that these concepts meet design requirements and differ from existing solutions [2]. In practical work, designers employ imagination and creativity to generate such concepts while retrieving relevant knowledge [2] based on different requirements. This knowledge ensures feasibility and stimulates the generation of new concepts [35]. However, conceptual design involves knowledge from various fields and disciplines. The effective retrieval and application of such diverse knowledge impose significant challenges for designers [6].

Researchers have developed methods and tools to support conceptual design through knowledge processing and learning. Data-driven tools, for instance, utilize methods like semantic networks [711], knowledge graphs [12,13], and knowledge bases to store and organize knowledge. These tools assist designers in retrieving [13] and understanding knowledge, highlighting the significance of effective knowledge utilization. However, there remains a need for theoretical models that can structure the process of knowledge-supported innovation [14]. Among the most notable theoretical contributions in this field is the concept–knowledge model (C–K model) [15].

The C–K theory offers a structured process model for conceptual design [16] by dividing the design process into two spaces: the concept space (C space) and the knowledge space (K space). The theory also proposes operators (C K, C C, K K, and K C) that govern the co-evolution of concepts and knowledge as an effective framework for concept generation. Building upon this framework, researchers have developed some knowledge-supported design tools [14,17,18], which store knowledge in structured formats for easy understanding by designers [14,17]. However, the knowledge representations in these studies depend on specific algorithms and data structures (such as nodes in knowledge graphs [14]), making it challenging for designers to quickly learn these forms to retrieve and filter knowledge. Additionally, in conceptual design, designers need to understand various aspects of knowledge, including functionality, appearance, and technology [19], and update this knowledge through reasoning. Therefore, the practice of merely providing and displaying knowledge in previous studies is insufficient.

To address the aforementioned research gaps, we propose expanding C–K theory and utilizing large language models (LLMs) [20] to develop an enhanced C–K method. Trained on massive dataset, LLMs can provide knowledge across different interdisciplinary design fields [21,22]. Their natural language processing [23] and context learning capabilities [24] also enable LLMs to process structured knowledge and concepts within the C–K theory and performing reasoning process on it. Notably, LLMs require only natural language input, thus avoiding the previous challenge where designers needed to learn knowledge representation forms. These capabilities prompt us to utilize LLMs as a technical solution for our method. LLMs also have inherent limitations, such as hallucination effects [25], when dealing with complex tasks or difficulty generating new ideas in creative work [26]. This necessitates the development of a more descriptive C–K method to better capture the complex concept design process. Thus, we redefine the C and K spaces and operators to align with concept and knowledge forms in the design. This facilitates the use of LLMs to identify the relationships between concepts and knowledge, resulting in an enhanced C–K method leveraging LLMs. The main contributions of this research can be described as follows:

  • An enhanced C–K method for the conceptual design. This method distinguishes the properties of concepts in the C–K theory and defines the supportive role of knowledge for these properties, forming a structured connection between knowledge and concepts, thereby facilitating concept exploration.

  • An efficient and convenient method for knowledge generation and concept exploration. By leveraging LLMs, we achieve flexible knowledge retrieval and allow designers to reason with the knowledge based on their needs. We also use LLMs to ideate new concepts, and then evaluate and refine them using knowledge or user requirement information.

  • A case study on wearable devices within the rehabilitation field to validate our method. The study indicates that by employing our method, designers acquire interdisciplinary knowledge, utilizing this knowledge to generate, evaluate, and refine concepts. The final design concepts are innovative and feasible within the context of rehabilitation.

2 Related Works

2.1 Conceptual Design.

Conceptual design plays an essential role as it includes critical decision-making processes and the creation of innovative design concepts [27]. To systematize the uncertain [28] and challenging [29] process of concept generation, a variety of design methods have been established. For example, the TRIZ (Theory of Inventive Problem Solving, a method for systematic innovation) method translates design issues into standard contradictions, thereby stimulating the formulation of solutions [30]. The Kansei approach uses the mapping between user emotions and the esthetic aspects of a product to inspire designers to propose appearance design concepts [31,32]. The function–behavior–structure (FBS) framework proposes that all designs entail transformations across function (articulating the design’s objective), behavior (defining its operation), and structure (characterizing its composition) [33].

Compared to TRIZ’s focus on solving predefined contradictions [30] or Kansei’s emphasis on emotional design [32], the C–K theory provides a more flexible and comprehensive framework that accommodates the dynamic and iterative nature of design thinking. In the C–K theory, concepts are defined as design propositions composed of properties, while knowledge represents established conclusions about these properties [19]. The C–K theory presents the design as an iterative process that alternates between the knowledge space and the concept space, with operators including concept concept, knowledge knowledge, knowledge concept, and concept knowledge to expand both the knowledge and concept spaces [34,35]. It is a general descriptive model with a strong logical foundation that can accurately describe this process (and even the changes in concepts and knowledge within other design methods [36]), while also facilitating the co-evolution of concepts and knowledge.

Beside the mentioned approaches involving standardized processes, the advancement of information technology and the evolving demands of users have spurred the development of data-driven tools for the concept design. A data-driven approach has been developed to identify functions suitable for integration into a product platform, utilizing patent data and function network analysis [37]. Researchers have also explored automated methods for representing design knowledge as semantic networks [7]. These networks are believed to provide a more comprehensible, detailed, informative, and structured representation of design knowledge in specialized fields such as engineering design [7]. Furthermore, knowledge graphs have also been utilized for concept design. For instance, a comprehensive technology knowledge graph has been constructed by extracting terms from patent texts and defining their semantic relationships for engineering design and innovation [38]. The introduction of a smart conflict resolution model, known as mKGCD (Multi-Knowledge Graph Contrastive Design, a method using multiple knowledge graphs for collaborative design), which employs a multi-layer knowledge graph for concept design, has been proposed to effectively align with designers’ requirements and retrieve appropriate knowledge [39]. In summary, these data-driven tools highlight the efficacy of utilizing data related to design tasks in the concept design process.

In addition, there are also researches that integrate these data-driven tools with the C–K theory. For instance, researchers have explored the construction of knowledge graphs to aid the four C–K operators in fulfilling multiple personalized requirements within the design of smart product-service systems [14,40]. However, this approach necessitates defining different knowledge graph ontologies and relationships for specific domain. This requires designers to invest time in learning the ontologies and to consider how to utilize relationships for retrieval, which presents a certain threshold. Besides, DesignNAR, which is compatible with the C–K theory, is a design assistant capable of observing the external design representation, providing suggestions to designers, and adapting its suggestion behavior based on the designers’ reactions to those suggestions [17]. Despite its potential benefits, DesignNAR requires significant initial setup and continuous adjustment to effectively align with the designers’ evolving needs and preferences. Overall, the construction of these data-driven C–K tools is often time-consuming and designers face an additional learning curve when acquiring proficiency in these specialized data-driven tools.

2.2 Large Language Models in the Conceptual Design.

Drawing upon an extensive textual data from books, encyclopedias, and the internet, transformer-based [41] LLMs such as PaLM (Pathways Language Model, a large-scale language model by Google), GPT-4 (Generative Pre-trained Transformer, a model for text generating), and LLaMa (Large Language Model Meta AI, Meta's family of large language models) possess natural language understanding and processing capabilities including text summarization, modification, and generation [20,42]. Products based on these models, such as ChatGPT, have also achieved multi-turn dialog and memory functionality [43]. These abilities enable LLMs to participate in various stages of the design process. Specifically, LLMs can extract functions and components from user requirement information to serve as product design objectives [44]. They can also help designers extract domain-specific design information from design documents [45] or generate design concepts based on different requirements [22].

Researchers have also attempted to enhance conceptual design methods by leveraging the knowledge and natural language capabilities of LLMs. Wang et al. [46] introduced a task-decomposed approach that facilitates collaboration between designers and LLMs in the application of the FBS method. In another advancement, Jiang and Luo [47] proposed AutoTRIZ, a system that automates and enhances the TRIZ problem-solving methodology by leveraging the broad knowledge and advanced reasoning capabilities of LLMs. Additionally, a novel methodology leveraging the GPT-2 and GPT-3 has been proposed to leverage the knowledge and reasoning from textual data, converting them into innovative concepts in a clear and understandable language [48]. Collectively, these design support tools and approaches have confirmed the feasibility of employing large language models to assist the conceptual design, offering designers an abundant reservoir of knowledge and a robust platform for the generation of innovative concepts.

LLM applications in conceptual design also face some challenges. LLMs may exhibit hallucination effects when dealing with complex tasks, generating incorrect information or misinterpreting facts, which poses challenges for their role in providing design knowledge. Additionally, since their training datasets are based on existing knowledge, LLM-generated creative concepts may appear diverse in form but are inherently fixed and limited. In this paper, we will examine the presence of these phenomena and extend the C–K theory to more accurately describe the concept design process, minimizing the occurrence of hallucinations during tasks, while leveraging human–machine collaboration to stimulate designers’ innovative concepts.

3 Methodology

3.1 An Expansion of the Concept–Knowledge Theory.

To address the aforementioned research gaps, we have expanded the C–K theory, adapting it to better suit the practical requirements of conceptual design while integrating it with the capabilities of LLMs. In this research, our primary focus lies in the definition of the concept space (C space) and the knowledge space (K space), as well as the four operators between concepts and knowledge.

Redefinition of Concept and Knowledge. In the C–K theory, a single concept C is represented by a proposition composed of multiple properties (i.e., Cn:P1,P2,Pn) [15]. These properties are considered undifferentiated and unstructured, while knowledge is defined as propositions with logical states (true or false) [19]. However, in the context of conceptual design, designers must consider a wide range of properties [19]. There are several references on how to describe these properties. The FBS model categorizes them into function, behavior, and structure [33], while some design theory books discuss design components from the perspectives of functional interrelationship, working interrelationship, constructional interrelationship, and system interrelationship [49]. We believe that the properties of functionality, appearance, and technical aspects can encompass more information considered in the design and provide greater assistance in concept generation. The product’s utility, usage methods, and interaction behaviors can be unified under function property. Similarly, the material, structure, and other physical information can be described by appearance (structure) properties. Since it is necessary to help designers understand technical principles and design standards, a dedicated technical property should be added. This classification method also creates links between properties: functionality can be supported by appearance and technical aspects, while appearance and technical aspects can inspire functional ideas for designers. In the expanded C–K theory, each concept should consist of these three properties.

Moreover, for designers, merely judging a concept as true or false does not provide sufficient information. This necessitates knowledge that goes beyond mere judgment and offers additional insights to facilitate innovation thinking. In the expanded C–K theory, knowledge should correspond to the definitions of the aforementioned concepts, encompassing explanatory knowledge about properties such as technical and appearance aspects.

Operators for Knowledge Retrieval and Reasoning. In the C–K theory, the C K operator is defined as making additions to or removals from the properties of a concept C to form propositions, which are then evaluated as true or false to generate knowledge [50]. Similarly, the K K operator involves reasoning from propositions to derive further knowledge [50]. However, the knowledge generated often involves properties that are highly similar to those in the original concept. In the conceptual design, designers often seek knowledge that not only aligns with design requirements but also encompasses new properties, such as emerging technologies and functionalities within the field. This demands faster and more flexible knowledge retrieval to support the refinement of their concepts. In line with the aforementioned expansion, the C K operator should be used to quickly retrieve relevant knowledge about appearance, structure, technology, etc. based on several properties of the concept. This is possible with the support of LLMs’ massive knowledge repository. In addition, the knowledge about different properties can be challenging for designers to quickly grasp, and the knowledge acquired in a single instance is often insufficient. Designers may need to retrieve additional relevant knowledge to aid their understanding and serve as supplementation. In this research, we propose utilizing LLMs’ question answering (QA) [51] and context understanding capabilities [52] to generate knowledge, assisting designers in retrieving and reasoning about relevant knowledge, thereby serving as a K K operator.

Operators for Concept Generation, Evaluation, and Refinement. The C–K theory defines the C C operator to modify concepts by adding or removing properties, while the K C operator is realized by modifying properties with knowledge support. However, in conceptual design, designers need to be more active and divergent, and they may come up with significantly different concepts in a short period of time, which cannot be represented by simple property changes. Building on the idea of knowledge-driven innovation in the C–K theory, we enhance the C C and K C operators by utilizing the relationships between properties based on the redefinition of C and K. Using these relationships means that knowledge related to properties can be leveraged to deduce potential new concepts using the reasoning capabilities of LLMs [53], realizing the K C operator. For the C C operator, we believe it should correspond to the concept refinement process in real design, namely the evaluation and improvement of concepts. Based on our previous expansion, the context understanding capabilities [24] of LLMs can be used to evaluate whether a concept meets the requirements and refine the concept by modifying different properties. Since this involves the iteration of innovative concepts, such refinement should be primarily led and decided by designers. In addition, in real-world conceptual design scenarios, the initial concept is often a commercial requirement. This requires us to implement a special C C operator, which maps the initial concept composed of business requirements to a concept composed of properties such as functions and appearance.

3.2 Our Enhanced Concept–Knowledge Method Leveraging Large Language Models.

Based on the expansion approach outlined above, we can describe an enhanced C–K method utilizing LLMs (also shown in Fig. 1). As specific technological solutions for implementation, the most advanced LLMs, such as GPT-4 and LLaMA, are applicable to our method. In the following sections, we will introduce the details about the method.

Fig. 1
Our expanded C–K method, demonstrating the roles of both LLMs and human designers within the method
Fig. 1
Our expanded C–K method, demonstrating the roles of both LLMs and human designers within the method
Close modal

3.2.1 Definitions of C Space and K Space.

The C space encompasses design concepts at various stages of iteration. We employ two-dimensional indexing to distinguish and represent concepts. Specifically, Cn denotes the nth proposal introduced during concept exploration, while Cnm signifies its evolution through m iterations of modification and refinement of Cn. As shown in Fig. 1, the data structure for a typical concept is represented as Cn(F,A,T), where

  • F (function) encompasses functional properties such as utility, usage methods, and interaction modes of the product.

  • A (appearance) encompasses the product’s appearance and structure.

  • T (technology) encompasses core technologies, operational principles, design standards, and other technological factors supporting the product.

These three properties can comprehensively describe a concept, and there exist potential relationships between properties [19], including the requirements of functionality on appearance and structure, as well as the inspirational influence of structure and technology on potential functionalities. Notably, the initial design requirements for product design are defined as C0. The data structure of C0 is denoted as C0(R1,R2,,Rn), where each Rn represents a description of a business requirement or user need. This requirement typically specifies potential functional or appearance properties (otherwise it cannot form C0), which can be easily transformed into a complete Cn(F,A,T). Therefore, we regard it as a special type of concept.

The K space encompasses various knowledge acquired during the conceptual design process. The data structure of knowledge is defined as Kni (or more specifically indicating its relation to properties, Kni(F/A/T)), indicating ith piece of knowledge to a specific function, appearance, or technology properties of the concept Cn. Specially, knowledge reasoned from Kni can be labeled as Kni1, Kni2, and so on. Its content may include simplified explanations of properties, as well as supplementary knowledge from relevant domains, aiming to facilitate designers’ understanding of the knowledge and acquisition of related information.

3.2.2 Large Language Model-Aided C K Operator: Knowledge Retrieval Based on Natural Language Processing.

The C K operator serves as the primary source of knowledge in our method. Designers rely on this operator to acquire knowledge for Cn(F,A,T). This knowledge assists designers in understanding the F/A/T properties within Cn and inspires potential concepts based on this understanding. When implementing this operator, the input consists of concepts Cn described in natural language (see Fig. 1). LLMs match knowledge related to F/A/T properties within their knowledge base, formed through training on large-scale datasets, and output it as Kni. The prompts used by the operator are presented in Table 1. The acquired knowledge is typically relevant to the concept Cn, serving as essential information for designers to understand and refine the concept. This operator can be represented as a pattern: Pattern I

Table 1

The prompt templates used to conduct operators: C0C, CK, KK, KC, CC

OperatorPrompt template
C0CDesign Requirements: [C0]; Design Concept: Function-, Appearance-, Technology-
Please refer to the provided form of the concept and create a design concept that meets the design requirements, ensuring the innovation of the concept.
CKDesign Requirements: [C0]; Design Concept: [Cn]
Please provide relevant knowledge from the aspects of function, appearance, and technology based on the above design requirements and concepts.
KKKnowledge: [Kni(F)]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge regarding function, please deduce and reason out new knowledge in the aspects of mapping, visibility, feedback, constraints, and consistency by following the steps below: (see more details in APPENDIX)
The explanations for the five aspects are as follows: (see more details in APPENDIX)
Knowledge: [Kni(A)]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge about appearance, please follow these steps to deduce and reason out new knowledge in the aspects of aesthetic appearance, component layout, connection methods, and material selection by following the steps below: (see more details in APPENDIX)
The explanations for the four aspects are as follows: (see more details in APPENDIX)
Knowledge: [Kni(T)]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge about technology, please follow these steps to deduce and reason out new knowledge in the aspects of engineering principles, technical specifications, and production processes by following the steps below: (see more details in APPENDIX)
The explanations for the three aspects are as follows: (see more details in APPENDIX)
KCKnowledge: [Kni]; Design Concept: Function-, Appearance-, Technology-
Please use the knowledge provided above to inspire an innovative and feasible design concept following the provided form.
CCDesign Concept: [Cn].
Evaluation Criteria: (1)Requirement Satisfaction: Sub-criteria listed as [Requirement1], [Requirement2] [Requirementn] (2)Knowledge Satisfaction: Sub-criteria listed as [K(F)], [K(A)], [K(T)]. Scoring Method: 4-point scale (see more details in APPENDIX)
Please evaluate the above design concept based on the evaluation criteria, using the scoring method, and present the results in a tabular format.
Please provide improvement directions for the aspects in the evaluation table that have received negative scores (1, 2), and then propose an improved design concept in the following format. “Design Concept: Function-, Appearance-, Technology-”
OperatorPrompt template
C0CDesign Requirements: [C0]; Design Concept: Function-, Appearance-, Technology-
Please refer to the provided form of the concept and create a design concept that meets the design requirements, ensuring the innovation of the concept.
CKDesign Requirements: [C0]; Design Concept: [Cn]
Please provide relevant knowledge from the aspects of function, appearance, and technology based on the above design requirements and concepts.
KKKnowledge: [Kni(F)]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge regarding function, please deduce and reason out new knowledge in the aspects of mapping, visibility, feedback, constraints, and consistency by following the steps below: (see more details in APPENDIX)
The explanations for the five aspects are as follows: (see more details in APPENDIX)
Knowledge: [Kni(A)]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge about appearance, please follow these steps to deduce and reason out new knowledge in the aspects of aesthetic appearance, component layout, connection methods, and material selection by following the steps below: (see more details in APPENDIX)
The explanations for the four aspects are as follows: (see more details in APPENDIX)
Knowledge: [Kni(T)]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge about technology, please follow these steps to deduce and reason out new knowledge in the aspects of engineering principles, technical specifications, and production processes by following the steps below: (see more details in APPENDIX)
The explanations for the three aspects are as follows: (see more details in APPENDIX)
KCKnowledge: [Kni]; Design Concept: Function-, Appearance-, Technology-
Please use the knowledge provided above to inspire an innovative and feasible design concept following the provided form.
CCDesign Concept: [Cn].
Evaluation Criteria: (1)Requirement Satisfaction: Sub-criteria listed as [Requirement1], [Requirement2] [Requirementn] (2)Knowledge Satisfaction: Sub-criteria listed as [K(F)], [K(A)], [K(T)]. Scoring Method: 4-point scale (see more details in APPENDIX)
Please evaluate the above design concept based on the evaluation criteria, using the scoring method, and present the results in a tabular format.
Please provide improvement directions for the aspects in the evaluation table that have received negative scores (1, 2), and then propose an improved design concept in the following format. “Design Concept: Function-, Appearance-, Technology-”

 C K:

  INPUT Cn(F,A,T)

  MATCH Knowledge as ,:Kni

  RETURN {Kn1,Kn2,,Kni}

3.2.3 Large Language Model-Aided K K Operator: Knowledge Question-Answering and Knowledge Reasoning.

In our definition of knowledge, we consider that knowledge spanning different domains may pose some understanding challenges for designers, and generating knowledge in a one-off manner is often insufficient to support conceptual design. Therefore, within the K K operator, we allow designers to explore and reason about relevant knowledge according to their own needs, corresponding to the following two distinct approaches, which can also be seen in Fig. 1.

One approach is knowledge-based questioning and answering, which aids in knowledge comprehension. In the C K operators, LLMs retrieve and output lots of knowledge, spanning numerous fields. However, designers may find it challenging to grasp this knowledge without consulting background information. As a model with the capability for multi-turn dialog and contextual understanding, LLMs can explain knowledge and provide background information to assist designers in comprehension through interactive Q&A sessions. In this process, designers input knowledge they do not comprehend and pose questions in natural language to LLMs, which promptly respond. Since LLMs’ responses contain additional knowledge, this is also considered as a K K operator generating new knowledge.

Another approach involves reasoning with specified directions for knowledge. Unlike QA approach designed to help designers understand knowledge, this approach aims to retrieve and reason about other knowledge related to the current one. Based on relevant product design theories [54,55], we have provided knowledge reasoning guidelines for conceptual design concerning the three properties of functionality, appearance, and technology. Our guidelines include potential directions and steps for reasoning about this knowledge, serving as prompts for reasoning (see prompts in Table 1). During the use of our method, designers can modify these prompts in accordance with the actual requirements. They then input the knowledge requiring reasoning, providing it to LLMs. This process employs the chain-of-thought reasoning method [56], where LLMs utilize the reasoning directions and steps provided by the prompts to generate more comprehensive and in-depth knowledge. Considering that function and appearance are supported by technical properties, it is recommended that designers first perform reasoning on the technical properties. This approach ensures that they have a thorough understanding of the properties when acquiring knowledge, thereby reducing the difficulty of comprehension.

The reasoning approach of the K K operator can be described as follows (the QA approach depends on the designer’s needs and cannot be described using a fixed pattern):

Pattern II

 K K (reasoning):

  INPUT Kni(F/A/T)

  INPUT Instruction,:,: //Given by the designers

  RETURN Kni1(F/A/T)

3.2.4 Large Language Model-Aided K C Operator: Ideation Based on the Relationships Between Knowledge andConcepts.

Within our defined properties, product functions often rely on both appearance (or structure) and technological implementation. This relationship is twofold: on one hand, it is constraining, meaning that functions cannot be achieved without both structure and technology; on the other hand, within divergent thinking processes, this relationship aids designers in discovering potential new properties. Building upon this foundation, we have implemented two distinct K C operators. These operators facilitate the ideation of new concepts by leveraging the relationship between knowledge and concepts.

One approach involves utilizing the constraint relationships among F/A/T. Functions necessitate both structural and technological support. Therefore, knowledge about functions often indicates that a product should utilize specific technologies or structures for implementation. Correspondingly, knowledge regarding appearance and structure may also require the utilization of certain technologies. That is, F imposes constraints on A/T, and F/A imposes constraints on T. We employ LLMs by inputting knowledge related to F/A and the corresponding concept Cn. Leveraging its contextual understanding and reasoning abilities (which, due to inputting old concepts, also involves a form of few-shot learning), LLMs engage in transformation of old concepts. It generates new appearance/technology properties to meet the requirements of relevant knowledge, producing reasonable new concepts. Designers are responsible for supervising this process and retain the final decision-making authority, requesting LLMs to regenerate concepts when the new ones do not meet design requirements.

Pattern III

 K C:

  INPUT Cn(F/A/T)

  INPUT Kni(F/A/T)

  GENERATE Cn+1(F/A/T) from Cn by Kni

  RETURN Cn+1(F/A/T)

Another approach involves leveraging the stimulating effect of F/A/T to engage in generating innovative concepts. Technical-related information not only supports the implementation of functionality but also stimulates the generation of new functional/appearance properties. Similarly, appearance or structural information can stimulate functions properties, i.e., T can stimulate F/A, and A can stimulate F. This process can be a collaborative process between LLMs and humans. By inputting knowledge related to A/T, LLMs can generate new concepts, which can then be filtered by humans. However, as LLMs are pre-trained models, its output often correlates with old product concepts within the training data. Therefore, designers need to judge whether the concepts provided by LLMs are highly similar to existing products. If such a situation arises, they are required to employ divergent thinking to explore more potential functions and appearance (or structures), integrating them as properties into concepts to facilitate innovative concept generation.

3.2.5 Large Language Model-Aided C C Operator: Concept Evaluation and Refinement.

We employ the C C operator as a primary approach to concept refinement from Cn to Cnm, situated in the later stages of the conceptual design process. Throughout this process, LLMs evaluate Cn(F,A,T) based on two dimensions: whether the functional/appearance/technology (F/A/T) within concept Cn can address commercial requirements in C0(R1,R2,), and whether they align with the knowledge Kni. LLMs do not make direct decisions; rather, it outputs relationships between F/A/T and their relevance to commercial requirements Rn or alignment with knowledge, providing explanations in natural language. Human designers use these relationships and explanations to filter and make decisions regarding the generated concepts. If unsatisfied, instructions can be given to LLMs to generate updated concepts based on these relationships. However, due to the limitations of the model’s dataset, LLMs often produce concepts that better meet the requirements but struggles to provide innovative concepts not present in the dataset. Human designers need to manually edit concepts, employing divergent thinking to introduce more innovative F/A/T properties (similar to the K C operator process) to generate new concepts. Designers can then input these new concepts the LLMs, execute the C C operator again, and obtain evaluation results.

Compared to manual evaluation methods that require considering relevant knowledge and assessment criteria, the aforementioned C C operator delegates the evaluation task to LLMs, preventing designers from having to handle excessive information and reducing cognitive load. Additionally, during the operator process, designers perform the innovation tasks while LLMs are responsible for evaluation, forming a closed-loop and effective human–artificial intelligence (AI) collaboration method.

The C C operator also encompasses a special case, where there is a mapping from the initial requirement, C0, to the initial concepts. LLMs can identify property-level user needs for individual components and properties of a product [44]. Therefore, we employ prompt instructions to request it to map C0 to various sets of different F/A/T properties, thereby generating initial concepts. The two types of prompts utilized by the C C operator can be seen in Table 1.

The C C operator can be represented as following patterns:

Pattern IV

 C C:

  INPUT Cn(F/A/T)

  EVALUATE Cn by {Kni, C0}

  GENERATE Cnm(F/A/T) //Human–AI collaboration

  RETURN Cnm(F/A/T)

Pattern V

C0C1:

  INPUT C0{R1,R2,,Rn}

  MATCH F/A/T by {R1,R2,,Rn}

  RETURN C1(F/A/T)

4 Workflow of Our Method

Building upon the implementation of operators using LLMs, we have developed the C–K workflow. Figure 2 illustrates the workflow and reflects the roles of the four operators we defined within it. Drawing from the double diamond model [57] in product design and previous C–K design researches, we partition the workflow into three stages: requirement analysis, ideation, and refinement and decision. This division aligns with the practical context of conceptual design and facilitates the use of the operators we defined, with more details as follows.

Fig. 2
An illustration of the workflow, demonstrating the definitions and operators from our C–K method
Fig. 2
An illustration of the workflow, demonstrating the definitions and operators from our C–K method
Close modal

4.1 Requirement Analysis Stage.

The requirement analysis is a crucial stage in product design, wherein designers gather user requirements for subsequent concept generation. Our method takes predefined initial user requirements as input. As shown in Fig. 2, designers need to execute the C C operator, leveraging LLMs to map these requirements onto several potential functional/appearance/technological properties, generating a range of initial concepts. This process minimizes the cost associated with analyzing user requirements.

4.2 Ideation Stage.

In this stage, various knowledge relevant to the functionality, structure, and technical properties of the concepts is retrieved. Following retrieval, new knowledge is reasoned based on the generated knowledge, stimulating the generation of new concepts.

As shown in Fig. 2, after completing the C C operator in the previous stage, designers should proceed with the C K operator. LLMs will generate knowledge Kni related to the initial concepts (C1, C2, …,Cn). Since these concepts are mapped from requirements, this knowledge often serves as essential background information necessary for designers. At this point, the subsequent course of action depends on the designer’s background knowledge and comprehension. They may proceed with the K K operator, generating new Kni1 through questioning with LLMs or knowledge reasoning, thereby aiding comprehension or acquiring additional relevant knowledge. If designers have a thorough comprehension of the knowledge, they can skip this step.

In the aforementioned process, all knowledge K is directly or indirectly derived from a certain concept Cn, so the supportive role of Kni on Cn can be explicitly indicated. Building upon this, the K C operator can be executed. Consistent with the aforementioned method, two possible approaches exist here, requiring designers collaborate with LLMs to propose these new concepts Cn.

4.3 Refinement and Decision Stage.

In this stage, the refinement of concepts is conducted to attain detailed and feasible concepts, ultimately leading to the concept Cnm as the final solution.

Following the ideation stage, where a relatively comprehensive concept Cn has been generated, the C C operator should be executed. Initially, LLMs, in conjunction with relevant knowledge Kni and the initial requirement C0, evaluate Cn(F,A,T), outputting whether the F/A/T adequately addresses the requirement C0 or aligns with the knowledge. In cases where designers perceive the concept as infeasible and requiring refinement, LLMs modify the concept based on instructions, or designers manually edit the concept.

Within the workflow, designers have the flexibility to backtrack steps and execute previous operators according to their discretion. For instance, as shown in Fig. 2, during the refinement and decision stage, designers can revert to the ideation stage’s K K operator to acquire additional knowledge supplementation.

5 Case Study

5.1 Backgrounds and User Requirements.

Chronic diseases pose a significant global health challenge, accounting for 74% of deaths worldwide, according to the World Health Organization. With the rising prevalence of chronic diseases and the growing awareness of public health, wearable medical devices are becoming increasingly crucial due to their ability to continuously monitor physiological data in real-time and empowering healthcare professionals to deliver personalized care. Underscoring this demand, the official survey report on the wearable medical devices market reveals that the global market size reached USD 26.8 billion in 2022, with a projected compound annual growth rate of 25% for 2023–2032.

In the rehabilitation training field, particularly for orthopedic and neurological conditions, force-sensor-based wearable medical devices have emerged as the most popular form. Among them, hand-worn wearable medical devices are commonly used for hand bone and nerve rehabilitation. These devices mainly take the form of gloves or hand braces equipped with sensors and force support components. They can monitor physiological data in real-time, such as muscle pressure in specific areas. The main issue is that, given the long-term nature of hand rehabilitation and the variety of training movements, hand-worn wearable medical devices should provide a more comfortable and flexible wearing experience, rather than fixed braces or gloves. Therefore, our initial requirement is to design rehabilitation devices for patients needing hand rehabilitation that can accurately measure training data while being more convenient to wear. This case study was conducted over a three-week period by a team of two randomly selected industrial design graduate students.

The demand for enhanced comfort and flexibility in wearable devices extends across various medical fields, including disciplines such as material science and rehabilitation medicine that often exceed the expertise of individual designers. Leveraging the proposed C–K method offers a promising solution for efficiently evolving hand-worn wearable medical devices. As a technological solution for our method, we selected the state-of-the-art LLMs in our case study, specifically GPT-4 [20].

5.2 Hand-Worn Wearable Medical Device Design Using Our Concept–Knowledge Method

5.2.1 Requirement Analysis.

Guided by the workflow shown in Fig. 2, the design process starts by the precise definition of design requirement C0. Drawing upon the insights from this section, design requirement C0 is formulated as follows: “To develop a hand-worn wearable medical device for rehabilitation patients that enables accurate muscle activity monitoring, comfortable and flexible wearability, and facilitates rehabilitation training.” Subsequent to defining C0, the initial concept C1 was generated through the application of the C C operator by GPT-4 and refined by the designers. As shown in Fig. 3, concept C1 includes three functional properties, two appearance properties, and two technical properties. Specifically, in terms of function, the proposed device exhibits adaptability, seamlessly adjusting to fit the skin during various rehabilitation exercises. Additionally, it incorporates data analysis capabilities to provide feedback on rehabilitation progress. Furthermore, the device offers auxiliary heat and cold therapy functionalities. Esthetically, the device adopts a detachable modular design, enabling customization to suit specific treatment needs. Integrated light emitting diode (LED) lights serve to indicate the muscle areas requiring movement. Technically, the device employs miniature sensors to monitor physiological data like pressure. Notably, the first functional property was modified from the original generated concept of automatic pressure and support adjustment. Because pressure and support levels are tailored by healthcare professionals based on a comprehensive assessment of patients’ recovery stage, physical capabilities, and other relevant factors in the prevailing practice.

Fig. 3
Illustration of the case study process, showcasing the concepts and selectively presented knowledge items as applied by the designers. Due to the extensive content involved in the entire process, this figure only illustrates the key steps and details. The complete, detailed list of knowledge items can be found in the appendix.
Fig. 3
Illustration of the case study process, showcasing the concepts and selectively presented knowledge items as applied by the designers. Due to the extensive content involved in the entire process, this figure only illustrates the key steps and details. The complete, detailed list of knowledge items can be found in the appendix.
Close modal

5.2.2 Ideation.

During the ideation stage, 14 knowledge items related to function, appearance, and technology were generated by applying the C K operator to each property of concept C1. Figure 3 shows the K points that the designers considered useful. All generated K points are included in the appendix. To gain more comprehensive knowledge for ideation, the designers input K12(F), K18(A), and K111(T) into the K K operator to perform further reasoning. For example, for flexible electronic materials (K12(F)), reasoning and deduction led to the knowledge that devices using flexible electronic materials should have intuitive visible user instructions, such as color and texture (K121(F)), and provide multimodal feedback, such as tactile and auditory feedback (K123(F)). Figure 3 shows a set of useful reasoned knowledge items. K186(A), K1115(T), K1117(T), and K1118(T) mention material performance requirements, such as being hypoallergenic and durable. To identify materials that meet these requirements, the designers interacted with GPT-4 through QA to obtain detailed information. Ultimately, they selected medical-grade silicone as the most suitable material for the product.

After reasoning over the generated knowledge, the designers selected more insightful items, specifically those that align with the design requirements and involve new properties. They then applied the K C operator to derive an innovative concept, C2, which aligns with the design properties and criteria mentioned in the knowledge about rehabilitation products (for example, C2 effectively satisfies the multimodal feedback mechanism proposed by K123(F) and the flexible materials suggested by K186(A)). As described earlier regarding this operator, some properties of the resulting concept C2 are derived from the LLMs (such as intelligent software), while others stem from the designers’ exploration (such as multimodal feedback, considered an important interactive feature). The designers integrated these properties, generating concept C2. Figure 3 illustrates the utilized knowledge items and concept C2. Concept C2 integrates color-coded indicators to differentiate distinct wearable areas and color-changing LED lights indicate the specific muscle groups requiring activation. Additionally, it provides haptic feedback in the form of gentle vibrations to confirm the completion of a set of movements. Besides, concept C2 features a modular design, including three individual modules: a palm module, a joint module, and a wrist module. Leveraging miniature sensors and a flexible printed circuit board (PCB), the device utilizes medical-grade silicone to ensure comfortable pressure monitoring. The integrated intelligent software analyzes rehabilitation data and generates comprehensive progress reports.

5.2.3 Refinement and Decision.

In the final stage, the designers evaluated the concept C2 against two primary criteria: fulfillment of the initial design requirements and satisfaction of knowledge requirements. As shown in the C C operator in Table 1, the designers input each requirement and knowledge item as sub-criteria into the prompt. Based on this prompt, GPT-4 evaluated the concept using the Harris Profile method [58], assigning scores on a four-point scale from 2 to +2. The assessment revealed that concept C2 performed well against most criteria. However, the fulfillment of knowledge requirements related to the adaptive interface was identified as an area for improvement. To address this issue, the designers continued to apply the C C operator, adding new properties to arrive at the final iterated concept C3. This process involved multiple rounds of evaluation and refinement of C2, including the application of the previously used K K operator. (For example, to concretize the modular and bionic design in C2, the designers conducted a QA process on relevant knowledge K182(A). GPT-4’s responses provided directions for bionic design, such as analogical design based on animal skin or plant structures, which inspired the designers to consider a fish-scale structure.) This process involved multiple rounds of evaluation and refinement of C2, with the primary stages shown in Fig. 3.

5.2.4 Description of the Design Scheme.

The final design concept C3, derived using our proposed C–K method, can be described as follows (key information is also shown in Fig. 4): The product will be a medical device designed for the rehabilitation training of patients with chronic diseases. Its primary function is monitoring muscle status in real-time during training. Additionally, it provides personalized exercise plans and data statistics to optimize users’ health management and quality of life. During training, it offers real-time feedback to remind users of their training status. This feedback is multimodal and easy to understand (e.g., visual signals and tactile vibrations during training).

Fig. 4
Our final design concept and the supporting relationship between function, technology, and appearance
Fig. 4
Our final design concept and the supporting relationship between function, technology, and appearance
Close modal

To achieve these functionalities, the product features a modular design with different modules that can be worn on various body parts, comprising skin-adhering sensors. The overall appearance is a scale-like structure, ensuring both accurate measurement and ease of wearing. Technologically, this structure is supported by advanced sensing technology and flexible circuits. For personalized rehabilitation plans, the product is supported by intelligent sensing and medical decision-making technologies, with data synchronization achieved through integration with the trainer’s software.

Based on the aforementioned design concept, we developed a prototype to assess its feasibility, specifically to determine whether the technologies and appearance involved in the concept could fulfill the intended functions as envisioned. As shown in Fig. 5(a), our prototype comprises both hardware and software interfaces. The hardware component is divided into a thumb module and a wrist module, which are connected by magnetic attraction. Different wearable modules are color-coded, with red for the thumb part and blue for the wrist part, to provide intuitive wearing instructions. During rehabilitation training, trainers can communicate with patients about the muscles need to be activated through the LED lights on the device and provide vibration feedback at the end of a set of actions. As shown in Fig. 5(c), we utilize a fish-scale-like structure to interconnect the sensors, thereby augmenting the device’s flexibility and ensuring sensor stability. Importantly, it also facilitates conductivity between different sensors. In addition, the device is equipped with a flexible PCB and multiple miniature sensors (as illustrated in Fig. 5(b)), which can accurately monitor the pressure data of different parts during the recovery process. The physiological data collected by the sensors will be transmitted to the intelligent software on the trainer’s side, providing personalized progress reports and recovery status, as shown in Fig. 5(d).

Fig. 5
The design of a hand-worn wearable medical device: (a) the prototype embodies several design properties: multimodal feedback, modular design, and wear instructions; (b) advanced sensing technology and flexible circuits: made of medical silicone, equipped with a flexible PCB and mini sensors; (c) fish-scale structure: Fish-scale design for flexible fit, efficient cooling, and enhanced conductivity, and (d) intelligent software with adaptable interface for synchronized data in rehab training
Fig. 5
The design of a hand-worn wearable medical device: (a) the prototype embodies several design properties: multimodal feedback, modular design, and wear instructions; (b) advanced sensing technology and flexible circuits: made of medical silicone, equipped with a flexible PCB and mini sensors; (c) fish-scale structure: Fish-scale design for flexible fit, efficient cooling, and enhanced conductivity, and (d) intelligent software with adaptable interface for synchronized data in rehab training
Close modal

To test the feasibility of the final concept, we validated the viability of our hardware prototype through technical methods. We employed 3D printing technology to form a fish-scale structure and flexible PCB, integrating high-sensitivity pressure sensors and conductive pads within the structure. This resulted in a basic configuration capable of detecting pressure data. Since the sensors used in the prototype are standard components with reliable performance, their functionality is generally trustworthy. However, due to the unique structure of our product, the sensors might be affected by deformation and vibrations. We conducted a pressure response test to verify whether the sensors could function effectively within the fish-scale structure.

Specifically, we placed the sensors within the fish-scale structure at various positions corresponding to practical use on the wrist, palm, and finger joint. We then applied pressure (simulating real muscle pressure values, 0–10 N) to these positions, measured the electrical signals output by the sensors, and obtained three distinct sets of data for analysis. Additionally, the data were fitted to third-degree polynomials using the least-squares fitting method to derive the calibration model for mapping pressure to voltage. The measurement results are shown in Fig. 6. The results demonstrate that different sensor groups exhibited similar logarithmic growth patterns, with sensor output voltage increasing with pressure, which aligns with the expected output pattern. Although the electrical signals output by different groups appeared to vary significantly, they all showed similar trends under different pressures. This variation is likely due to discrepancies in the PCB manufacturing process and differences in the conductive pads of the sensors. When the sensor positions were swapped, the results were comparable. Since each sensor unit was individually calibrated, this did not affect the ability of the soft sensors to measure pressure. Therefore, integrating sensors into a fish-scale structure enables accurate muscle pressure measurements at various positions, validating the functionality and technology of our design concept.

Fig. 6
Pressure response of three sensor units
Fig. 6
Pressure response of three sensor units
Close modal

5.2.5 Expert Evaluation of K Space.

LLMs, including GPT-4, can generate hallucinations, such as incorrect or nonsensical information [25], making the knowledge within our K space unreliable. We have taken numerous measures to address these knowledge-based errors, including applying the “chain of thought” technique when designing operators, and incorporating concept and knowledge content into prompt instructions for GPT-4 to emulate (this involved providing examples of concepts and knowledge, requiring GPT-4 to generate outputs in a similar format. In fact, this implemented a few-shot learning approach, which can reduce the hallucination effect [59]). According to previous studies [56,60], these methods significantly reduce the frequency of hallucination. In order to verify these practices and to ensure that our method did not mislead the designers, we have performed additional expert evaluations.

In the case study, GPT-4 generated 70 pieces of structured knowledge and approximately 40 pieces of unstructured knowledge (in the K K operator QA mode, where the knowledge is embedded within the dialog as unstructured information). Two designers conducted thematic analysis and screening of these pieces of knowledge. They individually coded each piece, assigned semantic labels, and eliminated redundant or overly simple information based on the coding results. Ultimately, they identified 41 key pieces of knowledge for evaluation. All of these pieces of knowledge were presented in a title-content format, including the general description provided by GPT-4 as well as specific information involved (detailed contents can be found in the appendix). Additionally, the evaluation materials given to the experts encompassed backgrounds and user requirements related to our case studies, enabling them to review the validity of the knowledge in its specific context. Drawing inspiration from human evaluation methods often employed in hallucination assessment for generative question answering [61], we employed four PhDs in design or clinical medicine (three majoring in design and one in medicine) to evaluate the knowledge. Each evaluator received the aforementioned evaluation materials, and was asked to use a 5-point Likert scale in assessing whether each piece of knowledge was correct. The criteria for judging correctness included whether the knowledge contradicted facts (the most critical factor) and whether the information about design standards and technical principles was valid within the design context. Higher scores indicated greater validity; for example, knowledge scoring 1 point contained significant factual errors, while knowledge scoring 3 points might state objective facts but be irrelevant to the medical design context. The evaluation process lasted for about 60 min. To ensure the accuracy of the evaluation, the evaluators were permitted to reference external materials.

Fig. 7
Distribution chart of expert evaluation scores
Fig. 7
Distribution chart of expert evaluation scores
Close modal

The results of the evaluation are shown in Table 2 and Fig. 7. Overall, the knowledge obtained from the case study was mostly assessed as completely correct (5 points) or right (4 points), with only a small proportion of knowledge (each expert’s result was less than 15%, averaging 8.5%) being judged as completely wrong (1 point) or wrong (2 points), which may indicate an hallucination. The consistency of experts’ scoring results was also considered high (intra-class correlation coefficient = 0.596), thus, they were deemed fair. The evaluation confirmed that most of the knowledge provided by our method in the case study was correct and in line with the design requirement. Notably, the incorrect knowledge entries were often not completely fabricated or inconsistent with the facts, but were wrong in our design scenario. For example, K15(F) temperature control technology, GPT-4’s description includes “the application of miniature heating and cooling systems to achieve precise temperature regulation within the device.” This is feasible in many devices in the rehabilitation field, but it is hard to apply and implement in wearable devices. This could be due to GPT-4’s ability to understand the context but its inability to predict the application results of technology in the real-world. However, similar knowledge information can be easily identified. During the execution of the case study, the designer actively judged the irrationality of K15(F) and did not continue to perform operations on it. Therefore, we believe that this risk of hallucination is controllable.

Table 2

Results of the expert evaluation

EvaluatorAverage score of knowledgeProportion of hallucinationVariance
Evaluator 14.474.8%0.82
Evaluator 24.394.8%0.02
Evaluator 33.819.7%0.52
Evaluator 43.6114.7%0.68
Mean4.078.5%Null
EvaluatorAverage score of knowledgeProportion of hallucinationVariance
Evaluator 14.474.8%0.82
Evaluator 24.394.8%0.02
Evaluator 33.819.7%0.52
Evaluator 43.6114.7%0.68
Mean4.078.5%Null

Note: Knowledge scored 2 (incorrect) or 1 (severely incorrect) are considered to have experienced hallucination.

5.3 Discussions and Limitations.

Based on the aforementioned test data and the prototypes developed, we have validated the feasibility of the final concept in the field of rehabilitation. This section will discuss how our method specifically operates and the remaining limitations.

Utilizing the extensive knowledge base of GPT-4, our C–K method offers knowledge from various backgrounds. In the practical context of design within commercial settings [62], designers often find themselves grappling with interdisciplinary tasks due to their lack of related background knowledge. During the conceptual design process of our case, designers rely on GPT-4 for knowledge retrieval to compensate for their lack of domain knowledge. Meanwhile, our C–K method facilitated efficient concept generation and stimulation. For example, the fish-scale structure is crucial for our prototype to adapt to patient requirements without compromising sensors’ performance. When using our method, designers apply the K C operator based on knowledge related to appearance and functionality (flexible material and adaptive interface) to generate the concept C2, which includes the property of modular bionic design. During the QA process with GPT-4 about bionic design (K K operator), they associated this with the key information of the fish-scale structure. If this process relied on other methods, it would require repeated searches for relevant knowledge and more thinking steps.

In addition, our method provides an effective human–computer interaction model. Before using our method, the designers should learn the prompts and our operators. Once they mastered these, they only needed to supplement the prompts with natural language and execute the operators. We also observed the performance of human–computer interaction during the case study. The concept exploration took the two designers a total of 2 days. To implement the operators, they engaged in 32 dialogs with GPT-4, with an average generation speed of approximately 15 s per dialog, and the designers needed to input about 39 characters per prompt (standard Chinese characters). In interviews with the designers, they noted that our method could avoid the complexity of searching and filtering through search engines or knowledge bases. Additionally, the designers mentioned that the knowledge generated was structured and concise, eliminating the need to spend time memorizing different pieces of knowledge.

Compared to previous LLM-based tools, our method offers more comprehensive and effective support for design knowledge and concept generation. Existing tools often rely on specific conceptual design methods, such as the bio-inspired method, where studies have used LLMs to generate biological knowledge and assess design concepts based on this knowledge [22,63]. However, methods on generating bio-inspired concepts using existing biological knowledge through LLMs are relatively scarce. Other studies have enhanced the TRIZ problem-solving methodology by leveraging LLMs [47], successfully using structured knowledge to generate concepts but without providing a reliable method for acquiring knowledge. In contrast, the C–K theory offers a universal framework for the co-evolution of concepts and knowledge. We have structured the definitions within the C–K theory, enhancing the relationship between concepts and knowledge, and utilized LLMs to support flexible knowledge retrieval and concept generation, providing a valuable complement and reference for similar studies. On the other hand, existing tools commonly utilized structured forms of knowledge (e.g., the FBS model [64]) and applied knowledge generation concepts, involving properties such as appearance and technology. This demonstrates that our method to supporting concept generation using LLMs, particularly the implementation of operators like C K and K C, has the potential to be generalized to related fields.

Based on the above discussion, we recommend applying our method in the design consulting industry, as well as in interdisciplinary product design contexts such as healthcare and mechanical engineering. By using our operators, designers can acquire knowledge and understand the specific backgrounds of different industries, thereby avoiding the need for extensive time investment in team training and reducing the initial costs of design projects. Additionally, design teams can manage knowledge and concepts based on our definitions of the C-space and K-space, quickly forming a structured design knowledge base to promote knowledge sharing and reuse.

However, we have identified several limitations in our method. First, LLMs exhibit some common issues in the concept design process. In our case study, GPT-4’s suggestions showed limited innovation when faced with complex tasks. For example, in discussions about machine learning algorithms, GPT-4 supplied comprehensive and precise details, actively presenting a wide array of specific algorithms and practical advice. However, when “micro sensor technology” became the topic, the diversity of the information supplied noticeably decreased, limiting to the principles of the technology and several examples. This could potentially be due to the fact that the training data of the model includes more content related to algorithms, and the subject of micro sensors is less frequently collected. Moreover, our expanded C–K theory also introduces some limitations. To avoid excessive complexity while accurately describing concepts, we use function, appearance, and technology to represent all properties. However, LLM-based tools need to consider the specific knowledge structures within particular contexts. For example, the bio-inspired design tools [63] adopt knowledge structures based on biological knowledge, biological characteristics, and application scenarios, which may perform better in related contexts. In future work, a key research focus will be on how to further mitigate the impact of LLMs’ performance and improve the quality of generated knowledge and concepts.

6 Conclusions and Future Works

This study introduces a method for conceptual design based on the C–K theory and LLMs, which aids designers in retrieving knowledge and exploring potential concepts. The method redefines the C space and the K space in the C–K theory, with the properties of concept C divided into function, appearance, and technology, requiring knowledge to correspond to these properties, thereby facilitating a structured connection between concepts and knowledge. Based on this, our method realizes the four operators of the C–K theory, including knowledge retrieval and reasoning operators (C K and K K), as well as concept generation, evaluation, and refinement operators (C C and K C). In the case study, the effectiveness of the aforementioned method and the quality of provided knowledge have been demonstrated.

Despite this, our method still has some limitations, such as the hallucination effect of LLMs leading to errors in knowledge and the generation of biased knowledge. Based on this, future research could focus on how to use the C–K theory to provide more comprehensive and in-depth professional knowledge. In fact, the LLMs we utilize have the potential to combine with knowledge engineering methods like knowledge graph, realizing LLMs-based graph construction, embedding, and retrieval [65], thereby increasing the efficiency of domain knowledge graph construction. This method perhaps can achieve better balance between the credibility of knowledge and cost of knowledge retrieval. Another possibility is to use technical methods to monitor the quality of generated knowledge, such as enabling LLMs to self-supervise [66], or designing low-cost expert intervention approaches to remind designers of potential hallucination effects and knowledge biases.

Acknowledgment

This work was supported by the Zhejiang Provincial Medical and Health Science and Technology Project under Grant No. 2022KY491.

Conflict of Interest

There are no conflicts of interest.

Data Availability Statement

The authors attest that all data for this study are included in the paper.

Appendix A: A Complete Prompt Template for Operators

See the complete prompts in Tables 3 and 4.

Table 3

The complete prompt templates used to conduct operators: C0C, CK, KK

OperatorPrompt template
C0CDesign Requirements: [C0]; Design Concept: Function-, Appearance-, Technology-
Please refer to the provided form of the concept and create a design concept that meets the design requirements, ensuring the innovation of the concept.
CKDesign Requirements: [C0]; Design Concept: [Cn]
Please provide relevant knowledge from the aspects of function, appearance, and technology based on the above design requirements and concepts.
KKKnowledge: [Kni{F}]
(I don’t understand the knowledge above. Please explain it to me in understandable language.)
Based on the above knowledge regarding function, please deduce and reason out new knowledge in the aspects of mapping, visibility, feedback, constraints, and consistency by following the steps below:
How can the mapping relationship between operations and outcomes be achieved in functionality?
To support effective mapping, how should visibility be demonstrated in interface design to ensure operations are predictable and easy to understand?
Given the enhanced visibility of operations, how should the feedback mechanism be designed to ensure users have immediate awareness of their operation results?
To reduce operational errors and improve interface security, how should constraints be implemented in the design?
To maintain operational consistency, how can the design ensure that similar operations yield similar results on the interface?
The explanations for the five aspects are as follows:
Mapping: The potential relationship between the controls and their effects.
Visibility: The design should clearly indicate the possible operations.
Feedback: The necessary feedback that users may require from the product.
Constraints: Limiting the potential for errors through design restrictions.
Consistency: Similar operations should produce similar results.
Knowledge: [Kni{A}]
(I do not understand the knowledge above. Please explain it to me in understandable language.)
Based on the above knowledge about appearance, please follow these steps to deduce and reason out new knowledge in the aspects of aesthetic appearance, component layout, connection methods, and material selection by following the steps below:
What aesthetic standards does this appearance conform to?
How should components be arranged to achieve this form?
How should components be connected to achieve the desired layout?
What materials are suitable to meet the above aesthetic requirements?
The explanations for the four aspects are as follows:
Aesthetic Standards: The potential design forms and aesthetic standards in the relevant field.
Component Layout: How components should be arranged within the product or structure to achieve optimal performance and user experience.
Connection Methods: How components are connected, such as through welding, threaded connections, adhesives, etc.
Material Selection: Choosing appropriate materials to provide the necessary structural strength, weight ratio, and durability.
OperatorPrompt template
C0CDesign Requirements: [C0]; Design Concept: Function-, Appearance-, Technology-
Please refer to the provided form of the concept and create a design concept that meets the design requirements, ensuring the innovation of the concept.
CKDesign Requirements: [C0]; Design Concept: [Cn]
Please provide relevant knowledge from the aspects of function, appearance, and technology based on the above design requirements and concepts.
KKKnowledge: [Kni{F}]
(I don’t understand the knowledge above. Please explain it to me in understandable language.)
Based on the above knowledge regarding function, please deduce and reason out new knowledge in the aspects of mapping, visibility, feedback, constraints, and consistency by following the steps below:
How can the mapping relationship between operations and outcomes be achieved in functionality?
To support effective mapping, how should visibility be demonstrated in interface design to ensure operations are predictable and easy to understand?
Given the enhanced visibility of operations, how should the feedback mechanism be designed to ensure users have immediate awareness of their operation results?
To reduce operational errors and improve interface security, how should constraints be implemented in the design?
To maintain operational consistency, how can the design ensure that similar operations yield similar results on the interface?
The explanations for the five aspects are as follows:
Mapping: The potential relationship between the controls and their effects.
Visibility: The design should clearly indicate the possible operations.
Feedback: The necessary feedback that users may require from the product.
Constraints: Limiting the potential for errors through design restrictions.
Consistency: Similar operations should produce similar results.
Knowledge: [Kni{A}]
(I do not understand the knowledge above. Please explain it to me in understandable language.)
Based on the above knowledge about appearance, please follow these steps to deduce and reason out new knowledge in the aspects of aesthetic appearance, component layout, connection methods, and material selection by following the steps below:
What aesthetic standards does this appearance conform to?
How should components be arranged to achieve this form?
How should components be connected to achieve the desired layout?
What materials are suitable to meet the above aesthetic requirements?
The explanations for the four aspects are as follows:
Aesthetic Standards: The potential design forms and aesthetic standards in the relevant field.
Component Layout: How components should be arranged within the product or structure to achieve optimal performance and user experience.
Connection Methods: How components are connected, such as through welding, threaded connections, adhesives, etc.
Material Selection: Choosing appropriate materials to provide the necessary structural strength, weight ratio, and durability.
Table 4

The complete prompt templates used to conduct operators: KK, KC, CC

OperatorPrompt template
KKKnowledge: [Kni{T}]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge about technology, please follow these steps to deduce and reason out new knowledge in the aspects of engineering principles, technical specifications, and production processes by following the steps below:
What are the engineering principles of this technology?
What technical specifications are required to achieve these principles?
What production processes are needed to manufacture according to these technical specifications?
The explanations for the three aspects are as follows:
Engineering Principles: The physical and chemical principles involved in the product.
Technical Specifications: The relevant technical standards and performance requirements, such as dimensional accuracy, durability, and power requirements for electronic products.
Production Processes: The methods and steps required to manufacture the designed product, including material processing, assembly techniques, and quality control.
KCKnowledge: [Kni]; Design Concept: Function-, Appearance-, Technology-
Please use the knowledge provided above to inspire an innovative and feasible design concept following the provided form.
CCDesign Concept: [Cn].
Evaluation Criteria: (1)Requirement Satisfaction: Sub-criteria listed as [Requirement1], [Requirement2] [Requirementn] (2)Knowledge Satisfaction: Sub-criteria listed as [K{F}], [K{A}], [K{T}]. Scoring Method: 4-point scale as follows: 2: The design concept performs poorly on this criterion, failing to meet the requirements and requiring significant improvement. 1: The design concept performs moderately on this criterion, partially meeting the requirements but with room for improvement. +1: The design concept performs well on this criterion, largely meeting the requirements. +2: The design concept performs excellently on this criterion, fully meeting or exceeding expectations.
Please evaluate the above design concept based on the evaluation criteria, using the scoring method, and present the results in a tabular format.
Please provide improvement directions for the aspects in the evaluation table that have received negative scores (1, 2), and then propose an improved design concept in the following format. “Design Concept: Function-, Appearance-, Technology-”
OperatorPrompt template
KKKnowledge: [Kni{T}]
(I do not understand the knowledge mentioned above. Please explain it to me in plain and understandable language.)
Based on the above knowledge about technology, please follow these steps to deduce and reason out new knowledge in the aspects of engineering principles, technical specifications, and production processes by following the steps below:
What are the engineering principles of this technology?
What technical specifications are required to achieve these principles?
What production processes are needed to manufacture according to these technical specifications?
The explanations for the three aspects are as follows:
Engineering Principles: The physical and chemical principles involved in the product.
Technical Specifications: The relevant technical standards and performance requirements, such as dimensional accuracy, durability, and power requirements for electronic products.
Production Processes: The methods and steps required to manufacture the designed product, including material processing, assembly techniques, and quality control.
KCKnowledge: [Kni]; Design Concept: Function-, Appearance-, Technology-
Please use the knowledge provided above to inspire an innovative and feasible design concept following the provided form.
CCDesign Concept: [Cn].
Evaluation Criteria: (1)Requirement Satisfaction: Sub-criteria listed as [Requirement1], [Requirement2] [Requirementn] (2)Knowledge Satisfaction: Sub-criteria listed as [K{F}], [K{A}], [K{T}]. Scoring Method: 4-point scale as follows: 2: The design concept performs poorly on this criterion, failing to meet the requirements and requiring significant improvement. 1: The design concept performs moderately on this criterion, partially meeting the requirements but with room for improvement. +1: The design concept performs well on this criterion, largely meeting the requirements. +2: The design concept performs excellently on this criterion, fully meeting or exceeding expectations.
Please evaluate the above design concept based on the evaluation criteria, using the scoring method, and present the results in a tabular format.
Please provide improvement directions for the aspects in the evaluation table that have received negative scores (1, 2), and then propose an improved design concept in the following format. “Design Concept: Function-, Appearance-, Technology-”

Appendix B: A Complete List of Knowledge Items in the Case Study

All knowledge below is generated by GPT-4.

K11(F): biomechanics and ergonomics: understanding the biomechanical principles of hand movements, including joint range of motion, muscle activation patterns, etc., to ensure designs can adapt to various activities.

K12(F): Flexible electronics materials: exploring bendable, stretchable electronic materials and sensors to allow devices to conform to the skin and maintain performance stability during various movements.

K13(F): data analysis and machine learning: utilizing data analysis techniques and machine learning algorithms to analyze collected health data, providing accurate feedback on rehabilitation progress and personalized suggestions.

K14(F): user interface design: developing intuitive user interfaces to enable patients and medical professionals to easily understand feedback data and progress reports.

K15(F): temperature control technology: researching and applying micro heating and cooling systems to achieve precise temperature regulation within devices.

K16(F): pain management and rehabilitation medicine: understanding the physiological mechanisms of heat therapy and cold therapy in relieving pain, reducing swelling, and promoting recovery.

K17(A): modular design principles: studying modular design methods to achieve highly customizable and maintainable product structures.

K18(A): rapid disassembly and connection technology: exploring user-friendly rapid disassembly and connection mechanisms to ensure modules are easy to replace and clean.

K19(A): human–computer interaction: understanding how to provide effective visual feedback through interactive LED indicators to enhance user experience.

K1(10)(A): LED technology: researching compact and energy-efficient LED solutions to achieve compact design and long-term operation of devices.

K1(11)(T): microsensor technology: in-depth understanding of microsensors for monitoring pressure and hand movements, including their working principles, data accuracy, and durability.

K1(12)(T): flexible PCB design: developing flexible PCBs for wearable devices to ensure device reliability and comfort.

K1(13)(T): wireless charging technology: researching wireless charging solutions suitable for wearable devices, considering their charging efficiency and compatibility.

K1(14)(T): low-power wireless communication: utilizing low-power bluetooth (BLE) or other wireless technologies to achieve real-time synchronization of device data and remote access.

K121(F): visibility—transparent usage indication: devices made with flexible electronic materials should be designed with intuitive visual or tactile markings indicating how to properly wear and operate the device. For example, color coding or texture differences can indicate different wearing positions or operating interfaces.

K122(F): visibility—integrated display: utilizing flexible display technology, the device surface can display simple icons or indicators, providing immediate operational feedback and guidance.

K123(F): feedback—multimodal feedback mechanism: combining tactile, visual, and auditory feedback, leveraging the characteristics of flexible materials to provide users with immediate responses to their actions. For example, the device can provide feedback through slight vibrations, sound signals, or color changes, indicating successful operations or monitored health parameters.

K124(F): feedback—real-time health monitoring feedback: leveraging its sensing capabilities, the device provides real-time feedback on user health status and rehabilitation progress, such as muscle activity levels or the accuracy of rehabilitation exercises.

K126(F): mapping—close mapping of physical actions to digital responses: through precise sensing technology, directly mapping users’ physical actions (such as hand movements) to digital feedback in the device enhances user experience and operational intuitiveness.

K127(F): constraint—intelligent error prevention: by design and programming, limiting the possibility of erroneous operations. For example, the device can identify improper wearing or operation and guide users to correct actions through feedback mechanisms such as warning sounds or prompt messages.

K128(F): constraint—adaptive interface: the device interface can automatically adjust based on user habits and frequency, hiding less-used functions to reduce operational complexity and the likelihood of errors.

K129(F): consistency—standardization of interface and operations: ensuring similar operations in the device produce similar feedback, both within different parts of the device and across different devices. This increases user learning efficiency and reduces confusion.

K1210(F): consistency—cross-device consistency: if this rehabilitation device is part of a product series, its operational logic, interface design, and feedback methods should remain consistent with other products to facilitate users’ quick adaptation to new devices.

K181(A): component layout—modular design: through modular design, each component or module can independently perform specific functions. This not only facilitates rapid disassembly and assembly but also allows for product customization according to user needs. Design should ensure uniform interfaces between modules to achieve seamless connections.

K182(A): component layout—biomimetic layout: referring to structures in nature such as honeycombs, tree branches, etc., to optimize component layout for the best weight-to-strength ratio. This biomimetic layout can improve product structural stability while reducing weight.

K183(A): connection methods—magnetic connection: using magnetism to achieve rapid connection and disconnection of components. This method is not only convenient for user operation but also ensures accurate alignment, reducing assembly errors.

K184(A): connection methods—sliding rails and latch mechanisms: designing sliding rails and latch mechanisms to achieve simple yet secure component connections. This method allows users to install and remove components through simple push-pull actions without requiring tools.

K185(A): material selection—composite materials: choosing lightweight composite materials such as carbon fiber or glass fiber-reinforced plastics to provide high strength and durability while maintaining product lightness. These materials also offer good corrosion resistance and abrasion resistance, suitable for frequent disassembly and cleaning requirements.

K186(A): material selection—flexible materials: for parts that need to conform to the human body, flexible materials like silicone can be chosen to enhance wearing comfort. These materials should be durable and easy to clean.

K187(A): esthetics and appearance—sleek modern design: adopting a sleek and modern design language, emphasizing smooth curves and simple color schemes. This design is not only esthetically pleasing but also facilitates cleaning and maintenance.

K1111(T): manufacturing processes—precision manufacturing and micromachining technology: the production of microsensors and flexible circuit boards involves high-precision manufacturing technologies, including photolithography, etching, and printed circuit technology. These processes can produce extremely small electronic components and wires, requiring high material processing accuracy and consistency.

K1112(T): manufacturing processes—special treatment of flexible materials: considering the special nature of flexible PCBs, assembly and connection technologies suitable for flexible substrates need to be developed, such as using flexible adhesives and low-temperature soldering processes to avoid material damage.

K1113(T): manufacturing processes—quality control and testing: to ensure the reliability and durability of microsensors and flexible PCBs, strict quality control measures, including but not limited to electrical testing, mechanical testing, and environmental adaptability testing, need to be implemented during the production process.

K1114(T): technical specifications—dimensional accuracy and flexibility requirements: the design of microsensors and flexible PCBs needs to meet strict dimensional accuracy requirements to ensure correct installation and functional realization within compact spaces. Meanwhile, material and structural design should ensure sufficient flexibility to adapt to different wearing and usage environments.

K1115(T): technical specifications—durability and protection ratings: products should be designed to be wear-resistant, waterproof, and dustproof to adapt to long-term wearing and various usage conditions. This requires the selection of durable materials and the adoption of appropriate encapsulation technologies to improve product protection ratings.

K1116(T): technical specifications—power and energy efficiency requirements: to ensure long-term use of wearable devices, the design of microsensors and flexible PCBs needs to optimize power consumption and improve energy efficiency. This may involve the use of low-power electronic components and optimizing circuit designs to reduce energy consumption.

K1117(T): engineering principles—physical principles and biocompatibility: the design of microsensors needs to be based on the physical properties of their monitoring objects, such as pressure, temperature, or conductivity, while considering biocompatibility to ensure prolonged contact with the skin does not cause discomfort or allergic reactions.

K1118(T): engineering principles—mechanical stability of flexible circuits: when designing flexible PCBs, consideration needs to be given to their mechanical stability and electrical performance after multiple bending or stretching cycles. This involves principles of materials science and mechanical engineering, such as using reinforced composite materials to improve durability.

K1119(T): engineering principles—thermal management: microsensors and electronic components generate heat during operation. Therefore, the design of flexible PCBs needs to consider effective thermal management solutions, such as using materials with good thermal conductivity and designing structures conducive to heat dissipation to prevent device overheating and ensure safety and comfort.

References

1.
Chou
,
J.-R.
,
2014
, “
An Ideation Method for Generating New Product Ideas Using Triz, Concept Mapping, and Fuzzy Linguistic Evaluation Techniques
,”
Adv. Eng. Inform.
,
28
(
4
), pp.
441
454
.
2.
Liu
,
X.
,
Li
,
Y.
,
Pan
,
P.
, and
Li
,
W.
,
2011
, “
Research on Computer-Aided Creative Design Platform Based on Creativity Model
,”
Expert Syst. Appl.
,
38
(
8
), pp.
9973
9990
.
3.
Siddharth
,
L.
, and
Chakrabarti
,
A.
,
2018
, “
Evaluating the Impact of Idea-Inspire 4.0 on Analogical Transfer of Concepts
,”
AI Edam
,
32
(
4
), pp.
431
448
.
4.
Srinivasan
,
V.
,
Song
,
B.
,
Luo
,
J.
,
Subburaj
,
K.
,
Elara
,
M. R.
,
Blessing
,
L.
, and
Wood
,
K.
,
2018
, “
Does Analogical Distance Affect Performance of Ideation?
,”
ASME J. Mech. Des.
,
140
(
7
), p.
071101
.
5.
Luo
,
J.
,
Sarica
,
S.
, and
Wood
,
K. L.
,
2021
, “
Guiding Data-Driven Design Ideation by Knowledge Distance
,”
Knowl. Based Syst.
,
218
, p.
106873
.
6.
Hao
,
J.
,
Yan
,
Y.
,
Gong
,
L.
,
Wang
,
G.
, and
Lin
,
J.
,
2014
, “
Knowledge Map-Based Method for Domain Knowledge Browsing
,”
Decision Support Syst.
,
61
, pp.
106
114
.
7.
Sarica
,
S.
,
Han
,
J.
, and
Luo
,
J.
,
2023
, “
Design Representation as Semantic Networks
,”
Comput. Ind.
,
144
, p.
103791
.
8.
Chen
,
L.
,
Wang
,
P.
,
Dong
,
H.
,
Shi
,
F.
,
Han
,
J.
,
Guo
,
Y.
,
Childs
,
P. R.
,
Xiao
,
J.
, and
Wu
,
C.
,
2019
, “
An Artificial Intelligence Based Data-Driven Approach for Design Ideation
,”
J. Visual Commun. Image Represent.
,
61
, pp.
10
22
.
9.
He
,
Y.
,
Camburn
,
B.
,
Liu
,
H.
,
Luo
,
J.
,
Yang
,
M.
, and
Wood
,
K.
,
2019
, “
Mining and Representing the Concept Space of Existing Ideas for Directed Ideation
,”
ASME J. Mech. Des.
,
141
(
12
), p.
121101
.
10.
Shi
,
F.
,
Chen
,
L.
,
Han
,
J.
, and
Childs
,
P.
,
2017
, “
A Data-Driven Text Mining and Semantic Network Analysis for Design Information Retrieval
,”
ASME J. Mech. Des.
,
139
(
11
), p.
111402
.
11.
Chen
,
L.
,
Cai
,
Z.
,
Jiang
,
Z.
,
Long
,
Q.
,
Sun
,
L.
,
Childs
,
P.
, and
Zuo
,
H.
,
2023
, “
A Knowledge-Based Ideation Approach for Bio-inspired Design
,”
Proc. Des. Soc.
,
3
, pp.
231
240
.
12.
Huet
,
A.
,
Pinquié
,
R.
,
Véron
,
P.
,
Mallet
,
A.
, and
Segonds
,
F.
,
2021
, “
Cacda: A Knowledge Graph for a Context-Aware Cognitive Design Assistant
,”
Comput. Ind.
,
125
, p.
103377
.
13.
Jia
,
J.
,
Zhang
,
Y.
, and
Saad
,
M.
,
2022
, “
An Approach to Capturing and Reusing Tacit Design Knowledge Using Relational Learning for Knowledge Graphs
,”
Adv. Eng. Inform.
,
51
, p.
101505
.
14.
Li
,
X.
,
Chen
,
C.-H.
,
Zheng
,
P.
,
Wang
,
Z.
,
Jiang
,
Z.
, and
Jiang
,
Z.
,
2020
, “
A Knowledge Graph-Aided Concept–Knowledge Approach for Evolutionary Smart Product–Service System Development
,”
ASME J. Mech. Des.
,
142
(
10
), p.
101403
.
15.
Hatchuel
,
A.
, and
Weil
,
B.
,
2009
, “
CK Design Theory: An Advanced Formulation
,”
Res. Eng. Des.
,
19
(
4
), pp.
181
192
.
16.
Hatchuel
,
A.
,
Le Masson
,
P.
, and
Weil
,
B.
,
2011
, “
Teaching Innovative Design Reasoning: How Concept–Knowledge Theory Can Help Overcome Fixation Effects
,”
Artif. Intell. Eng. Des. Anal. Manuf.
,
25
(
1
), pp.
77
92
.
17.
Kazakci
,
A. O.
,
2005
, “
Designar: An Intelligent Design Assistant Based on CK Design Theory
,”
ASME 2005 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference
,
Long Beach, CA
,
Sept. 24–25
, pp.
7
15
.
18.
Gillier
,
T.
,
Piat
,
G.
,
Roussel
,
B.
, and
Truchot
,
P.
,
2010
, “
Managing Innovation Fields in a Cross-Industry Exploratory Partnership With C–K Design Theory
,”
J. Product Innov. Manage.
,
27
(
6
), pp.
883
896
.
19.
Weisbrod
,
G.
, and
Kroll
,
E.
,
2018
, “
Idea-Configuration-Evaluation (ICE): Development and Demonstration of a New Prescriptive Model of the Conceptual Engineering Design Process Based on Parameter Analysis and C–K Theory
,”
Res. Eng. Des.
,
29
(
2
), pp.
203
225
.
20.
Achiam
,
J.
,
Adler
,
S.
,
Agarwal
,
S.
,
Ahmad
,
L.
,
Akkaya
,
I.
,
Aleman
,
F. L.
,
Almeida
,
D.
, et al.,
2023
, “Gpt-4 Technical Report,” preprint arXiv:2303.08774.
21.
Zhu
,
Q.
, and
Luo
,
J.
,
2023
, “
Generative Transformers for Design Concept Generation
,”
ASME J. Comput. Inf. Sci. Eng.
,
23
(
4
), p.
041003
.
22.
Zhu
,
Q.
,
Zhang
,
X.
, and
Luo
,
J.
,
2023
, “
Biologically Inspired Design Concept Generation Using Generative Pre-trained Transformers
,”
ASME J. Mech. Des.
,
145
(
4
), p.
041409
.
23.
Honovich
,
O.
,
Shaham
,
U.
,
Bowman
,
S. R.
, and
Levy
,
O.
,
2023
, “
Instruction Induction: From Few Examples to Natural Language Task Descriptions
,” 6
1st Annual Meeting of the Association for Computational Linguistics
,
Toronto, Canada
,
July 9–14
, pp.
1935
1952
.
24.
Liu
,
J.
,
Shen
,
D.
,
Zhang
,
Y.
,
Dolan
,
W. B.
,
Carin
,
L.
, and
Chen
,
W.
,
2022
, “
What Makes Good In-Context Examples for GPT-3?
Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures
,
Dublin, Ireland and Online
, pp.
100
114
.
25.
Maynez
,
J.
,
Narayan
,
S.
,
Bohnet
,
B.
, and
McDonald
,
R.
,
2020
, “
On Faithfulness and Factuality in Abstractive Summarization
,”
Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics
,
Online
,
July
, pp.
1906
1919
.
26.
Anderson
,
B. R.
,
Shah
,
J. H.
, and
Kreminski
,
M.
,
2024
, “
Homogenization Effects of Large Language Models on Human Creative Ideation
,”
Proceedings of the 16th Conference on Creativity & Cognition
,
Chicago, IL
,
June 23
, pp.
413
425
.
27.
French
,
M. J.
,
Gravdahl
,
J.
, and
French
,
M.
,
1985
,
Conceptual Design for Engineers
,
Springer
,
Heidelberg
, pp.
1
12
.
28.
Dietrich
,
A.
,
2004
, “
The Cognitive Neuroscience of Creativity
,”
Psychon. Bull. Rev.
,
11
(
12
), pp.
1011
1026
.
29.
Daly
,
S. R.
,
Seifert
,
C. M.
,
Yilmaz
,
S.
, and
Gonzalez
,
R.
,
2016
, “
Comparing Ideation Techniques for Beginning Designers
,”
ASME J. Mech. Des.
,
138
(
10
), p.
101108
.
30.
Altshuller
,
G.
,
Shulyak
,
L.
, and
Rodman
,
S.
,
1999
,
The Innovation Algorithm: TRIZ, Systematic Innovation and Technical Creativity
,
Technical Innovation Center
,
Worcester, MA
.
31.
Nagamachi
,
M.
,
1995
, “
Kansei Engineering: A New Ergonomic Consumer-Oriented Technology for Product Development
,”
Int. J. Ind. Ergon.
,
15
(
1
), pp.
3
11
.
32.
Nagamachi
,
M.
,
2002
, “
Kansei Engineering as a Powerful Consumer-Oriented Technology for Product Development
,”
Appl. Ergon.
,
33
(
3
), pp.
289
294
.
33.
Gero
,
J. S.
, and
Kannengiesser
,
U.
,
2014
,
The Function-Behaviour-Structure Ontology of Design
,
Springer
,
London
, pp.
263
283
.
34.
Hatchuel
,
A.
, and
Weil
,
B.
,
2003
, “
A New Approach of Innovative Design: An Introduction to CK Theory
,” DS 31: Proceedings of ICED 03: The 14th International Conference on Engineering Design, Stockholm.
35.
Hatchuel
,
A.
, and
Weil
,
B.
,
2009
, “
CK Design Theory: An Advanced Formulation
,”
Res. Eng. Des.
,
19
(
4
), pp.
181
192
.
36.
Kroll
,
E.
,
2013
, “
Design Theory and Conceptual Design: Contrasting Functional Decomposition and Morphology With Parameter Analysis
,”
Res. Eng. Des.
,
24
, pp.
165
183
.
37.
Song
,
B.
,
Luo
,
J.
, and
Wood
,
K.
,
2019
, “
Data-Driven Platform Design: Patent Data and Function Network Analysis
,”
ASME J. Mech. Des.
,
141
(
2
), p.
021101
.
38.
Sarica
,
S.
,
Luo
,
J.
, and
Wood
,
K. L.
,
2020
, “
Technet: Technology Semantic Network Based on Patent Data
,”
Expert Syst. Appl.
,
142
, p.
112995
.
39.
Huang
,
Z.
,
Guo
,
X.
,
Liu
,
Y.
,
Zhao
,
W.
, and
Zhang
,
K.
,
2023
, “
A Smart Conflict Resolution Model Using Multi-layer Knowledge Graph for Conceptual Design
,”
Adv. Eng. Inform.
,
55
, p.
101887
.
40.
Cong
,
X.
,
Xiang
,
Z.
,
Li
,
X.
,
Dong
,
L.
,
Li
,
J.
, and
Chen
,
J.
,
2023
, “
A Knowledge Graph-Based Interaction Design Method for Smart Product-Service System Development
,”
J. Mech. Eng.
,
59
(
11
), pp.
94
105
.
41.
Vaswani
,
A.
,
Shazeer
,
N.
,
Parmar
,
N.
,
Uszkoreit
,
J.
,
Jones
,
L.
,
Gomez
,
A. N.
,
Kaiser
,
L. u.
, and
Polosukhin
,
I.
,
2017
, “
Attention is all you need
,”
Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17
,
Long Beach, CA
,
Dec. 4–9
, Curran Associates Inc., pp.
6000
6010
.
42.
Chowdhery
,
A.
,
Narang
,
S.
,
Devlin
,
J.
,
Bosma
,
M.
,
Mishra
,
G.
,
Roberts
,
A.
, and
Barham
,
P.
,
2023
, “
Palm: Scaling Language Modeling With Pathways
,”
J. Mach. Learn. Res.
,
24
(
240
), pp.
1
113
.
43.
Wu
,
T.
,
He
,
S.
,
Liu
,
J.
,
Sun
,
S.
,
Liu
,
K.
,
Han
,
Q.-L.
, and
Tang
,
Y.
,
2023
, “
A Brief Overview of Chatgpt: The History, Status Quo and Potential Future Development
,”
IEEE/CAA J. Autom. Sin.
,
10
(
5
), pp.
1122
1136
.
44.
Han
,
Y.
, and
Moghaddam
,
M.
,
2021
, “
Eliciting Attribute-Level User Needs From Online Reviews With Deep Language Models and Information Extraction
,”
ASME J. Mech. Des.
,
143
(
6
), p.
061403
.
45.
Qiu
,
Y.
, and
Jin
,
Y.
,
2023
, “
Document Understanding-Based Design Support: Application of Language Model for Design Knowledge Extraction
,”
ASME J. Mech. Des.
,
145
(
12
), p.
121401
.
46.
Wang
,
B.
,
Zuo
,
H.
,
Cai
,
Z.
,
Yin
,
Y.
,
Childs
,
P.
,
Sun
,
L.
, and
Chen
,
L.
,
2023
, “
A Task-Decomposed AI-Aided Approach for Generative Conceptual Design
,”
International Design Engineering Technical Conferences and Computers and Information in Engineering Conference
,
Boston, MA
,
Aug. 20–23
, Volume 87349, p.
V006T06A009
.
47.
Jiang
,
S.
, and
Luo
,
J.
,
2024
, “Autotriz: Artificial Ideation With Triz and Large Language Models,” preprint arXiv:2403.13002.
48.
Zhu
,
Q.
, and
Luo
,
J.
,
2023
, “
Generative Transformers for Design Concept Generation
,”
ASME J. Comput. Inf. Sci. Eng.
,
23
(
4
), p.
041003
.
49.
Pahl
,
G.
,
Beitz
,
W.
,
Feldhusen
,
J.
, and
Grote
,
K.-H.
,
1996
,
Engineering Design: A Systematic Approach
,
Springer
,
London
.
50.
Hatchuel
,
A.
,
Le Masson
,
P.
, and
Weil
,
B.
,
2004
, “
CK Theory in Practice: Lessons from Industrial Applications
,”
DS 32: Proceedings of DESIGN 2004, the 8th International Design Conference
,
Dubrovnik, Croatia
,
May 18
, pp.
245
258
.
51.
Chang
,
Y.
,
Wang
,
X.
,
Wang
,
J.
,
Wu
,
Y.
,
Yang
,
L.
,
Zhu
,
K.
, and
Chen
,
H.
,
2024
, “
A Survey on Evaluation of Large Language Models
,”
ACM Trans. Intell. Syst. Technol.
,
15
(
3
), pp.
1
45
.
52.
Dai
,
D.
,
Sun
,
Y.
,
Dong
,
L.
,
Hao
,
Y.
,
Ma
,
S.
,
Sui
,
Z.
, and
Wei
,
F.
,
2023
, “
Why Can GPT Learn In-Context Language? Models Secretly Perform Gradient Descent as Meta-optimizers
,”
Findings of the Association for Computational Linguistics: ACL 2023
,
Toronto, Canada
,
July
, pp.
4005
4019
.
53.
Kojima
,
T.
,
Gu
,
S. S.
,
Reid
,
M.
,
Matsuo
,
Y.
, and
Iwasawa
,
Y.
,
2022
, “
Large Language Models are Zero-Shot Reasoners
,”
Adv. Neural Inform. Process. Syst.
,
35
, pp.
22199
22213
.
54.
Ulrich
,
K. T.
, and
Eppinger
,
S. D.
,
2016
,
Product Design and Development
,
McGraw-Hill
,
New York
.
55.
Norman
,
D. A.
,
2002
,
The Psychology of Everyday Things
,
MIT Press
,
Cambridge, MA
.
56.
Wei
,
J.
,
Wang
,
X.
,
Schuurmans
,
D.
,
Bosma
,
M.
,
Chi
,
E.
,
Le
,
Q. V.
, and
Zhou
,
D.
,
2022
,
Advances in Neural Information Processing Systems
,
S.
Koyejo
,
S.
Mohamed
,
A.
Agarwal
,
D.
Belgrave
,
K.
Cho
, and
A.
Oh
, eds.,
MIT Press
,
Cambridge, MA
, pp.
24824
24837
.
57.
Rugman
,
A. M.
, and
D’Cruz
,
J. R.
,
1993
, “
The “Double Diamond” Model of International Competitiveness: The Canadian Experience
,”
Manage. Int. Rev.
,
33
, pp.
17
39
. https://www.jstor.org/stable/40228188
58.
Van Boeijen
,
A.
,
Daalhuizen
,
J.
,
Van Der Schoor
,
R.
, and
Zijlstra
,
J.
,
2014
,
Delft Design Guide: Design Strategies and Methods
,
Delft University of Technology
.
59.
Tonmoy
,
S.
,
Zaman
,
S.
,
Jain
,
V.
,
Rani
,
A.
,
Rawte
,
V.
,
Chadha
,
A.
, and
Das
,
A.
,
2024
, “A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models,” preprint arXiv:2401.01313.
60.
Song
,
C. H.
,
Wu
,
J.
,
Washington
,
C.
,
Sadler
,
B. M.
,
Chao
,
W.-L.
, and
Su
,
Y.
,
2023
, “
LLM-Planner: Few-Shot Grounded Planning for Embodied Agents With Large Language Models
,”
Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)
,
Paris, France
,
Oct. 2–6
, pp.
2998
3009
.
61.
Ji
,
Z.
,
Lee
,
N.
,
Frieske
,
R.
,
Yu
,
T.
,
Su
,
D.
,
Xu
,
Y.
,
Ishii
,
E.
,
Bang
,
Y. J.
,
Madotto
,
A.
, and
Fung
,
P.
,
2023
, “
Survey of Hallucination in Natural Language Generation
,”
ACM Comput. Surv.
,
55
(
12
), pp.
1
38
.
62.
Heiss
,
L.
, and
Kokshagina
,
O.
,
2021
, “
Tactile Co-design Tools for Complex Interdisciplinary Problem Exploration in Healthcare Settings
,”
Des. Stud.
,
75
, p.
101030
.
63.
Chen
,
L.
,
Jiang
,
Z.
,
Xia
,
D.
,
Cai
,
Z.
,
Sun
,
L.
,
Childs
,
P.
, and
Zuo
,
H.
,
2024
, “
BidTrainer: An LLMs-Driven Education Tool for Enhancing the Understanding and Reasoning in Bio-inspired Design
,”
Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems
,
Honolulu, HI
,
May 11
, pp.
1
20
.
64.
Chen
,
L.
,
Jing
,
Q.
,
Tsang
,
Y.
,
Wang
,
Q.
,
Sun
,
L.
, and
Luo
,
J.
,
2024
, “
DesignFusion: Integrating Generative Models for Conceptual Design Enrichment
,”
ASME J. Mech. Des.
,
146
(
11
), p.
111703
.
65.
Yu
,
S.
,
Huang
,
T.
,
Liu
,
M.
, and
Wang
,
Z.
,
2023
, “
Bear: Revolutionizing Service Domain Knowledge Graph Construction With LLM
,”
International Conference on Service-Oriented Computing
,
Rome, Italy
,
Nov. 28 – Dec. 1
, pp.
339
346
.
66.
Shinn
,
N.
,
Cassano
,
F.
,
Gopinath
,
A.
,
Narasimhan
,
K.
, and
Yao
,
S.
,
2023
, “Reflexion: Language Agents With Verbal Reinforcement Learning,”
Advances in Neural Information Processing Systems
, Vol.
36
,
A.
Oh
,
T.
Naumann
,
A.
Globerson
,
K.
Saenko
,
M.
Hardt
, and
S.
Levine
, eds.,
MIT Press
,
Cambridge, MA
, pp.
8634
8652
.