RIS theory

The RIS theory seeks to explain how different innovation actors within a region interact and collaborate to enable knowledge sharing and collective innovation, thereby promoting regional economic growth and industrial upgrading and providing a theoretical foundation for regional innovation policy. Cooke [12] pioneered the typological framework of RIS by classifying regional innovation systems into three categories—grassroots, networked, and dirigiste (or interventionist)—based on governance structures and knowledge interaction mechanisms. These types correspond respectively to innovation models driven by informal knowledge exchange among local small and medium-sized enterprises, coordinated knowledge networks led by multiple actors from academia, industry, and research institutions, and top-down resource allocation and innovation guidance promoted by governments or large enterprises. Building on this framework, Asheim and Isaksen [13] proposed a threefold classification of territorially embedded systems, regional network systems, and regionalized national innovation systems, thereby enhancing the empirical applicability of RIS typologies. Simmie and Martin [14] introduced the concept of path dependence into RIS research, emphasizing that the evolution of urban and regional innovation systems is strongly shaped by historical structures and existing industrial foundations. Cooke [12] likewise highlighted the dynamic coupling between regional economic structures and innovation behavior, which jointly drives processes of path lock-in and evolutionary transformation. From this perspective, evolutionary economics has become a key theoretical foundation for understanding institutional learning, organizational agglomeration, and knowledge trust mechanisms within RIS. Furthermore, Audretsch and Belitski [15] extended RIS theory from an innovation ecosystem perspective, conceptualizing it as a self-organizing system composed of diverse actors such as firms, universities, and financial institutions, within which regional institutional settings, cultural contexts, and infrastructure jointly form the ecological background supporting value co-creation. In the context of the global knowledge economy, Asheim and Coenen [16] further developed the concept of the “learning region”, arguing that a region’s capacity to absorb, integrate, and transform both codified and tacit knowledge constitutes the core driving force behind the evolution and upgrading of regional innovation systems.

Differentiated development of urban high-tech industries

Amid the accelerated concentration of top-tier scientific and technological resources in core regions, spatial heterogeneity in regional technological development has become a focal topic in technology management, regional economics, and the geography of innovation. Existing studies typically approach this issue through four lenses: output statistics, textual co-occurrence, collaboration networks, and spatial visualization. In terms of quantitative output, for example, Yaozong Zhu [17] shows via patentometric analysis that the concentration of high-value patents along the eastern seaboard significantly exceeds that of central and western regions, vividly reflecting uneven regional innovation. Yet output gaps alone cannot reveal technological structure and knowledge linkages. To further resolve regional technological features, collaboration-network analysis offers another perspective. Chen [18] uses VOSviewer to construct cross-regional keyword co-occurrence networks, with knowledge maps that display each region’s advantage topics and the strength of their linkages, demonstrating that co-word analysis effectively identifies structural differences in technology. Abbasiharofteh et al. [19] build co-inventor patent networks and find that regions with dense cross-community connections are more likely to generate breakthrough combinations of technologies, indicating that network topology substantially shapes innovation outcomes. After introducing geographic information systems to spatially project collaboration networks, Weinan Gu [20] maps the relationships among regional innovation actors onto geographic coordinates, uncovering a pattern in which innovation activities are highly concentrated within major metropolitan areas while peripheral regions exhibit relatively sparse ties.

Studies of industrial technology development based on scientometric and patent-analysis methods

With the accelerating pace of knowledge production and technological diffusion, scientometric data have become a key information source for tracing the evolution of industrial technologies and innovation trends [21, 22]. Methodologically, scholars have employed co-citation, co-word, and main-path network analysis to track the accumulation and transformation of technological knowledge. Chi-Yo Huang et al. [23], building on a co-citation network, used main-path analysis to delineate the core chains of knowledge inheritance within a field, thereby revealing the dominant routes of technological evolution. Patent data likewise underscore the value of quantitative analysis: Altuntas et al. [22] devised a technology life-cycle–diffusion model demonstrating that patent indicators can gauge technological maturity and forecast the probability of successful commercialization. Beyond elucidating evolutionary trajectories, scientific publications provide objective yardsticks for evaluating innovation capacity. Zhijing Zhu [24] argues that the volume of Science Citation Index (SCI) papers and patents, the frequency of highly cited works, and grant ratios collectively reflect the originality and dynamism of national or regional innovation. Such quantitative results are widely embedded in macro-policy and strategic planning: Gibson’s review [25] shows that scientometric analysis is increasingly integral to public policy, technology foresight, and research planning; Yaozong Zhu [26], using the geographic distribution of patents and networks of innovation actors, maps the competitive landscape among China, the United States, and Japan in key technologies and proposes recommendations for international collaboration and technological deployment. In sum, scientometrics can serve as a bridge across the “knowledge–technology–policy” chain: it not only traces the main lines of technological change but also supplies evidence for assessing innovation capacity and formulating industrial strategy. In terms of concrete techniques, scientometrics and patentometrics have developed a mature analytical paradigm. Zupic and Čater [27], via tools such as VOSviewer and CiteSpace, construct knowledge maps to reveal thematic evolution and the interaction patterns of innovation actors, thereby enabling a multidimensional, reproducible research workflow.

Synthesis

Research on differentiated trajectories of urban high-tech development and industrial technological evolution provides critical micro-level support and quantitative instruments for advancing RIS theory. From one perspective, cities exhibit relatively stable differences in key innovation dimensions, including dominant innovation actors (e.g., universities versus firms), industrial orientation (hardware-oriented manufacturing versus software- and platform-based activities), and knowledge transformation pathways (publication-driven versus patent-driven innovation). These dimensions align closely with the core logic underlying classical RIS typologies—namely grassroots, networked, and interventionist systems. By operationalizing and quantifying these variables, cities can be systematically positioned within specific RIS configurations, thereby enabling more structured and reproducible empirical testing and longitudinal tracing of RIS evolution. From another perspective, bibliometric and patent-based approaches offer effective means to uncover inter-regional knowledge flow structures and technological evolution pathways, providing high-dimensional empirical evidence for examining internal mechanisms, heterogeneous development patterns, and path-dependent dynamics within RIS.

Despite notable progress in theoretical development and methodological exploration, important limitations persist in existing research paradigms and analytical depth. On the one hand, traditional approaches—such as surveys, interviews, case studies, and aggregate statistical analysis—are constrained by sample coverage, processing efficiency, and reliance on subjective judgment, limiting their ability to capture the complexity and multi-scale nature of contemporary technological systems. These challenges are further compounded by data-related issues. Bibliometric analyses commonly depend on author-provided keywords and textual metadata; in the absence of standardized technical vocabularies, heterogeneous expressions fragment semantically identical concepts and distort frequency distributions [28]. Likewise, policy-oriented analyses often rely heavily on expert interpretation, making them susceptible to cognitive bias, information asymmetry, and experience-driven reasoning, which may introduce subjectivity into policy conclusions.

On the other hand, much of the literature on regional technological differentiation remains focused on macro-level output comparisons, with insufficient attention to specific technological categories, hierarchical knowledge structures, and collaborative relationships. The connection between quantitative measurement results and policy interpretation is frequently weak, lacking mechanisms that meaningfully integrate regional contexts, industrial structures, and technological trajectories. Consequently, the dynamic interactions among regions, industries, and technologies remain underexplored, limiting the ability of existing studies to inform finely targeted policy interventions and to support evidence-based strategies for regional collaborative innovation.

Research framework

This study adopts a co-occurrence–based analytical approach and leverages VOSviewer to construct and visualize scientific knowledge maps. These maps are combined with an expert-curated technical taxonomy to support semantic interpretation, while LLMs are introduced to enhance the precision of data processing and the interpretive coherence of results. Based on this design, we propose an integrated four-part framework—bibliometric statistics, co-word analysis, network analysis, and LLM-assisted interpretation—to systematically examine differentiated patterns and evolutionary features of urban high-tech development (the overall workflow is summarized in Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. The overall research framework.

The system takes bibliometric and patent data as input and outputs visualized networks, statistical indicators, and LLM-based semantic interpretations.

https://doi.org/10.1371/journal.pone.0348590.g001

First, guided by an expert-compiled technical taxonomy encompassing foundational, core, and frontier technologies, we collected scientific publications and patent records from 2016 to 2025 using the WOS and the national patent information platform. These data were cross-tabulated along the two dimensions of region and industry to construct a multi-source dataset suitable for comparative analysis. A central objective at this stage was to delineate technological objects and category boundaries using an authoritative lexicon, thereby ensuring that all subsequent quantitative analyses were conducted within a consistent and comparable analytical frame.

Second, data preprocessing focused on keyword cleansing and standardization to address noise and orthographic inconsistency in the raw data. Individual English tokens were first extracted from the keyword sets, and terms already covered by the expert taxonomy—including plural variants—were retained to avoid the inadvertent removal of core technical concepts. LLM-based semantic recognition was then used to identify and exclude abbreviations or model labels lacking substantive technological meaning (e.g., SVM, ResNet-50), compensating for the limitations of purely rule-based filtering. Remaining tokens were treated as non-informative and removed from the relevant metadata fields. For normalization, equivalence sets were generated for taxonomy entries so that heterogeneous expressions referring to the same concept could be merged into standardized forms. An increasing frequency threshold (starting at two occurrences) was applied to retain keywords until the candidate set was reduced to no more than, and as close as possible to, 200 terms. Equivalence merging was performed only when a concept appeared in at least two distinct lexical forms.

Third, in the modeling and measurement stage, cleansed keywords and author–affiliation information were used to construct a co-word network, an institutional collaboration network, and an author collaboration network in VOSviewer. The co-word network was used to identify topical hotspots and cluster structures, while the collaboration networks illustrated each city’s research partnerships and the organizational structure of core research groups. Building on the three-tier technical taxonomy, multidimensional bibliometric statistics were then conducted. These included an annual analysis of keyword frequencies from 2016 to 2025 to trace the structural evolution of foundational, core, and frontier technologies; an identification of the five most frequent frontier keywords over the most recent five-year period to capture short-term momentum; and an assessment of the annual proportion of frontier technologies relative to all three tiers to measure regional “frontier intensity”. Frontier intensity refers to a quantitative indicator measuring the share of frontier-level technologies among all identified technologies (foundational, core, and frontier) in a given region and time period. However, the taxonomy-based temporal analysis conducted in this stage requires a sufficient density of author keywords that can be reliably mapped onto the three-tier technical classification. Empirically, in the SC domain, keyword–taxonomy alignment was extremely sparse, resulting in limited informational content for robust time-series comparison across regions. For instance, even in Wuhan—the city with the largest publication volume in this domain—only a single frontier-level keyword (non-volatile memory) was identified over the past decade. To avoid noise amplification and over-interpretation, SC was therefore not included in this specific analytical step. Subsequent taxonomy-driven temporal analyses focus on AI, ICV, and FOC, where knowledge articulation is richer and alignment with the taxonomy ensures analytical stability and interpretability. Importantly, this stage-specific exclusion reflects data constraints rather than a reassessment of SC’s analytical relevance.

It should be noted that SC was intentionally included in the initial data collection and descriptive analysis as a technologically mature and capital-intensive industry occupying a distinct position in the technology life cycle. Its limited representation in taxonomy-based trend analysis reflects the domain’s structural characteristics—namely, a weaker reliance on diversified scientific keywords and a greater emphasis on process- and firm-specific knowledge—rather than analytical inconsistency.

Finally, LLM-based semantic support via ChatGPT-DeepResearch was used to interpret and contextualize the quantitative results. It linked observed hotspots, evolutionary trajectories, and collaboration patterns to regional industrial and policy contexts, and helped identify representative authors and research contributions. Rather than replacing quantitative evidence, LLM-assisted interpretation synthesized results across analytical layers to inform policy-relevant insights.

Overall, the framework remains firmly grounded in structured data analysis and conventional statistical procedures, while enhanced semantic coherence supports more systematic interpretation and cross-contextual reasoning. This design improves interpretive depth and explanatory clarity without intervening in the underlying statistical logic, measurement process, or reproducibility of the quantitative results.

Data Source

The dataset comprised three components: an expert-defined technical taxonomy, scientific publications, and patents. The taxonomy was developed with intellectual support from the Wuhan Science and Technology Strategy Research Institute, jointly established by the Wuhan Municipal Science and Technology Bureau and Wuhan University of Technology. The institute has established extensive collaborative relationships with eleven high-level laboratories across Hubei Province, including the National Key Laboratory of Optical Communication Technology and Networks, the Hubei Optics Valley Laboratory, and the Wuhan Institute of AI. Drawing upon domain reports produced through these collaborations, industry experts distilled the taxonomy from sectoral analyses and policy documents and organized it into a three-tier scheme—foundational, core, and frontier technologies—serving as the baseline knowledge base throughout the analytical workflow. The technical taxonomy used in this study is presented in Table 1, and the complete expert-defined taxonomy is provided in S1 File. The taxonomy functioned in two principal ways. First, it served as a standardized keyword set against which keywords extracted from the WOS were systematically matched and classified, thereby reducing subjectivity and harmonizing definitions. Second, it provided an operable hierarchical labeling mechanism that supported evolutionary analysis and cross-regional comparison. Under this mechanism, annual frequencies of the three technology tiers were compared to trace overall technological trajectories; rolling counts of frontier-level terms were conducted to identify recent hotspots; and the annual share of frontier technologies within the total of all three tiers was calculated to gauge each region’s degree of frontier intensity. In this way, the data foundation was both authoritative and reproducible, supplying quantifiable support for revealing regional differences and the cadence of technological evolution.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Distribution of the Authoritative Technical Taxonomy Across Four Industries.

https://doi.org/10.1371/journal.pone.0348590.t001

Publication data were obtained from the WOS Core Collection and restricted to SCI and SSCI journals for the period 2016–2025 to ensure scholarly authority and international comparability. The retrieval strategy prioritized WOS subject categories (WC); where WC coverage was insufficient, topic searches (TS) were constructed around four domains—AI, FOC, ICV, and SC—with review literature and the expert taxonomy used to filter keywords with unambiguous technological referents. For geographic filtering, the “author affiliation address” field was limited to Wuhan, Chengdu, Hangzhou, and Tianjin, yielding a total of 39,205 publication records across the four cities and four domains, including bibliographic data, author keywords, and author and affiliation metadata.

Patent data were drawn from the National Key Industries Patent Information Service Platform and retrieved using a two-stage strategy. In the first stage, domain-specific technical keyword queries were constructed and applied within the corresponding first-level industry categories of the patent retrieval system. In the second stage, the search results were constrained by the “address” field to restrict patent applicants to the target regions. After screening, a total of 9,945 patent records meeting the criteria were retained for the four industries across the four cities.

LLM selection and roles

During data processing, LLMs were introduced as auxiliary tools to support keyword cleaning, normalization, and interpretive analysis. It should be emphasized that this study does not treat different LLMs as independent analytical instruments for comparison. Instead, they are employed as language-processing interfaces assigned to specific tasks at different stages of the workflow. Their use does not affect subsequent bibliometric analysis or statistical modeling, nor does it introduce differences attributable to model choice. Because the substantive conclusions are derived from the bibliometric patterns, network structures, and the cleaned keyword system under fixed analytical procedures, rather than from any comparison between LLMs, this task-specific division of labor does not alter the study’s core findings.

For reasons of general applicability and accessibility, DeepSeek—a widely used AI model in China—was selected for keyword denoising and normalization. In the keyword-processing stage, DeepSeek (deepseek-chat) was accessed via an API to perform tasks with relatively high structural constraints and explicit rule definitions. These tasks included identifying and removing abbreviations or shorthand expressions lacking clear technological content, as well as assisting in the generation of equivalent expressions for terms in the expert-defined technical taxonomy. This API-based approach is well suited for batch processing and repetitive single-item judgments, ensuring consistency and reproducibility in keyword cleaning and normalization. By contrast, the construction of equivalence-mapping tables for bibliographic keywords was conducted through an interactive web-based interface. This task requires integrated semantic judgment across multiple alternative expressions referring to the same concept and the generation of mappings that remain interpretable to human readers, making it unsuitable for simple batch API calls.

In addition, the ChatGPT-DeepResearch function was employed to support semantic interpretation and analytical writing. Deep Research is a multi-step analytical agent integrated into ChatGPT that performs staged information retrieval and synthesis by combining model-internal knowledge with cross-document and online information sources, rather than relying solely on fixed training data. When additional context or clarification was required, Deep Research retrieved relevant information from diverse sources, including news reports, academic literature, and statistical data, and synthesized insights through planned multi-step reasoning processes. Public benchmark evaluations indicate that this function exhibits strong performance in complex reasoning tasks. For example, in the Humanity’s Last Exam benchmark released in February 2025, Deep Research achieved an accuracy of 26.6%, outperforming the second-ranked OpenAI o3-mini (High) model, which scored 13.0%. Compared with conventional LLM interactions that yield brief responses, Deep Research typically produces structured long-form outputs organized by sections and supplemented with traceable source references. Accordingly, it is well suited for extending bibliometric findings with semantic interpretation and for providing multi-perspective analytical support in discussions of urban high-tech industrial differentiation and policy implications.

Data processing

The keywords used in this study were batch-exported by VOSviewer from WOS records, including both author-provided keywords and indexed keywords. To reduce contingency and noise, unless otherwise specified, the analysis was restricted to keywords with an occurrence frequency of at least two, thereby ensuring a minimum level of scholarly recognition and representativeness. On this basis, keyword denoising and normalization were performed.

For denoising, close inspection of the raw data showed that non-informative or non-technical terms were predominantly single English tokens (e.g., model). We first protected terms already included in the expert taxonomy (and their plural forms), then used the semantic recognition capability of LLMs (prompt available in S2 File) to identify and remove abbreviations or model labels lacking substantive technical content (e.g., SVM, ResNet-50). Remaining single-word tokens were uniformly deleted (synchronously in the ID and DE fields), yielding a cleaner candidate set.

For normalization, to address frequency fragmentation caused by synonymous variants, we first had the LLMs (prompt available in S3 File) pre-generate equivalence sets for taxonomy entries, covering symbol splitting, hyphen/space alternation, singular–plural conversion, US/UK spelling variants, and acronym/full-form parenthetical combinations. We then applied an increasing threshold T (starting at 2) to filter high-frequency candidates, capping the set at no more than—and as close as possible to—200 entries. Next, we aggregated synonyms using a “first-occurrence priority” rule to fix the canonical label and produce a JSON-structured equivalence map (prompt available in S4 File). This “rules + LLMs” hybrid strategy balanced standardization and variant coverage, providing a reproducible, low-noise keyword basis for subsequent analyses.

Table 2 compares changes in high-frequency keyword rankings before and after denoising and normalization. The results show that, prior to processing, the most frequent keywords were dominated by generalized methodological terms such as model, algorithm, and system. Although these terms appeared frequently, they carry weak technological specificity and are of limited value for characterizing concrete technological pathways. If such raw keyword sets were directly used for VOSviewer-based co-occurrence analysis, the resulting network structure would be dominated by methodological noise, reducing interpretability and potentially obscuring keywords with genuine technological significance.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Comparison between traditional keyword statistics and LLM-assisted normalization.

https://doi.org/10.1371/journal.pone.0348590.t002

In addition, the coexistence of synonymous variants and formal alternatives substantially fragmented keyword frequencies, leading to an underestimation of important technical concepts in the original rankings. For example, in the unnormalized list, neural-networks, neural networks, and neural-network appeared 152, 111, and 110 times, respectively, each ranking outside the top ten when counted separately. After equivalence merging, however, the aggregated frequency of this concept increased markedly, elevating its rank to second overall. This comparison demonstrates that LLM-assisted denoising and normalization not only reduce interference from irrelevant high-frequency terms but also effectively restore the structural importance of key technological concepts in co-occurrence analysis.

By contrast, after automated processing using the “rules + LLMs” hybrid strategy, the overall keyword ranking structure changed substantially. On the one hand, generalized terms lacking clear technological content were significantly down-ranked or removed. On the other hand, concrete technical units (e.g., CNN, genetic algorithm) and functional technological concepts (e.g., feature extraction, attention mechanism) occupied more central positions in the high-frequency list. These results indicate that the introduction of LLMs does not replace statistical analysis; rather, it enhances the semantic discriminability and structural consistency of keywords, thereby providing a more reliable technological-semantic foundation for subsequent co-word modeling and network analysis.

Analytical methods

Differentiated Analysis and Interpretation.

Building on the VOSviewer-based network modeling of keywords, authors, and institutions, as well as the quantitative statistics of the three categories of technological keywords derived from the technical taxonomy, this study further employs the Deep Research function within ChatGPT to conduct differentiated analysis and integrative interpretation. Fig 2 illustrates the prompt-engineering workflow used for this stage, and the complete prompt is provided in S5 File. This stage is grounded in the bibliometric results and focuses on identifying and explaining structural differences across cities in terms of technological configuration, research hotspots, and collaboration structures through systematic synthesis and comparison.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. ChatGPT Prompt Engineering.

https://doi.org/10.1371/journal.pone.0348590.g002

Specifically, the analysis examines the performance of Wuhan, Hangzhou, Chengdu, and Tianjin across AI, FOC, ICV, and SC, and organizes the findings in a city-specific and industry-specific structure. The resulting conclusions are developed along four dimensions: first, summarizing the relative strengths and weaknesses of each city in technological layout, hotspot evolution, and collaboration patterns; second, identifying logical linkages and structural differences through cross-city comparison; third, discussing how regional industrial policies and support environments influence technological development trajectories and the growth paths of research actors; and fourth, selecting representative authors or institutional research outputs as illustrative cases to substantiate the academic impact and practical relevance of regional scientific production.

On this basis, the study derived policy implications for each of the four cities through a structured, city-by-city interpretive process. Policy discussions were organized in a “city-specific subheading plus narrative” format and were explicitly grounded in the Regional Differentiation Summary for each city. Rather than presenting standalone recommendations, each policy discussion was developed as a reasoned extension of the empirically observed patterns and RIS-based interpretations, linking identified structural constraints and opportunities—such as imbalances in innovation actor configurations, uneven technological evolution, or weak collaboration mechanisms—to corresponding policy leverage points. By maintaining this explicit linkage between empirical analysis, theoretical interpretation, and policy reasoning, the study established a coherent analytical-to-policy logic that ensures internal consistency between findings and policy implications.

Temporal configuration of scientific publications and patents

To systematically characterize the relationship between scientific knowledge production and technological application across cities in high-tech industries, this study integrates publication and patent data over the period 2016–2025 from a temporal perspective. Publication output reflects the accumulation of scientific knowledge in specific technological fields, whereas patent activity represents the advancement of technological application and industrialization. Examining the two within a unified temporal framework facilitates the identification of how different cities configure the timing and pace of research expansion and technological application within the same industry, as well as the heterogeneity of these configurations.

As shown in Fig 3, most industries exhibit a sustained increase in publication output during the study period. In particular, after 2019, research output in most cities entered a phase of accelerated growth. In contrast, patent activity shows markedly different temporal response patterns across cities. Rather than following publication growth in a synchronized manner, the relationship between research expansion and technological application displays diverse temporal configurations, indicating that the two processes are not mechanically aligned.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Temporal evolution of publication and patent outputs across cities and industries.

Literature data were obtained from the WOS Core Collection and Patent data were drawn from the National Key Industries Patent Information Service Platform.

https://doi.org/10.1371/journal.pone.0348590.g003

In the AI domain, all four cities experienced substantial growth in scientific publications, but the temporal relationship between publications and patents differed markedly across cities. In Hangzhou, patent activity remained relatively synchronized with the expansion of scientific output, with patent growth broadly keeping pace with increasing publication volume. By contrast, in Wuhan and Chengdu, patent growth generally lagged behind publication expansion, suggesting a temporal pattern in which scientific production preceded technological application. Tianjin, while smaller in overall scale, showed more stage-specific fluctuations. Taken together, these patterns suggest that cities differ considerably in how research outputs and technological application are temporally aligned in frontier technologies.

A different pattern appeared in the SC domain, where the publication–patent relationship showed characteristics distinct from those of the other industries. Publication output remained relatively limited and grew at a moderate pace, whereas patent activity in some cities was more concentrated in specific periods. Notably, Wuhan repeatedly recorded patent counts exceeding publication output during the study period, suggesting a stronger orientation toward technological application than academic publication. Other cities, by contrast, showed more limited patent growth and weaker temporal coupling between publications and patents. This pattern suggests that, in a technologically mature and capital-intensive industry such as storage chips, inter-city differences in the balance between research input and technological application may be more pronounced.

The ICV domain showed broadly similar trajectories in publication growth across cities, but patent activity diverged more clearly in both timing and intensity. Wuhan and Chengdu displayed relatively strong patent activity in the earlier stages of the study period, followed by gradual stabilization, whereas Hangzhou showed more pronounced acceleration in the middle and later stages. Tianjin, despite its smaller overall scale, also concentrated its patent growth peaks in the later period. These differences point to varying rhythms of technological exploration and industrialization across cities and suggest that the development paths of growth-stage industries are not temporally synchronized.

As a more mature industry, FOC showed a relatively stronger coupling between publications and patents over time, although structural differences across cities remained evident. Wuhan consistently maintained high levels of both scientific output and patent activity throughout the study period, suggesting a relatively stable relationship between research and application. Other cities showed varying degrees of divergence in the scale and volatility of patent growth, indicating heterogeneous temporal configurations.

Further evidence comes from the annual evolution of the patent-to-publication configuration ratio. As shown in Fig 4, emerging and growth-stage industries such as AI and ICV displayed greater temporal volatility in this ratio, with cities differing markedly in the timing between research expansion and technological application. By contrast, in more mature industries such as FOC and SC, the ratio remained relatively stable overall, although structural differences across cities persisted. Notably, although publication output in most industries rose simultaneously across cities in 2025, the timing and magnitude of patent responses still varied substantially, suggesting a continuing non-synchronous relationship between research expansion and technological application.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Temporal patterns of patent–publication ratios across cities and industries.

https://doi.org/10.1371/journal.pone.0348590.g004

Overall, inter-city differences in high-tech development are reflected not only in the scale of scientific or patent outputs, but more importantly in how these outputs are temporally configured and evolve over time. Such temporal configuration differences provide a critical perspective for understanding differentiated trajectories of urban high-tech development.

City-level research hotspots in high-tech industries

By analyzing co-occurring author keywords, we identified city- and industry-specific research hotspots, yielding a clearer view of each region’s research orientations and technological trends. The ten most frequently occurring keywords are listed in Table 3, and the co-occurrence networks of author keywords across the four cities and four technological domains are illustrated in Fig 5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Author keywords.

https://doi.org/10.1371/journal.pone.0348590.t003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Co-occurrence network of author keywords.

The figure covers four cities (Wuhan, Hangzhou, Chengdu, and Tianjin) and four technological domains (AI, FOC, ICV, and SC), illustrating their respective research hotspots and interconnections.

https://doi.org/10.1371/journal.pone.0348590.g005

Across AI, all four cities converged on core algorithmic topics such as “deep learning” and “neural network”, indicating sustained attention to foundational methods. Their emphases nonetheless differed: Wuhan gave greater prominence to swarm-intelligence approaches such as “particle swarm optimization” and “differential evolution”; Hangzhou focused more on emerging architectures, especially “graph neural network” and “Transformer”; Chengdu highlighted “group decision-making” and “aggregation operator”, suggesting a distinctive concentration in multi-agent decision-making; and Tianjin emphasized “reinforcement learning” and “deep reinforcement learning”, pointing to an orientation toward dynamic decision-making and adaptive learning.

In SC, “atomic layer deposition” appeared as a shared core across cities, underscoring its importance in semiconductor manufacturing. Wuhan placed more emphasis on “phase-change memory” and “non-volatile memory”; Hangzhou on “through-silicon via” and “electrical characteristics”; Chengdu on “resistive switching” and “logic gate”; and Tianjin on “ReRAM” and “NAND flash”, indicating differentiated attention to storage, packaging, device-level fundamentals, and application-oriented memory technologies.

A similar pattern appeared in ICV, where “model predictive control” and “autonomous driving” were common hotspots. Wuhan emphasized “autonomous vehicles” and “internet of vehicles (IoV)”; Hangzhou highlighted “predictive model” and “voltage control”; Chengdu focused on “resource allocation” and “computational modeling”; and Tianjin gave more prominence to “real-time system” and “deep reinforcement learning”, reflecting different orientations in connectivity, system modeling, optimization, and responsive control.

In FOC, all four cities shared attention to “fiber-optic sensor” and “optical sensor”, highlighting sensing as a key technological focus. Wuhan emphasized “fiber Bragg grating” and “refractive index”; Hangzhou concentrated on the “Mach–Zehnder interferometer” and “optical switch”; Chengdu highlighted “free-space optical communication” and “visible light communication”; and Tianjin focused on “photonic crystal fiber” and “fiber polarization”.

Future Development Trends in Urban High-Tech Industries

By tallying author keywords mapped to the three-tier technical taxonomy and analyzing their temporal trajectories, we delineated regional technological pathways and potential inflection points across industries, thereby providing objective, quantitative support for trend assessment. The temporal evolution of AI, ICV, and FOC across the four regions is illustrated in Figs 6–9.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Technological trend chart of the three technology categories in Wuhan.

https://doi.org/10.1371/journal.pone.0348590.g006

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Technological trend chart of the three technology categories in Hangzhou.

https://doi.org/10.1371/journal.pone.0348590.g007

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Technological trend chart of the three technology categories in Chengdu.

https://doi.org/10.1371/journal.pone.0348590.g008

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Technological trend chart of the three technology categories in Tianjin.

https://doi.org/10.1371/journal.pone.0348590.g009

For AI, all four cities had broadly shifted from a convolution-centered paradigm toward attention mechanisms and LLMs, with frontier shares rising to around 19%, except Tianjin, which remained slightly lower (16%). Differences were mainly visible in frontier composition and application orientation. Hangzhou moved earlier and more deeply into LLMs and multimodality, with greater emphasis on Transformer engineering and scenario-based deployment, and ranked highest in both the number and share of frontier keywords. Wuhan followed a dual trajectory of foundational and frontier development, with fewer LLM-related terms and relatively greater emphasis on theoretical modeling and visual perception. Chengdu showed frequent attention-mechanism terms and, although it lagged behind Hangzhou in large-model and cross-modal engineering, displayed a more concentrated frontier profile in decision-making and representation optimization. Tianjin remained smaller in scale but explored differentiated directions such as brain–computer interfaces and artificial general intelligence (AGI).

In the ICV domain, all four cities advanced within an integrated perception–decision framework, although the overall frontier level remained relatively low (all ≤ 14% in 2025). Their pathways differed in pace and emphasis. Hangzhou integrated frontier algorithms such as Transformer more extensively into autonomous-driving perception and decision modules and combined this with open-road testing; frontier terms such as “LiDAR” and “Transformer” appeared more frequently than in peer cities. Wuhan focused on advanced driver-assistance systems, IoV security, and LiDAR, with frontier development emerging later and showing a stronger emphasis on safety and verifiability. Chengdu emphasized communication–control integration, including on-board communications and edge-computing scheduling, and showed activity in verifiable modules such as steer-by-wire, while maintaining active international collaboration. Tianjin progressed mainly through standards and testing systems, with more limited incorporation of frontier algorithms, suggesting a steadier engineering-oriented pathway.

In FOC, all four cities shared a common trend toward advances in photodetector technologies, but their development routes differed. Wuhan maintained the strongest advantages in both scale and breadth, with a wider frontier portfolio that also extended into inter-satellite communication. Hangzhou followed a more focused strategy, concentrating on satellite laser communication and quantum random numbers, and recorded the fastest rise in frontier share. Chengdu remained comparatively strong in core-technology areas such as transmission and network architecture but showed less activity in frontier fields, leaving its frontier level below Wuhan and Hangzhou. Tianjin appeared more application-oriented, focusing on fiber-optic sensing and optical-fiber delay-line scenarios, with a limited frontier keyword count and a more gradual engineering-led trajectory.

Analysis of Collaboration Patterns in Urban High-Tech Industries

In the evolution of regional industrial technology, research collaboration serves as a key mechanism for catalyzing knowledge diffusion and concentrating innovation, exerting far-reaching effects on capability trajectories and regional competitive advantage. To systematically uncover the collaboration structures and network properties of cities within high-tech industries, we conducted analysis at two levels: the organizational level—using organizational co-occurrence to examine inter-institutional coordination and regional collaboration intensity—and the individual level—using author co-occurrence to depict cooperation structures and the characteristics of core research groups.

Organizational Co-Occurrence Analysis

From a RIS perspective, organizational collaboration structures reflect the configuration of dominant R&D actors and the institutional logic through which knowledge is produced and mobilized. By examining organizational co-occurrence networks, this analysis identifies which actors—universities, research institutes, enterprises, or state-affiliated organizations—occupy central positions across industries and cities, thereby revealing distinct RIS configurations in practice.

Analyzing the co-occurrence networks of collaborating organizations thus clarifies each city’s principal partners, tie strengths, and overall collaboration structure. The top five organizations by document count are reported in Table 4, and Fig 10 visualizes the organizational co-occurrence networks across the four cities and technological domains.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Collaborating organizations.

https://doi.org/10.1371/journal.pone.0348590.t004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Co-occurrence networks of industrial collaboration organizations across four cities and four technological domains.

https://doi.org/10.1371/journal.pone.0348590.g010

By analyzing the co-occurrence network of collaborating organizations, Wuhan relied heavily on leading local universities across the four industries. Wuhan University and Huazhong University of Science and Technology occupied core positions in AI and SC, while Wuhan University of Technology and China University of Geosciences sustained influence in ICV and FOC. This pattern indicated a collaboration model centered on local flagships, in which “inter-university alliances + research-institute cooperation” drove industrial progress. As shown in Wuhan’s organizational co-occurrence map (Fig 10), Huazhong University of Science and Technology and the Chinese Academy of Sciences (UCAS) maintained close ties with Hong Kong Polytechnic University, Nanyang Technological University, the University of Sydney, and other international universities, and the overall network exhibited pronounced clustering. Different colors denoted relatively stable collaboration communities: in AI, a green cluster was centered on Wuhan University, linking the central-China region and selected international universities; a red cluster centered on Huazhong University of Science and Technology, connecting key national universities and research institutes; and the blue and orange clusters included UCAS and China University of Geosciences, engaging in cross-disciplinary and cross-domain collaboration. This clustered structure reflected both division of labor and complementarity among institutions, as well as diversity and differentiation of collaboration themes. Overall, Wuhan exhibits a university- and research-institute-led RIS, in which local flagship universities function as stable knowledge anchors, while national institutes and international partners extend the system’s scientific depth and global connectivity.

In Hangzhou, collaboration was led by Zhejiang University and spanned AI, SC, ICV, and FOC, forming a cross-industry, broad-participation pattern. Hangzhou Dianzi University, Zhejiang University of Technology, and China Jiliang University provided strong support in FOC and ICV and maintained stable ties with top domestic research bodies. Some teams actively expanded international exchanges, facilitating the wide circulation and translation of research results. As Fig 10 further shows, in AI, Zhejiang University and Hangzhou Dianzi University not only occupied core positions but also maintained close cooperation with enterprises such as Alibaba Group and Ant Group; collaborations with Nanyang Technological University and the University of Sydney enhanced openness and frontier orientation. In FOC, Zhejiang University and Zhejiang Lab led research while linking Westlake University, the Chinese Academy of Sciences, and Hong Kong Polytechnic University, forming a dual track of “established research bodies + emerging research forces.” In ICV, local universities were tightly connected and maintained strong ties with Tsinghua University, Beihang University, and other key national institutions, highlighting an engineering orientation and industrialization potential. Collaboration in SC was relatively localized while maintaining exchanges with select international partners, reflecting steadiness alongside outward extension. Hangzhou’s collaboration structure reflects an industry-integrated RIS, where leading universities remain central knowledge producers, while major platform firms are deeply embedded in research networks, enabling strong market feedback and frontier-oriented knowledge translation.

In Chengdu, the organizational co-occurrence network was anchored by the University of Electronic Science and Technology of China and Sichuan University, which wielded strong influence in AI, SC, and FOC; Southwest Jiaotong University (SWJTU)—the “cradle of China’s railway engineers”—played a central role in ICV. Participation by Southwestern University of Finance and Economics, Chengdu University of Information Technology, and other institutions gave Chengdu’s network a multi-layered, composite character. This model strengthened the city’s research base in information, communications, and intelligent technologies while enabling cross-disciplinary innovation. Across the four networks, the AI graph showed three primary nodes bridged by economics/management and information-science universities, indicating interdisciplinary synergy. In FOC, UCAS was deeply embedded and formed clusters with Swinburne University of Technology, ITMO University, and other international partners—an open pattern of “national capacity + international channels.” In ICV, SWJTU occupied the transportation core and, together with local strengths, formed joint linkages of “mechanism-based control × data-driven methods.” In SC, UCAS and enterprise nodes such as Huawei Technologies were directly embedded, underscoring “university–institute–enterprise” collaboration. Overall, Chengdu’s pattern could be summarized as “aggregation around local flagships—reinforcement by national institutes—extension to international partners—enterprise nodes embedded”, ensuring both scale and depth while providing a robust, open basis for cross-disciplinary collaboration and technology transfer. Chengdu represents an academically strong but weakly coupled RIS, characterized by solid university-led research capacity, partial integration of national institutes, and a growing yet still fragmented linkage to industrial actors.

In Tianjin, research collaboration was primarily led by Tianjin University and Nankai University, which together covered core directions across the four industries. The co-occurrence maps showed stable links with national-level bodies such as UCAS and Beihang University, alongside partnerships with Hong Kong Polytechnic University, the University of Oxford, the University of Tokyo, and other international universities—evidence of cross-regional and international orientation. In ICV, Tianjin University maintained close ties with industry institutions such as the China Automotive Technology and Research Center, reflecting a strong application focus. In FOC and SC, Hebei University of Technology and Tianjin University of Technology further strengthened intersections with engineering and materials research. Overall, Tianjin’s collaboration model combined the advantages of local university clusters with cross-regional connectivity, balancing academic inquiry with industry demand in an application-oriented approach. Tianjin’s organizational network aligns with a policy-oriented RIS, in which government-supported universities and national research bodies coordinate innovation activities in strategically prioritized and application-driven domains.

Author Co-Occurrence Analysis

By examining the co-occurrence networks of authors, we further identified clear inter-city differences in the micro-structure of regional innovation networks and gained deeper insight into each city’s research layout and comparative strengths across the four domains. Table 5 presents the top five authors ranked by document count, and Fig 11 illustrates the corresponding author co-occurrence networks across the four cities and four technological domains.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Collaborating authors.

https://doi.org/10.1371/journal.pone.0348590.t005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Co-occurrence networks of industrial collaboration authors across four cities and four technological domains.

https://doi.org/10.1371/journal.pone.0348590.g011

In Wuhan, the AI and FOC networks appeared relatively dense and hierarchical, with a small number of highly central authors surrounded by many smaller contributors. In AI, Zeng, Zhigang ranked first with 246 papers and 8,916 citations, while Ma, Jiayi recorded 15,835 citations from 139 papers, suggesting high impact among leading scholars. In FOC, although the total output was much lower than in AI, the network remained closely connected, with Liu, Deming standing out at 210 publications. By contrast, ICV and SC showed a more multi-community structure, with stronger intra-group ties and weaker cross-group integration. Several authors in these two fields combined relatively modest output with comparatively high citation impact, suggesting the importance of original or application-oriented contributions.

A different pattern appeared in Hangzhou, where collaboration was less concentrated around a single core and more shaped by multiple medium-sized centers and bridging authors. In AI, Zheng, Guang occupied a central position with 99 papers and 5,784 citations, but no single-author dominance was evident. Authors such as Yang, Yi, Li, Xi, and Wu, Fei formed a relatively stable collaborative group, each with around 75 papers and more than 2,100 citations, indicating a comparatively balanced leadership structure. In FOC, Dai, Daoxin and Shi, Yaocheng anchored the main core with 109 and 77 papers, respectively, while cross-cluster ties linked materials, devices, and systems. The ICV and SC networks were also relatively weakly centralized, with multiple specialized communities and a visible degree of international collaboration. Overall, Hangzhou’s author collaboration structure was characterized by multi-core coordination and inter-cluster bridging.

In Chengdu, all four domains showed a multi-community and weak-centralization pattern, with several clusters of comparable size connected by a limited number of bridging authors. In AI, Xu, Zeshui stood out with 362 papers and 11,840 citations, the highest publication count among the four cities, while authors such as Shen, Heng Tao played linking roles across clusters. In FOC, a core–periphery belt emerged: Qiu, Kun sat at the center, connecting device-oriented teams (e.g., Yan, Lianshan; Pan, Wei) with application-focused micro-clusters through long ties. In ICV, the international scholar Rodriguez, Jose appeared again with differentiated collaboration paths: on the Chengdu side, the focus was on method-level innovation—leveraging FCS-MPC, RL, data-driven approaches, ESO, and event-triggered control to enhance robustness and performance—whereas on the Hangzhou side, emphasis lay on system-level deployment (e.g., VSG in islanded AC microgrids), using MPC and fuzzy-MPC to improve frequency and voltage dynamics and validate engineering feasibility through HIL. Both operated within an MPC framework yet complemented each other in control layer, technical levers, and validation settings. Notably, his Chengdu-side paper “Model-Predictive Control of Multilevel Inverters: Challenges, Recent Advances, and Trends” had 105 citations—substantially higher than the Hangzhou-side “Predictive Control of Voltage Source Inverter: An Online RL Solution” (39)—indicating deeper impact within Chengdu collaborations. In SC, multiple stable circles coalesced around Sun, Bai, with engineering-oriented teams such as Zhao, Yong forming tightly coupled, stable structures. Taken together, Chengdu’s author network can be described as a parallel multi-core structure supported by a small number of bridging nodes.

Tianjin showed a somewhat different profile. In AI, Cheng, Ming-Ming published 72 papers that received 10,026 citations, averaging about 140 citations per paper, the highest such figure among the four cities and fields, suggesting strong scholarly visibility despite a smaller output scale. By comparison, Chen, Shengyong published slightly more (73 papers) but with only 1,134 citations, further underscoring Cheng’s quality edge and the need for regional-level support. However, SC and ICV remained more limited in scale, with relatively balanced author distributions and multiple small communities rather than a dominant core. This produced a looser collaboration structure with weaker overall cohesion than in the other cities. Overall, Tianjin’s author collaboration pattern appeared more fragmented and weakly integrated, especially outside AI.

Differentiated analysis and interpretation of regional industrial innovation pathways

Given the breadth and depth of the analysis generated by ChatGPT-DeepResearch, this section presents only the principal conclusions and core viewpoints in summary form, distilling the main findings on regional differentiation, innovation pathway characteristics, and policy-relevant implications; full details are provided in the Appendix. The complete LLM-based regional industrial and technological differentiation analysis can be found in S6 File. All supplementary information is supported by traceable references (peer-reviewed literature, industry reports, policy documents, or database records) and is documented in the Appendix with citation markers and retrieval protocols to ensure transparency and replicability. This summary distills only the most representative conclusions along the established evidence chain; each statement maps back to corresponding bibliometric and network-analytic evidence in the main text/Appendix, avoiding interpretations beyond the data’s scope. Readers may use the Appendix source list and citation markers to locate specific city–industry evidence items (keywords, representative studies, and network-structure features) for verification.

Before turning to the city-level analysis, we briefly outline the differentiated regional innovation pathways observed across the four cities. Wuhan is characterized by a research- and state-coordinated system; Hangzhou by a platform- and firm-led, market-oriented system; Chengdu by a science-based, policy-supported system; and Tianjin by a transitional, application-oriented system.

Main contributions

Grounded in the RIS perspective, this study contributes to the literature by advancing a unified, data-driven framework for examining inter-city heterogeneity in high-tech development. The framework operationalizes key RIS components—dominant R&D actors, knowledge production and transformation mechanisms, and institutional collaboration structures—through the integrated use of bibliometric statistics, co-word analysis, collaboration network modeling, and LLM-assisted semantic interpretation. By embedding these elements within a single analytical workflow, the study provides a reproducible approach for empirically investigating how different urban innovation systems organize scientific research, technological application, and cross-actor coordination across industries, thereby moving beyond purely descriptive comparisons.

Methodologically, the study introduces a scalable pipeline for processing large-scale textual innovation data in RIS-oriented research. Rule-based procedures are combined with the capabilities of large language models to construct a standardized and noise-reduced keyword system, improving the reliability of hotspot identification, temporal trend analysis, and cross-regional comparison. The incorporation of a three-tier technical taxonomy—foundational, core, and frontier technologies—enables the differentiation of technological roles within innovation systems, allowing structural change and frontier orientation to be quantitatively assessed rather than inferred qualitatively.

Empirically, the study provides comparative evidence on how distinct RIS configurations shape differentiated development trajectories across Wuhan, Hangzhou, Chengdu, and Tianjin in AI, FOC, ICV, and SC. Joint analysis of publications, patents, and organizational collaboration networks reveals how cities rely on different combinations of universities, research institutes, firms, and policy instruments to drive innovation. The LLM-assisted interpretive layer situates these quantitative patterns within regional industrial and policy contexts, facilitating the translation of empirical findings into policy-relevant insights while preserving the integrity of statistical inference.

Overall, these contributions strengthen the link between RIS theory and empirical measurement, enhance the interpretive depth of regional innovation analysis, and offer a transferable framework for studying urban technological differentiation across regions and sectors.

Scholarly and Practical Significance

This study yields both theoretical and practical significance for research on regional industrial and technological development. From a scholarly perspective, the integrated framework—combining bibliometric statistics, co-word analysis, collaboration network modeling, and LLM-assisted interpretation—extends existing approaches in scientometrics and knowledge mapping by enabling a more structured and interpretable analysis of regional innovation systems. The introduction of LLMs into data preprocessing improves the reliability of keyword-based analysis through systematic removal of non-informative terms, abbreviation recognition, and synonym normalization, addressing long-standing challenges in bibliometric studies related to semantic fragmentation and interpretive ambiguity. Rather than replacing conventional quantitative methods, this human–AI collaborative workflow enhances analytical consistency and semantic coherence, illustrating a feasible pathway for integrating LLMs into empirical innovation research.

At the same time, the three-tier technical classification of foundational, core, and frontier technologies provides a replicable semantic instrument for examining technological evolution within and across regions. By distinguishing different functional roles of technologies, this classification enables a more precise depiction of hierarchical structures and dynamic transitions in innovation systems, thereby strengthening the analytical linkage between technological change and RIS theory.

From a practical standpoint, the cross-regional and cross-industry analyses reveal differentiated configurations and evolutionary trajectories of high-tech development, offering evidence-based insights for regional policy design. By identifying frontier technologies and tracking their temporal dynamics, the framework supports forward-looking industrial planning and helps policymakers mitigate homogeneous competition and inefficient resource allocation. Moreover, the LLM-assisted interpretive layer translates complex patterns embedded in research hotspots and collaboration networks into policy-relevant insights, enhancing the accessibility and applicability of empirical findings for science-and-technology governance and strategic decision-making.

Limitations

Despite the systematic exploration undertaken in this study, several limitations should be acknowledged. First, the data sources are primarily drawn from the Web of Science and the national patent information platform. While these datasets provide authoritative and internationally comparable coverage of scholarly publications and formal technological outputs, they do not fully capture gray literature, conference proceedings, firm-level R&D reports, or unpublished patents. As a result, certain innovation activities—particularly those embedded in enterprises or informal knowledge channels—may be underrepresented, potentially introducing partial coverage bias.

Second, the construction of retrieval strategies inevitably influences the scope of the dataset. In domains where fine-grained and universally accepted classification schemes are lacking, keyword-based queries may either omit relevant records or introduce cross-domain noise. Although expert taxonomies and iterative refinement were applied to mitigate this issue, some degree of incompleteness or contamination in the retrieved corpus cannot be entirely ruled out.

Third, bibliometric and network-based approaches primarily capture output volume, co-occurrence patterns, and structural relationships. While these methods are effective for identifying technological distributions, collaboration structures, and evolutionary trends, they are less capable of directly assessing research quality, technological novelty, or real-world economic impact. Consequently, the indicators used in this study should be interpreted as proxies for innovation activity rather than comprehensive measures of technological performance.

Fourth, although large language models were introduced to support keyword standardization and semantic interpretation, their outputs may still be influenced by training data, prompt design, and inherent semantic ambiguity. LLM-assisted processing may therefore introduce residual noise or bias, underscoring the importance of treating such tools as complementary aids rather than autonomous analytical agents.

Finally, the proposed framework places greater emphasis on structural patterns and temporal trends than on explicitly modeling dynamic interactions among policy environments, industrial value chains, and market demand. This limits the framework’s capacity to capture short-term shocks or feedback mechanisms, and future research could benefit from integrating policy dynamics, firm-level data, and market indicators to enhance explanatory power and predictive relevance.