A cultural memory semiotics and function behavior structure model for digital inheritance and innovation in AI generated Huizhou woodcarving images

Abstract

As generative AI becomes embedded in cultural heritage visualization and design, it is crucial to understand how audiences evaluate AI-generated heritage images in terms of cultural continuity and innovation. This study investigates how cultural memory, design semiotics, and the Function–Behavior–Structure (FBS) framework jointly shapes perceived cultural heritage digital inheritance and innovation (CHDII), using AI-generated Huizhou woodcarving images as a case. We propose a Cultural Memory–Semiotics–FBS model in which material, functional, and symbolic memory dimensions influence semiotic sign types (icon, index, symbol), FBS constructs, and CHDII. Theory-driven modular prompts were used in Midjourney to generate a standardized set of 20 images, followed by a survey of 434 respondents. Structural equation modelling and multilayer perceptron neural networks were applied to test nine hypotheses and explore potential nonlinear effects. Results reveal two dominant pathways: FD–IN–F–B–S and SD–SY–S, while the material/iconic path is not significant. ANN analyses confirm the central roles of functional, behavioural, and symbolic variables. The findings position generative AI as a controllable medium for encoding cultural meaning and offer a reusable workflow for culturally informed digital heritage design.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-35360-5.

Introduction

Traditional visual arts and crafts are a vital component of the world’s cultural heritage. Beyond providing aesthetic pleasure, they support cultural learning, identity formation, and emotional bonding^1–3. Craft traditions such as wood carving, textiles, and metal ornamentation often encode local historical narratives, religious beliefs, and socio-cultural values into highly condensed visual forms and symbolic systems. These material and visual configurations not only constitute an important part of world cultural heritage, but also embody community identity and affective attachment^2–7. Recent studies increasingly recognize that the cultural, social, and psychological value of traditional visual arts depends largely on how their visual symbols and cultural memories are interpreted, re-encoded, and redesigned^8–10. Research in iconology and visual culture further shows that decorative motifs frequently layer moral norms, kinship structures, and regional identities on top of aesthetic form, thereby becoming key media for consolidating and transmitting collective memory^11,12.

The rapid development of digital media technologies, such as high-precision 3D modelling, digital twins, interactive displays, and generative artificial intelligence, has opened new possibilities for the re-presentation and redesign of traditional crafts. Designers can now reconfigure materials, textures, compositions, and narrative relationships within virtual environments^13–15. These advances demonstrate the considerable potential of digital technology to activate traditional visual symbols.

Despite this progress, existing practice still shows clear methodological limitations. Many digitization projects remain focused on “geometric recording” and “visual reproduction,” prioritizing data capture, fine-grained modelling, and surface texture reconstruction. In doing so, they tend to treat cultural heritage primarily as a three-dimensional object requiring high-fidelity restoration, rather than as a cultural medium that embodies complex symbolic and mnemonic mechanisms^16–18. With the rapid integration of generative AI into design workflows, increasing numbers of studies apply text-to-image models to the visual reconstruction of cultural heritage. However, most of these efforts rely on implicit, experience-based prompt adjustments and focus on producing visually appealing images. They rarely articulate a clear theoretical framework that systematically connects symbol extraction, structural mapping, and user evaluation. Even the emerging work that does address symbolic expression and digital design methodology often remains confined to technical descriptions of specific cases, making it difficult to generalize into transferable design–evaluation models^19–23. This reveals a core gap that has not been adequately addressed: most studies stay at the level of image reproduction and experience presentation, while lacking attempts to model cultural memory and symbolic structure within a clear theoretical framework and to translate such models into testable design pathways.

To address this gap, the present study adopts Huizhou wood carving from Anhui Province, China, as a representative case. Huizhou wood carving is widely found in architectural components, furniture, and everyday artifacts. Through techniques such as relief carving, openwork, and round carving, it depicts auspicious motifs, historical stories, and scenes from daily life^11,24–26. The motifs function not merely as decorative elements; they condense clan ethics, value systems, and local cosmologies, and thus possess strong narrative and symbolic qualities^27,28. We therefore treat Huizhou wood carving as a typical example of symbol-dense cultural heritage and integrate three complementary theoretical perspectives: cultural memory theory, design semiotics, and the Function–Behavior–Structure (FBS) framework. Cultural memory theory helps explain how traditional carvings store and activate collective memory through material, functional, and symbolic dimensions. Design semiotics provides an image-level analytic toolkit based on icons, indices, and symbols. The FBS framework, in turn, describes how design reasoning proceeds from intended functions, through behavioral representations, to specific structural forms. Taken together, these perspectives offer a systematic theoretical basis for explaining how traditional cultural symbols are re-encoded in digital images, how they are perceived by users, and how they contribute to the digital inheritance and innovation of cultural heritage.

Building on this integrated framework, the study addresses the following research questions:

RQ1

How are the three dimensions of cultural memory associated with the three types of design signs (iconic, indexical, symbolic) in participants’ evaluations of AI-generated images of Huizhou wood carving?

RQ2

How do different types of signs shape participants’ evaluations of function, behavior, and structure in AI-generated heritage design, and in turn relate to their perceptions of cultural heritage digital inheritance and innovation?

RQ3

Within the integrated Cultural Memory–Semiotics–FBS model, which dimensions and pathways most strongly predict CHDII when examined using a combined SEM–ANN approach?

This study makes three main contributions:

First, it proposes and empirically tests a falsifiable integrated model. We formalize a “Memory–Semiotics–FBS” digital heritage model that incorporates the three dimensions of cultural memory, the three types of design signs, and the three FBS components into a single structural equation framework, thereby capturing how these constructs jointly relate to the outcome variable of cultural heritage digital inheritance and innovation. Nine directional path hypotheses are explicitly specified and evaluated using standard errors and confidence intervals, providing a coherent theoretical structure for subsequent research.

Second, the study operationalizes abstract theoretical dimensions into a user-perceivable evaluation system. Drawing on the literature and expert interviews, we develop and validate a perception-oriented measurement scale that translates cultural memory, symbolic features, and FBS dimensions into measurable latent variables and items. This fills a gap in quantitative assessment for symbol-dense cultural heritage at the “design–user experience” interface and offers a generic tool for comparative studies across regions and craft types.

Third, the study establishes a reproducible workflow linking theory, prompts, AI images, and user evaluations. Using the integrated model as a blueprint, we explicitly encode each construct into structured prompts and employ a generative AI model to produce a standardized image set of Huizhou wood carving. In this workflow, AI is treated as a theory-constrained rendering engine rather than a black-box creative tool. Participants then evaluate the images under a unified visual context, and their responses are analyzed using structural equation modelling (SEM) in combination with artificial neural networks (ANN) to capture both linear and non-linear relationships. This closed loop—from theory to images to perception data—provides a methodological paradigm for future AIGC-based design experiments and A/B testing across different cultural contexts and heritage types.

To test the proposed model and hypotheses, we invited participants to view the standardized AI-generated image set of Huizhou wood carving derived from this framework and to complete a cross-sectional questionnaire assessing their perceptions of cultural memory, symbolic features, FBS dimensions, and CHDII. SEM was used to evaluate the directional positive associations among latent variables, while ANN was employed to examine variable importance and potential non-linear patterns. Given the cross-sectional design, all pathways in this study are interpreted as theoretically directional associations rather than strict causal relationships.

Literature review, research hypothesis and research model

Cultural memory theory

The concept of cultural memory was proposed by Jan Assmann, building on Maurice Halbwachs’s theory of social memory^29–32. Assmann distinguishes communicative memory, which is short-term and transmitted orally, from cultural memory, which is long-term and preserved through institutionalized forms such as writing, art, and rituals³². The notion of lieux de mémoire (“sites of memory”) was further developed by Pierre Nora and later elaborated by Aleida Assmann in their influential works^33,34. Within cultural memory theory, material objects, historical periods, and locations that fulfil social functions and embody symbolic meanings are regarded as sites of memory. These sites operate as media of communication and recollection, serving as symbols of identity and cues for remembrance within particular groups³⁵.

Cultural memory is commonly described as comprising three essential dimensions: material, functional, and symbolic³². The material dimension refers to the presence of cultural memory in tangible and visible forms—such as architecture, monuments, documents, artworks, and rituals—which not only preserve history but also provide a material basis for transmitting memory across generations. The functional dimension emphasizes the practical role of cultural memory in social life, focusing on how shared memories help groups maintain cultural identity and communicate values and social norms. The symbolic dimension concerns the representational content and meanings of cultural memory, highlighting how symbols and images convey deeper cultural significance and emotion^32,33.

Within this framework, cultural memory is closely linked to cultural heritage. Assmann argues that cultural memory provides a cohesive structure for the formation of national cultural identity through a “practical-formalized material world” and “symbolic-formalized rituals”³². Images play a central role in this process, perpetuating memory and offering crucial support for heritage preservation³⁶. Cultural heritage can thus function as an active instrument for reshaping problematic urban structures and enhancing urban areas socially and economically by mobilizing collective memory and identity²⁷. At the same time, the destruction and loss of cultural heritage lead to the gradual disappearance of cultural memory³⁷. Building on this perspective, Huang et al. further argue that the reconstruction of collective memory forms a fundamental basis for the sustainable development of cultural heritage¹³.

Semiotics theory

Semiotics is generally defined as the study of signs and symbols, including their meanings and uses^38,39. From this perspective, design can be regarded as a language and a system of symbols⁴⁰. Peirce argues that the essential function of signs is to make otherwise inefficient signifying relations effective⁴¹. If design is understood as a practice imbued with symbolic meaning, then the common objective of semiotics-based design theories is to identify a design language or at least a design grammar⁴². Many scholars maintain that semiotic theory holds significant value for solving design problems and should be placed at the core of design activities, providing a useful and applicable paradigm for design practice^42,43. In applied contexts, semiotics has been used to understand how designed environments stimulate user desires and shape meaning⁴⁴. Semiotic approaches have had a profound influence on communication studies—especially the study of organized meaning-making—as well as on design methodologies and pedagogies^42,45. Throughout design history, from early modern movements to contemporary practice, semiotic principles have served as tools for embedding deeper meanings into artifacts.

Artificial semiotics enables the application of semiotic knowledge and concepts to the analysis of visual systems and designed objects^42,43. Recent research also shows that the analytical tools of semiotics and discourse analysis are useful for examining public discourses about artificial intelligence⁴⁶. In product and interaction design, designers seek to convey purpose and meaning through visual appearance (aesthetics), pragmatic interaction elements (product semantics), and meanings that extend beyond physical interaction, that is, semiotic content⁴⁷.

Peirce’s tripartite division of signs—icons, indices, and symbols—has been widely adopted because it aligns well with functional requirements in design^48,49. Icons are signs characterized by similarity or imitation; there is a direct resemblance between the sign and its referent. Indices are signs defined by a causal or physical connection with their referents; they do not rely on similarity but indicate objects through an actual association. Symbols, in contrast, derive their meanings from social conventions or customs; no natural connection exists between a symbol and its referent, and meaning is established through shared cultural or linguistic habits^48,50.

There is substantial evidence that cultural memory theory and semiotics are closely connected. From a semiotic perspective, culture can be viewed as a hierarchy of interacting sign systems^39,51,52. Cultural memory is non-genetic and depends on a complex structure of such systems; its carriers are not merely external linguistic forms but the underlying semantic structures encoded within them^51,53. Collective memory is therefore embodied in publicly circulating symbols rather than residing solely in individual cognition⁵⁴. Building on Assmann’s work, Nora’s notion of lieux de mémoire (“sites of memory”) further illuminates how specific sites, objects, and images become focal points where collective memory is condensed and negotiated^54–56. Such sites simultaneously fulfil material, functional, and symbolic roles: they are tangible carriers in the landscape, they organize social practices and rituals, and they crystallize shared narratives and values^54,55. Interpreting these lieux de mémoire from a semiotic perspective helps to bridge the divide between material and ideal dimensions of culture and to specify how iconic, indexical, and symbolic meanings are layered within the same artifact^49,53,57. This perspective is particularly relevant for Huizhou wood carving, whose motifs operate as multi-layered signs that encode local histories, ethical codes, and cosmological views within a single visual–material system.

Existing research also clarifies how the three dimensions of cultural memory map onto different types of signs. Regarding the material dimension, Assmann shows that cultural memory relies on material symbols to preserve and convey historical narratives, while Eco’s analysis of similarity in signs and Peirce’s definition of icons as resemblance-based signs resonate with the way cultural artifacts visually embody shared meanings and experiences^32,48,50. The functional dimension of cultural memory, which emphasizes its practical role in social life, aligns with indexical signs: Assmann highlights how functional aspects allow direct cognition of practical attributes, and both Peirce and Eco point out that indices provide concrete references and practical cues within symbol systems^32,48,50. Finally, the symbolic dimension underscores the role of abstract, convention-based meanings in structuring collective identities and value systems. On this level, metaphorical and symbolic meanings are embedded in cultural memory^56,58, and symbols gain meaning through social conventions rather than natural connections⁵⁰. Thus, the symbolic dimension of cultural memory corresponds to conventional symbols in semiotics^38,41. In this sense, cultural memory theory and semiotic theory converge in asserting that symbols are central to the construction of meaning^32,39.

Taken together, these studies suggest that the material, functional, and symbolic dimensions of cultural memory are closely intertwined with visual sign systems in design. In this study, these dimensions are operationalized as three latent variables (MD, FD, SD), measured by items that capture the tangible carriers, practical functions, and symbolic meanings of Huizhou wood carving. Given the cross-sectional nature of the data, we treat the following relationships as theoretically directional associations rather than strict causal effects. Accordingly, we propose the following hypotheses:

H1

The material dimension (MD) of cultural memory theory is positively associated with icons (IC) in semiotics.

H2

The functional dimension (FD) of cultural memory theory is positively associated with index (IN) in semiotics.

H3

The symbolic dimension (SD) of cultural memory theory is positively associated with symbols (SY) in semiotics.

Function–Behavior–Structure (FBS) framework

The Function–Behavior–Structure (FBS) model is widely recognized as a mainstream design framework for supporting multi-level knowledge mapping and for analyzing conceptual transformations in the design process^59–61. It was originally proposed to address design challenges related to optimal decision-making through three core concepts: function, behavior, and structure⁵⁹. In this framework, three types of variables are used to describe different aspects of a design object. Function (F) variables describe the teleology of the object—what it is for. Behavior (B) variables describe the properties derived, or expected to be derived, from the structure, indicating what the object does. Structure (S) variables describe the components and their relationships within the object, defining what it is^59,60. Functions are attributed to behaviors, and behaviors are derived from structures, forming a chain from abstract intentions to concrete configurations⁵⁹.

Gero and Kannengiesser argue that the FBS framework marks an important milestone in design theory and can be used to describe virtually any design instance, regardless of domain or specific methods employed⁶⁰. With the continuous growth of design data and knowledge, numerous hybrid approaches inspired by the FBS model and related decision frameworks have been developed to accelerate the product design process. These include data-driven knowledge acquisition and evaluation processes^62–64, cognitive behavior-based problem-solving and behavioural design approaches^65,66, and the construction of design knowledge graphs and decision models under uncertainty⁶⁷.

Analysis of existing research further reveals a close connection between semiotic theory and the FBS framework. In semiotics, an index is a type of sign that maintains a direct, often causal, relationship with what it represents. Peirce describes indices as providing operational or orientational cues, and such cues can help clarify functional requirements in design^48,49,60. In the FBS framework, the function stage involves understanding real-world needs and indexical information⁵⁹. In design practice, indices are manifested as physical clues or markers that imply function, assisting both designers and users in identifying the intended functions of an artifact^68–70.

The iconic dimension of signs is also closely related to the behavior component in FBS. The form or imagery characteristics of icons can influence aesthetic and shape decisions during the design behavior stage^48,50,71. Gero notes that the behavior stage includes experimenting with and adjusting visual and morphological aspects of a design⁵⁹. As designers explore different forms, icons guide expectations about how the product will behave and how it will be experienced in use.

Finally, symbols are often used to embody social and cultural norms, simplifying complex information and making structures easier to understand^53,68. By using widely recognized symbols, designers can communicate function or behavior without lengthy explanations, thereby enhancing users’ understanding of a structure’s purpose^69,70. Symbols also encapsulate cultural memory, transcending language barriers and allowing users to intuitively grasp the intended function and behavior of structural elements^33,56.

These arguments indicate that indexical, iconic, and symbolic signs can be systematically mapped onto the function, behavior, and structure components of the FBS framework in heritage design. In our model, IN, IC, and SY are specified as latent variables reflecting perceived sign types, while F, B, and S are measured by items describing functional performance, behavioral responses, and structural form of AI-generated Huizhou wood-carving designs. On this basis, we conceptualize the following directional associations:

H4

Index signs (IN) in semiotics are positively associated with the function (F) variable in the FBS framework.

H5

Icon signs (IC) in semiotics are positively associated with the behavior (B) variable in the FBS framework.

H6

Symbol signs (SY) in semiotics are positively associated with the structure (S) variable in the FBS framework.

The FBS framework provides a systematic way of thinking about design and emphasizes that design is an iterative process^59–61. The eight processes described within the FBS framework are considered fundamental to all design activities, and their interactions are typically illustrated in FBS process diagrams(Fig. 1). Once functional requirements are clarified, they guide the specific behavioral processes in design. The final structural solution is directly determined by the outcomes of the behavior stage^59,60.

Fig. 1.

The eight processes described within the FBS framework59.

When AI-generated structural outcomes are produced, the value of digital heritage and innovation is ultimately judged by how these structures perform at visual, cultural, and functional levels^68–71. In the digital recreation of Huizhou woodcarving, the effectiveness of innovation and heritage preservation is therefore assessed by both the quality of the final structural configuration and its cultural significance within the broader system of cultural memory^31,58.

Drawing on these arguments, we conceptualize the following directional relationships within the FBS framework. Given the cross-sectional design of this study, these paths are treated as theoretically directional associations rather than strict causal effects:

H7

The function (F) variable in the FBS framework is positively associated with the behavior (B) variable.

H8

The behavior (B) variable in the FBS framework is positively associated with the structure (S) variable.

H9

The structure (S) variable in the FBS framework is positively associated with the digital preservation and innovation of cultural heritage (CHDII).

Research model

Based on a comprehensive review of the relevant literature, the structural model shown in Fig. 2 is proposed in this study. Consistent with the cross-sectional design, all paths in the model are formulated as theoretically directional associations informed by the above theoretical arguments.

Fig. 2.

Research hypothesis model.

Methods

Research instruments and AI-generated stimuli

This study adopted a cross-sectional survey design to examine how constructs from cultural memory theory, design semiotics, and the Function–Behavior–Structure (FBS) framework jointly shapes users’ evaluations of digital inheritance and innovation in Hui-style woodcarving. The survey combined exposure to a standardized set of AI-generated images with a structured questionnaire.

Hui-style woodcarving was selected as the focal case because it embodies rich symbolic motifs, narrative scenes, and structural patterns that are highly suitable for digital reinterpretation. To provide participants with a concrete and consistent visual context, we constructed a set of AI-generated Hui-style woodcarving images using Midjourney, a commercially available text-to-image generative AI system based on diffusion models. In this study, Midjourney was treated not as an autonomous creator, but as a controllable rendering engine driven by theory-informed textual prompts. The overall research workflow, from theory-informed prompt development to AI image generation, survey administration, and data analysis, is summarised in Fig. 3.

Fig. 3.

Overall research workflow from theoretical framework through prompt development, AI image generation and expert review, survey procedure, and SEM–ANN data analysis.

The stimulus development followed a two-stage procedure. First, the research team compiled a library of descriptive phrases grounded in the three dimensions of cultural memory theory (material, functional, symbolic) and the three semiotic sign types (icon, index, symbol), as well as general design descriptors related to composition, structure, and style. These phrases were combined into prompt templates that kept global style parameters constant (Hui-style woodcarving theme, viewpoint, lighting, level of detail, and resolution), while varying the semantic elements related to cultural memory and symbolic content.

To ensure that the AI-generated images were explicitly grounded in our theoretical model, we constructed theory-informed prompt templates rather than relying on ad-hoc or purely aesthetic prompt engineering. For each construct in the Cultural Memory–Semiotics–FBS framework (MD, FD, SD, IC, IN, SY, F, B, S), we compiled small sets of modular prompt components—short descriptive phrases capturing material carriers, functional roles, symbolic themes, sign types, and design attributes—that could be combined in a Midjourney-friendly format (e.g., material + scene + symbolic motif + function/composition). These phrase sets were then systematically assembled into structured prompts that guided the generation of Hui-style woodcarving images in Midjourney. Table 1 summarizes how each theoretical construct was operationalized into specific prompt components, with illustrative examples of the modular phrases used.

Table 1.

Mapping of theoretical constructs to prompt components for AI-generated Huizhou woodcarving images.

Module	Construct	Role in Prompt	Example Prompt Components
Cultural Memory	MD	Specifies material carriers and craft techniques	Huizhou-style wooden panel, fine wood grain, high-relief carving, openwork lattice, aged hardwood
Cultural Memory	FD	Describes usage scenarios and social functions	ancestral temple interior, clan hall entrance, ceremonial doorway, decorative beam in traditional house
Cultural Memory	SD	Highlights cultural themes and symbolic meanings	motifs of family prosperity, many descendants, longevity symbols, blessings and auspicious patterns
Semiotics	IC	Introduces concrete figurative objects that resemble their referents	lifelike human figures, birds and fish, dragons and phoenixes, vividly carved flowers and plants
Semiotics	IN	Adds contextual cues that point to actions, events or environments	scene of everyday labour, festive celebration, ritual gathering, smoke and lanterns in the background
Semiotics	SY	Uses conventional motifs and allegorical symbols	traditional lucky characters “fu” and “shou”, mythic beasts, allegorical story motifs, clan emblems
FBS	F	States the intended design purpose and role in the built environment	woodcarving used as door lintel ornament, spatial divider panel, focal altar backdrop in the hall
FBS	B	Describes expected experiential and perceptual effects	guiding the viewer’s gaze inward, creating strong depth and layering, inviting viewers to pause and look
FBS	S	Constrains overall composition and structural organization	multi-layered framed composition, clear geometric segmentation, vertical symmetry, balanced layout

Second, two preliminary groups of 20 images each were generated during the conceptual design stage. The first group was produced with prompts that explicitly integrated phrases derived from cultural memory theory and semiotics, emphasizing functional cues, narrative connotations, and symbolic meanings. The second group was generated using simpler descriptive prompts that referenced Hui-style woodcarving patterns in general, with minimal theoretical elaboration. These two groups were not used as separate experimental conditions in the main survey but served only as candidate sets for expert review and refinement.

Three senior experts in design and cultural heritage studies, together with two doctoral students in design, independently evaluated the 40 candidate images on visual clarity, recognizability of Hui-style woodcarving features, perceived cultural depth, symbolic richness, and structural organization. Based on their feedback, a final standardized stimulus set of 20 images was compiled by selecting and, when necessary, lightly adjusting images that best reflected clear structural composition and medium-to-high levels of cultural and symbolic content, while maintaining a coherent visual style. Figure 4 presents representative examples from this final standardized image set used in the main survey.

Fig. 4.

Representative examples from the standardized set of 20 AI-generated Huizhou woodcarving images used in the main survey: (A–C) everyday family gatherings and agricultural labour scenes; (D–F) auspicious and symbolic motifs such as cranes and pines, guardian qilin, and fish-and-lotus patterns.

During the main survey, all participants viewed the same standardized set of 20 AI-generated Hui-style woodcarving images. The two preliminary image groups were used only during the expert review and pilot refinement stage and did not constitute analytical factors in the structural equation or neural network models.

Measures and operationalization of constructs

The measurement instrument was designed to operationalize ten latent constructs derived from cultural memory theory, design semiotics, and the FBS framework in the context of digitally reinterpreted Hui-style woodcarving. The final questionnaire included 30 items, with three items per construct: material dimension (MD), functional dimension (FD), and symbolic dimension (SD) of cultural memory; index (IN), icon (IC), and symbol (SY) from design semiotics; function (F), behavior (B), and structure (S) from the FBS framework; and perceived cultural heritage digital inheritance and innovation (CHDII). All items were rated on a 5-point Likert scale ranging from 1 = “strongly disagree” to 5 = “strongly agree”.

Item development proceeded in several stages. First, an initial pool of 45 candidate items was compiled from established scales and prior studies on cultural memory, semiotics in design, and FBS-based evaluation of products and visual artifacts. Items were included if they (a) explicitly referred to material, functional, or symbolic aspects of cultural artifacts; (b) described users’ perceptions of iconic, indexical, or symbolic cues in visual design; or (c) captured evaluative judgements related to intended function, perceived behavior, structural form, and innovative character of design outcomes.

Second, each item was mapped onto one of the ten theoretical constructs based on its conceptual content. Two design scholars and a PhD student in design independently assigned the items to MD, FD, SD, IN, IC, SY, F, B, S, or CHDII using construct definitions derived from the literature review. Items that showed inconsistent assignments or ambiguous wording were discussed in a consensus meeting. Ambiguous or overlapping items were either revised for clarity or removed. This procedure resulted in a balanced set of three items for each construct.

The preliminary item set was then reviewed by three senior experts in design theory and cultural heritage. They assessed each item for clarity, cultural appropriateness, and alignment with the intended construct, and suggested minor wording changes to improve readability and avoid technical jargon for non-expert respondents. A pilot test with 10 participants familiar with Chinese cultural heritage and design topics was conducted to check comprehension, response variability, and completion time. Participants were invited to flag any confusing expressions or redundant questions; based on their feedback, a small number of phrases were simplified.

Because the main survey was administered in Chinese, all items were originally drafted in Chinese. To ensure conceptual equivalence for international reporting, a back-translation procedure was applied. A bilingual researcher independently translated the items from Chinese into English and then back into Chinese. Discrepancies were resolved through discussion until semantic consistency was achieved. The final Chinese version was used in data collection, while the English wording is presented in Table 2 together with the sources from which the items were adapted.

Table 2.

Question items and references.

Construct	Item codes	Example item (English wording)	Key references (examples)
Material dimension(MD)	MD1–MD3	“The materials and textures in these designs remind me of the traditional atmosphere of Hui-style woodcarving.”	[32,35,56]
Functional dimension (FD)	FD1–FD3	“The depicted scenes reflect functions and uses associated with Hui regional culture.”	[31,32,37]
Symbolic dimension (SD)	SD1–SD3	“The carved motifs convey symbolic meanings about family, morality, and local beliefs.”	[34,36,55]
Icon (IC)	IC1–IC3	“Some visual elements directly resemble real objects from Hui daily life.”	[39,48,50]
Index (IN)	IN1–IN3	“Certain details imply underlying stories or events rather than depicting them directly.”	[48,51,54]
Symbol (SY)	SY1–SY3	“Abstract patterns in the designs stand for broader cultural ideas or values.”	[50,56,58]
Function (F)	F1–F3	“The design clearly communicates its intended purpose.”	[59,60,68]
Behavior (B)	B1–B3	“The arrangement of elements suggests how the design should be used or experienced.”	[61,63,71]
Structure (S)	S1–S3	“The overall structure of the composition feels coherent and well-organized.”	[59,65,70]
Cultural heritage digital inheritance and innovation (CHDII)	CHDII1–CHDII3	“These digitally designed images successfully combine traditional Hui-style elements with contemporary aesthetics.”	[27,36,37]

Example items include: “The materials and textures in these designs remind me of the traditional atmosphere of Hui-style woodcarving” (MD), “The depicted scenes reflect functions and uses associated with Hui regional culture” (FD), and “The carved motifs convey symbolic meanings about family, morality, and local beliefs” (SD). For the semiotic constructs, illustrative items are “Some visual elements directly resemble real objects from Hui daily life” (IC), “Certain details imply underlying stories or events rather than depicting them directly” (IN), and “Abstract patterns in the designs stand for broader cultural ideas or values” (SY). For the FBS constructs, examples include “The design clearly communicates its intended purpose” (F), “The arrangement of elements suggests how the design should be used or experienced” (B), and “The overall structure of the composition feels coherent and well-organized” (S). Items for CHDII capture perceived effectiveness of digital inheritance and innovation, such as “These digitally designed images successfully combine traditional Hui-style elements with contemporary aesthetics.”

All items were coded so that higher scores indicate stronger agreement with the respective statements. In the subsequent analyses, each construct was modeled as a latent variable indicated by its three items. Reliability and validity of the measurement model (Cronbach’s alpha, composite reliability, and confirmatory factor analysis) are reported in Sect. “ Data analysis”.

An overview of the latent constructs, item codes, and example content is provided in Table 2. The full wording of all 30 items in Chinese and English is presented in Appendix A. Appendix A presents the full wording of all items in English. For brevity, example items in the main text are paraphrased to highlight their reference to the AI-generated Huizhou woodcarving images.

Samples and data collection procedure

The target population of this study comprised adults with design-related backgrounds and a basic interest in Chinese cultural heritage and visual design.

Data were collected via an online questionnaire hosted on a widely used survey platform in China. A non-probability convenience sampling strategy was adopted, with the survey link disseminated through social media platforms and university mailing lists. Participation was voluntary and anonymous.

Upon accessing the survey link, prospective respondents first viewed an information sheet outlining the purpose of the study, the approximate duration of participation, data protection measures, and their rights as participants. Only individuals who confirmed that they were 18 years or older and provided electronic informed consent were allowed to proceed.

The procedure consisted of three stages. First, participants were given a brief textual introduction to Hui-style woodcarving and informed that they would see a set of AI-generated images illustrating how traditional motifs might be reinterpreted in contemporary digital form. Second, they viewed the standardized stimulus set of 20 AI-generated Hui-style woodcarving images described in Sect. “Research instruments and AIgenerated stimuli”. The images were displayed within the online questionnaire as a single stimulus set, and participants were asked to browse all of them before proceeding. The instructions emphasized that they should focus on their overall impressions of the style, symbolic content, and design quality, rather than analyzing each image in isolation. All respondents viewed the same 20 images in the same order and format.

Third, participants completed the questionnaire. The first part collected demographic information, including age, gender, education level, and self-reported familiarity with Hui-style woodcarving and Chinese traditional culture. The second part comprised the 30 items measuring the ten latent constructs (MD, FD, SD, IN, IC, SY, F, B, S, CHDII) described above. Additional attention-check items were embedded to identify random or careless responding.

A total of 507 responses were initially recorded. After excluding cases with excessive missing data, implausibly short completion times, or failed attention checks, 434 valid questionnaires remained for analysis (see Table 3 for detailed demographic characteristics). The final sample size satisfies commonly recommended criteria for structural equation modeling, thereby supporting the robustness of the subsequent analyses.

Table 3.

Demographic characteristics of respondents.

Attribute	Items	Numbers	Percentage (%)
Gender	Male	206	47.46
	Female	228	52.53
Age	18–25	139	32.02
	26–35	203	46.77
	35 and above	92	21.19
Education	Undergraduate	349	80.41
	Master	62	14.28
	PhD	23	5.29
Profession	Design Students	204	47.00
	Design Teachers	91	20.96
	Designers	97	22.35
	Design Enthusiasts	42	9.67
Generative AI Experience	Yes	305	70.27
	No	129	29.72

Ethical principles were strictly adhered to throughout the study. Participation was fully voluntary, no identifying personal information was collected, and respondents could withdraw at any time by closing the survey window without penalty.

Data preprocessing and analysis overview

Before testing the structural relationships among the constructs, the dataset was preprocessed and screened. Cases with extensive missing values or patterned responses were removed as described in Sect. “Samples and data collection procedure”. Descriptive statistics (means, standard deviations, skewness, and kurtosis) were computed for all items to examine data distribution. All variables met commonly used thresholds for univariate normality, indicating suitability for covariance-based structural equation modeling.

Next, the measurement model was assessed using confirmatory factor analysis (CFA) to evaluate factor loadings, internal consistency, and convergent and discriminant validity of the ten latent constructs. Upon obtaining an acceptable measurement model, the hypothesized structural model was estimated to test the directional associations specified in H1–H9. Given the cross-sectional design, these hypotheses were formulated and interpreted as theoretically directional positive associations, rather than strict causal effects.

To complement the linear SEM analysis and explore potential nonlinear relationships, an artificial neural network (ANN) model was subsequently trained using the latent construct scores as inputs and CHDII as the output variable. Comparing SEM path coefficients and ANN-derived importance measures allowed us to triangulate the relative influence of cultural memory dimensions, semiotic constructs, and FBS-related design attributes on perceived digital inheritance and innovation. Detailed results of the descriptive statistics, CFA, SEM, and ANN analyses are presented in Sect.“ Data analysis”.

Data analysis

Descriptive statistics and normality test

We examined the univariate distributions of all observed variables. As summarized in Appendix B, means, standard deviations, skewness, and kurtosis all fell within commonly used cutoffs for approximate normality, indicating suitability for subsequent CFA and SEM analyses.

Measurement model

The measurement properties of the ten latent constructs are summarized in Tables 4 and 5. Standardized factor loadings range from 0.516 to 0.922 for all items except two indicators of the index dimension (IN1 and IN3), which show lower loadings (0.413 and 0.353) but were retained because they capture conceptually important aspects of indexical design and do not undermine overall model fit. All loadings are statistically significant (C.R. > 1.96, p < 0.001), supporting the associations between items and their intended constructs.

Table 4.

Results of construct validity and reliability analysis.

Construct	Code	Mean	Standard Deviation	Factor Loadings	C.R.	P	α	AVE	CR
Functional Dimension (FD)	FD1	3.147	1.005	0.536		***	0.832	0.549	0.751
	FD2			0.545	6.971	***
	FD3			0.566	7.090	***
Material Dimension (MD)	MD1	3.247	1.055	0.673		***	0.817	0.575	0.710
	MD2			0.522	8.971	***
	MD3			0.637	10.531	***
Symbolic Dimension (SD)	SD1	3.423	1.191	0.897		***	0.855	0.913	0.968
	SD2			0.922	29.828	***
	SD3			0.919	29.605	***
Index (IN)	IN1	3.165	0.932	0.413			0.893	0.433	0.635
	IN2			0.516	4.914	***
	IN3			0.353	4.217	***
Icon (IC)	IC1	2.828	1.071	0.553			0.876	0.657	0.795
	IC2			0.609	8.124	***
	IC3			0.742	8.371	***
Symbol (SY)	SY1	3.250	1.097	0.705			0.915	0.698	0.860
	SY2			0.686	12.554	***
	SY3			0.702	12.816	***
Function (F)	F1	3.187	0.998	0.552			0.798	0.545	0.748
	F2			0.538	7.396	***
	F3			0.545	7.446	***
Behavior (B)	B1	3.276	0.985	0.619			0.851	0.592	0.786
	B2			0.558	9.965	***
	B3			0.596	10.520	***
Structure (S)	S1	3.337	1.126	0.752			0.877	0.732	0.881
	S2			0.699	13.944	***
	S3			0.744	14.856	***
Cultural Heritage Digital Inheritance and Innovation (CHDII)	CHDII1	3.443	1.085	0.653			0.932	0.694	0.858
	CHDII2			0.711	11.965	***
	CHDII3			0.716	12.029	***

Note: α Cronbach’s alpha, CR composite reliability, AVE average variance extracted.

***Significant at 0.001 level (two-tailed).

Table 5.

Results of the correlation analysis and the discriminant validity comparison among dimensions.

	FD	MD	SD	IN	IC	SY	F	S	B	CHDII
FD	0.741
MD	0.289**	0.800
SD	0.260**	0.249**	0.955
IN	0.223**	0.218**	0.177**	0.658
IC	0.121*	0.193**	0.150**	− 0.047	0.809
SY	0.328**	0.327**	0.597**	0.220**	0.168**	0.835
F	0.292**	0.289**	0.191**	0.259**	− 0.072	0.320**	0.738
B	0.368**	0.372**	0.576**	0.249**	0.250**	0.656**	0.266**	0.856
S	0.393**	0.586**	0.325**	0.248**	0.164**	0.424**	0.461**	0.517**	0.769
CHDII	0.420**	0.483**	0.437**	0.308**	0.199**	0.518**	0.345**	0.603**	0.544**	0.833

Note: Bold (on diagonal) represents the square root of the variable’s AVE.

** The correlation is significant at the 0.01 level (two-tailed).

* The correlation is significant at the 0.05 level (two-tailed).

Cronbach’s alpha values range from 0.765 to 0.878, and composite reliability (CR) values from 0.710 to 0.968, exceeding the conventional thresholds (α ≥ 0.70, CR ≥ 0.70) and indicating satisfactory internal consistency for all dimensions^72,73. With the exception of the index construct (AVE = 0.433), average variance extracted (AVE) values lie between 0.545 and 0.913, meeting the recommended cutoff of 0.50 and suggesting adequate convergent validity. In addition, the square roots of AVE for all constructs exceed their inter-construct correlations (Table 5), which indicates good discriminant validity^73–75.

Structural model

The structural model demonstrates an acceptable overall fit to the data (Table 6): χ² = 957.62, χ²/df = 2.32, RMSEA = 0.056, CFI = 0.875, TLI = 0.863, and AGFI = 0.857, all falling within commonly used guidelines for SEM model adequacy^73,76,77. These indices support the robustness of the proposed structural relationships.

Table 6.

Model fit indices.

Model	X2	X2/DF	AGFI	TLI	CFI	RMSEA
Confirmatory Factor Model	543.753	1.531	0.904	0.951	0.958	0.033
Structural Equation Model	957.615	2.324	0.857	0.863	0.875	0.056
Standard Value	P > 0.05	1 ~ 5	> 0.8	> 0.8	> 0.8	< 0.08

Hypothesized relationships were evaluated by inspecting the path coefficients in the structural model (Table 7; Fig. 5). Eight out of nine paths are statistically significant, with only H1 (MD → IC; p = 0.057) not supported at the 0.05 level. In contrast, paths linking the functional, symbolic, and indexical dimensions to higher-order visual constructs are all positive and substantial: FD → IN (H2, β = 0.709), SD → SY (H3, β = 0.742), and IN → F (H4, β = 0.824; all p < 0.001).

Table 7.

Results of path coefficients hypotheses.

Hypothesis	Path	B	S.E.	C.R.	P	Result
H1	PD→ IC	0.353	0.079	4.733	0.057	Rejected
H2	MD→IN	0.709	0.109	4.345	***	Supported
H3	SD→SY	0.742	0.045	12.849	***	Supported
H4	IN→F	0.824	0.266	4.511	***	Supported
H5	IC→B	0.381	0.055	5.613	0.042	Supported
H6	SY→S	0.763	0.068	10.91	***	Supported
H7	F→B	0.813	0.129	6.957	***	Supported
H8	B→S	0.533	0.075	8.061	0.018	Supported
H9	S→CHDII	0.807	0.077	10.319	***	Supported

Note: B Unstandardized coefficient, S. E Standardized estimates; *** p < 0.001, **p < 0.01, *p < 0.05.

Fig. 5.

Results of path analysis (All hypotheses are supported, except H1). ***p < 0.001, **p < 0.01, *p < 0.05.

At the behavioral level, higher levels of perceived icon design are associated with higher levels of user behavior in the SEM model (IC → B; H5, β = 0.381, p < 0.05), and function shows a strong positive association with behavior (F → B; H7, β = 0.813, p < 0.001). Symbolic design is positively associated with perceived structure (SY → S; H6, β = 0.763, p < 0.001), and behavior is positively associated with structure (B → S; H8, β = 0.533, p < 0.05). Finally, perceived structure is strongly and positively associated with cultural heritage digital inheritance and innovation (S → CHDII; H9, β = 0.807, p < 0.001).

Overall, the pattern of results is consistent with a cascading mechanism in which functional–material–symbolic design features are linked to higher-level visual structures and user behavior, which in turn are associated with higher perceived digital inheritance and innovation of cultural heritage visual images.

Artificial neural network (ANN) analysis

To complement the linear SEM and explore potential nonlinear and non-compensatory effects among the key constructs, we estimated two multilayer perceptron (MLP) artificial neural network models in SPSS, following prior SEM–ANN studies in behavioral and technology-adoption research^78–80. Both models used a sigmoid activation function and ten-fold cross-validation, with 90% of the data allocated to training and 10% to testing. Model A predicted user behavior (B) from icon (IC) and function (F), whereas Model B predicted perceived structure (S) from behavior (B) and symbol (SY).

As shown in Table 8, training and testing root mean square error (RMSE) values range from 0.162 to 0.398 across the ten networks in each model, with average RMSEs of 0.352 (training) and 0.338 (testing) for Model A and 0.260 (training) and 0.249 (testing) for Model B. These relatively low and stable error levels indicate satisfactory predictive performance, consistent with previous ANN-based prediction studies^81,82. Sensitivity analysis (Table 9) further shows that function (F) and symbol (SY) emerge as the most influential predictors in Models A and B, respectively, reinforcing the SEM finding that functional and symbolic design dimensions play a central role in shaping user behavior and perceived structural qualities.

Table 8.

RMSE values for models A and B.

Neural network	Model A		Model B
	Input: Icon(IC), Function(F)		Input: Behavior(B), Symbol(SY)
	Output: Behavior(B)		Output: Structure(S)
	Training	Testing	Training	Testing
ANN1	0.359	0.249	0.253	0.253
ANN 2	0.356	0.343	0.262	0.240
ANN 3	0.358	0.353	0.262	0.162
ANN 4	0.347	0.366	0.259	0.263
ANN 5	0.341	0.356	0.269	0.357
ANN 6	0.351	0.360	0.248	0.274
ANN 7	0.371	0.348	0.250	0.244
ANN 8	0.337	0.398	0.261	0.173
ANN 9	0.347	0.319	0.254	0.266
ANN10	0.349	0.291	0.283	0.256
Mean	0.352	0.338	0.260	0.249
SD	0.010	0.042	0.010	0.054

Table 9.

Sensitivity analysis for models A and B.

Neural network	Model A (Output: B)		Model B (Output: S)
	IC	F	B	SY
ANN1	0.612	1.000	0.646	1.000
ANN 2	0.292	1.000	0.615	1.000
ANN 3	0.313	1.000	0.510	1.000
ANN 4	0.459	1.000	0.664	1.000
ANN 5	0.427	1.000	0.584	1.000
ANN 6	0.361	1.000	0.545	1.000
ANN 7	0.520	1.000	0.481	1.000
ANN 8	0.421	1.000	0.474	1.000
ANN 9	0.402	1.000	0.613	1.000
ANN10	0.484	1.000	1.000	0.898
Average relative importance	0.429	1.000	0.613	0.990
Normalized relative importance (%)	70.114	100.000	61.320	98.980

Discussion

Addressing the research questions

Building on the integrated SEM–ANN analysis, this section discusses how the empirical findings answer the three research questions and what they imply for cultural memory, design semiotics, and the FBS framework. Overall, the structural equation model confirmed eight of the nine hypotheses, and the neural network analyses yielded convergent importance rankings. Together, they empirically substantiate core propositions in cultural memory theory, semiotics, and FBS-based design reasoning—namely, that cultural symbols and functional cues act as key mediators between collective memory and perceived design quality, and that users’ evaluations depend on the coherent alignment of function, behavior, and structure in design outcomes^{32,50,60,68,70}. At the same time, the results refine and extend these theories to the context of AI-generated heritage images by showing which specific dimensions matter most for perceived digital inheritance and innovation. Given the cross-sectional design, these patterns are interpreted as theoretically directional associations rather than strict causal effects.

RQ1 asked how cultural memory theory relates to the digital inheritance and innovation of cultural heritage.

The finding that participants’ CHDII evaluations are driven predominantly by the functional and symbolic dimensions of cultural memory, whereas the material dimension plays a comparatively weaker role, is consistent with cultural memory research emphasising the practical and symbolic functions of lieux de mémoire in sustaining identity and value systems^32,35,55. Higher scores on FD and SD were associated with stronger perceptions of indexical and symbolic signs in the AI-generated images, and these, in turn, contributed to more positive CHDII judgements. By contrast, the material dimension did not significantly enhance the perceived iconicity of the motifs along the hypothesised pathway, echoing arguments that material carriers must be culturally “readable” to remain effective memory media.

On this basis, the present study extends cultural memory theory to digital, AI-mediated representations. It shows that, when encountering heritage through AI-generated images, users are less persuaded by literal reproduction of material details than by how convincingly a design signals “what this pattern does” and “what it stands for”. Functional cues that anchor scenes in recognisable social practices, and symbolic themes that condense shared beliefs, become the primary drivers of perceived cultural continuity and innovation. In other words, under digital conditions, functional and symbolic readability can override material resemblance as key pathways through which cultural memory informs evaluations of heritage design.

RQ2 examined how semiotics shapes or is associated with the digital inheritance and innovation of cultural heritage.

The confirmed pathways from FD and SD to IN and SY, and from IN, IC, and SY to F, B, and S, empirically support semiotic accounts which argue that cultural meaning is organised through layered systems of icons, indices, and symbols^48,50,58. Indexical signs, which point to practical functions and situational cues, were strongly linked to evaluations of function and behavior, while symbolic signs contributed directly to structural judgements. Iconic similarity, in contrast, mainly influenced behavioural and aesthetic responses and played a modest role in predicting CHDII, aligning with prior work that sees icons as necessary but not sufficient for deep meaning-making in design.

These results extend design semiotics by specifying, with quantitative evidence, how different sign types are differentially aligned with functional, behavioural, and structural aspects of design in a cultural-heritage context. Rather than merely making images “look like” traditional motifs, effective digital heritage design depends on an interplay of indexical and symbolic cues that show what actions are taking place, what roles are implied, and what values are encoded. Semiotics thus operates as a mediating “translation layer” between cultural memory and design evaluation, clarifying how culturally grounded signs are re-encoded in AI-generated images and how they shape users’ perceptions of digital inheritance and innovation.

RQ3 investigated how the FBS framework is linked to the digital inheritance and innovation of cultural heritage.

The supported F → B → S → CHDII chain corroborates FBS theory, which posits that clarified functions guide behavioural reasoning and that behaviours, in turn, determine structural solutions and user evaluations^59,60. In line with prior research on product semantics and user experience^68,70,71, the results show that participants’ global evaluations of AI-generated heritage designs depend on more than isolated motifs or narrative scenes. Designs that clearly communicate what they are for (F), how users might interact or move within them (B), and how components form a coherent whole (S) are more likely to be perceived as simultaneously respectful of tradition and innovatively reimagined.

The ANN sensitivity analysis reinforces this picture by assigning high importance to function-, behavior-, and symbol-related variables, while material/iconic cues played a comparatively modest role. In doing so, it confirms that the FBS framework provides an effective lens for understanding how cultural information, once encoded into signs, is further translated into perceived design quality and innovation in AI-generated heritage images. At the same time, the SEM–ANN combination highlights the non-linear, non-compensatory nature of these relationships: weak function or structure cannot be fully offset by attractive iconic details. This insight broadens existing FBS-based design research by positioning cultural significance—not just performance or usability—as a central criterion in evaluating digitally reimagined heritage.

Theoretical implications

Taken together, the findings point to several theoretical implications for cultural memory, design semiotics, and FBS-based design research.

One key implication is that digital heritage research should move from technology-centred models toward culture-centred models. Rather than treating cultural content as an optional “skin” applied to technologically sophisticated outputs, the analysis shows that CHDII is primarily shaped by the functional and symbolic dimensions of cultural memory and by their translation into indexical and symbolic signs. Cultural attributes are therefore not secondary decorations but constitute the main explanatory pathways in users’ evaluations of AI-generated heritage images.

A further implication is that the integrated Cultural Memory–Semiotics–FBS model offers a multi-level account of how cultural information flows from abstract meanings to concrete design structures. Instead of examining cultural symbols, user behavior, and structural composition as separate topics, the model specifies a sequential chain linking them: FD/SD → IN/SY → F/B/S → CHDII. This linkage bridges symbolic analysis and function–behavior–structure reasoning within a single quantitative framework and offers a way to operationalise theories of cultural memory and semiotics in design research, addressing longstanding calls to connect meaning-making and design cognition more systematically.

A final implication concerns the role of generative AI in theory-driven design workflows. The proposed pipeline—deriving modular prompt components from theoretical constructs, generating and filtering images, and then modelling user evaluations via SEM and ANN—demonstrates that text-to-image systems such as Midjourney can be used as controllable rendering engines constrained by cultural and design theory. This contributes to emerging discussions on “explainable” or “interpretable” AI in design by showing how symbolic, functional, and structural variables can be explicitly encoded in prompts and then empirically evaluated, rather than relying solely on ad-hoc or purely aesthetic prompt engineering.

Practical implications

Beyond theoretical contributions, the findings also offer practical guidance for designers, educators, and heritage institutions working with generative AI.

For design practice, the results suggest that successful AI-assisted heritage designs should prioritise the encoding of functional scenarios and symbolic narratives. When constructing prompts and curating AI outputs, designers might not only focus on enumerating surface attributes but also on describing social actions, roles, and values. Such prompts are more likely to generate images that users read as culturally meaningful rather than merely decorative.

For cultural heritage professionals and institutions, the Cultural Memory–Semiotics–FBS framework can be used as a checklist for evaluating digital projects. Curators and educators can ask: Does this AI-generated image convey recognisable functional contexts (FD)? Are indexical cues present that guide the viewer’s understanding of what is happening (IN)? Are key symbolic themes clearly and coherently encoded (SD, SY)? Is the overall structure organised in a way that supports learning, reflection, or identity formation (F/B/S)? Using these questions as evaluative criteria may help align digital outputs with broader goals of heritage transmission and public engagement.

For design education, the study suggests that training in generative AI should be integrated with, rather than separated from, training in cultural theory and design semiotics. Courses could employ exercises where students derive prompts from cultural memory constructs, generate images, and then critique them using semiotic and FBS perspectives. This would help future designers develop a more reflexive, theory-informed approach to AI tools, moving beyond “trial-and-error” prompting to culturally and ethically grounded design practice.

Limitations and future research

Several limitations of this study should be acknowledged, which also suggest directions for future work.

First, the empirical investigation focused on a single heritage case—Huizhou wood carving—and on AI-generated 2D images rather than 3D artefacts or immersive environments. Although the constructs in the Cultural Memory–Semiotics–FBS model were conceptualised at a general level, the specific item wording and visual stimuli inevitably reflect the characteristics of this craft tradition. Future research should replicate and extend the model in other symbol-rich heritage domains, such as religious murals, textile patterns, temple architecture, or ritual objects, to test its cross-cultural robustness and adaptability.

Second, the study employed a cross-sectional survey design without experimental manipulation or longitudinal follow-up. As a result, the SEM and ANN analyses identify associative patterns and theoretically directional pathways, but they do not establish causal effects. Future studies could adopt experimental designs—for example, systematically varying the presence or absence of functional and symbolic cues in AI-generated designs—or longitudinal designs tracking how repeated exposure to such images shape learning, attachment, and behavioural intention over time.

Third, the ANN analysis, while useful for capturing nonlinear patterns and ranking variable importance, was based on self-reported evaluations and a limited set of latent constructs. Behavioural and physiological measures, such as eye-tracking, interaction logs, or neural indicators, could be incorporated to provide a richer picture of how users actually attend to and process AI-generated heritage images. Comparing different machine-learning architectures or ensemble methods could also help assess the robustness of the findings and further clarify the relative contributions of cultural, semiotic, and structural factors.

Despite these limitations, the study demonstrates how a theory-driven SEM–ANN pipeline can be combined with generative AI to investigate digital heritage design, offering a methodological template for future interdisciplinary work at the intersection of cultural studies, design research, and human–AI interaction.

Conclusions

This study proposed and empirically tested an integrated Cultural Memory–Semiotics–FBS model to explain how AI-generated images of Huizhou wood carving support cultural heritage digital inheritance and innovation. Using a cross-sectional survey and a two-stage SEM–ANN analytic strategy with data from 434 participants, the study identified two dominant association pathways: a functional–index–FBS chain in which functional dimensions of cultural memory, perceived indexical signs, and evaluations of function, behavior, and structure jointly predicted CHDII, and a symbolic–symbol–structure chain linking symbolic dimensions of cultural memory and symbol signs to structural evaluations. In contrast, a material/icon pathway did not exhibit a significant effect on CHDII.

These findings highlight that users’ evaluations of AI-generated heritage designs depend more on how effectively functional contexts and symbolic meanings are encoded into visual signs and structural compositions than on the simple replication of traditional motifs. In other words, digital inheritance and innovation are driven less by “how similar the image looks” and more by “what actions and meanings the image makes thinkable”.

Practically, the study suggests that designers and cultural institutions should treat generative AI not as a black-box creator of decorative images, but as a controllable tool for re-encoding cultural meaning. Prompting strategies and selection criteria that emphasise functional cues, symbolic narratives, and coherent structural layouts are more likely to produce digital artefacts that audiences perceive as both culturally authentic and innovatively designed. More broadly, the proposed framework and workflow offer a reusable methodological pattern for future studies seeking to integrate cultural theory, design reasoning, and AI tools in the service of sustainable heritage transmission.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, Y.Q.; Methodology, Y.Q., X.P.; Software, Q.B.; Validation, Y.Q. and X.P.; Formal analysis, Y.Q. and X.P.; Investigation, Y.Q. and X.P.; Writing—original draft preparation, Y.Q.; Writing—review and editing, Y.Q., and Q.B.; Visualization, Y.Q. and S.Z.; Supervision, X.P. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data availability

All data generated or analyzed during this study are included in this article. The raw data are available from the First Author Yanran Qian upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Institutional review board statement

Upon review by the Hanyang University Institutional Review Board, the experimental design and research plan have been deemed scientifically sound, fair, and impartial, posing no harm or risk to participants. Participant.

recruitment is based on the principles of voluntariness and informed consent, with full protection of participants’ rights and privacy. The study involves no conflict of interest and does not violate ethical standards or legal regulations.

Informed consent

Informed consent was obtained from all participants involved in the study.