ABSTRACT. Building models to improve the predictive effectiveness of technology convergence and providing innovative perspectives and methods for technology convergence research. First, the Apriori algorithm is used to model co-occurrence relationships, effectively identifying diverse and directed technology fusion patterns and evolution paths. Second, link prediction models based on five graph neural network algorithms are constructed to extract node and topological features of the technology fusion network, enabling the prediction of fusion relationships. Using artificial intelligence technologies as empirical examples, the link prediction model based on GraphSAGE-GCN achieved an AUC of 0.84, and the consistency between the prediction results and actual data reached 0.7, thereby identifying emerging, strengthening, and declining technology links. Data and feature dimensions need improvement, high-order association mining is insufficient, and dynamic modeling needs further refinement. The GraphSAGE-GCN algorithm achieved the best overall performance in this link prediction task, demonstrating its advantage in uncovering high-value potential fusion opportunities.

ABSTRACT. Extracting deep technical intelligence from massive, unstructured scientific literature remains a challenging task for traditional bibliometric methods and topic models. This study proposes a framework coupling large language models (DeepSeek V3.2 and Qwen3-max) with domain knowledge graphs to transform unstructured text into computable knowledge. Knowledge is extracted via a constrained strategy to mitigate generative hallucinations, and the output is structured into fine-grained entity-relation triplets to support quantitative reasoning. The framework provides three representative analytical approaches: identifying latent technology pathways via standardized residuals, tracking milestone performance breakthroughs through heterogeneous indicator normalization, and profiling institutional competitiveness using integrated scale-quality metrics. It is validated through a case study on all-solid-state lithium batteries, achieving a mean score of 4.50/5.00 in expert evaluation. This methodology represents a systematic, automated, and scalable route to obtaining deep semantic insights and strategic forecasts from scientific literature.

ABSTRACT. Understanding technology evolution from the perspective of problems can reveal how core challenges shift over time and what drives technological progress. However, existing methods face two limitations: problem representations based on SAO structures or sentence-level extraction lack standardization for cross-document alignment, and citation networks reflect document relationships rather than problem inheritance. This study proposes a framework combining TRIZ contradiction parameters with hypergraph structures to analyze technology evolution. Large language models are employed to extract contradiction parameter pairs from patent texts, transforming technical problems into standardized representations. A dynamic hypergraph is then constructed with patents as hyperedges and contradiction components as nodes, preserving the multi-element relationships within each invention. Temporal analysis of the hypergraph identifies core problems across different periods and tracks shifts in problem focus. The framework is applied to the fuel cell vehicle domain, revealing how problem focus has migrated across different technological stages.

ABSTRACT. This research-in-progress examines how inventive strategy patterns reflected in patents relate to functional creativity and how both relate to patent value. We use TRIZ as an ex post taxonomy to map patented technical solutions to the 40 inventive principles and operationalize functional creativity using the Creative Solution Diagnosis Scale (CSDS), which emphasizes usefulness and effectiveness in addition to novelty. Because manual TRIZ and CSDS coding does not scale to large patent corpora, we implement an LLM-based pipeline that summarizes the novel inventive steps and then produces TRIZ principle scores and CSDS item ratings from this summary.

We test associations between (i) the number of TRIZ principles in a patent and functional creativity, (ii) the number of TRIZ principles and patent value indicators, and (iii) functional creativity and patent value indicators, using regression models with firm and issue-year fixed effects and patent-level controls. Patent value is measured using three commonly used proxies: a market-based value measure (KPSS), forward citations in a fixed five-year window, and renewals at 3.5 years.

Preliminary results show that patents mapped to a larger number of TRIZ principles tend to receive higher functional creativity scores. Functional creativity is positively associated with all three value proxies when included alongside TRIZ breadth. In contrast, the conditional association between TRIZ breadth and value differs across proxies once functional creativity is included: it is positive for forward citations, statistically indistinguishable from zero for KPSS, and negative for renewals. These findings are associational and motivate additional validation and robustness work, including prompt/model sensitivity checks, targeted human-coded validation, and analyses that move beyond principle counts to principle categories and creativity dimensions.

ABSTRACT. Technological evolution tends to move toward increasing complexity, driven by internal contradictions within technological systems. However, innovation is not a linear process but a nonlinear one involving the interaction of multiple factors, exhibiting dynamic changes over time. Therefore, this study proposes an analytical method integrating node dynamics and spanning trees to model the diffusion patterns of evolving terminology. Specifically, we construct a temporal network of terms based on co-occurrence relationships, extract an update tree of emerging terms, and introduce the Hawkes process to model the diffusion dynamics of terms and their associations over time. Furthermore, it analyses the future evolutionary trajectories of technical terms. Empirical analysis is conducted using 172,362 patents and 2,684,847 papers from the AI field, constructing patent and paper term networks, respectively. The results demonstrate that the proposed method performs robustly, particularly in node and edge prediction tasks within patent term networks. Compared with baseline methods, it achieves improvements of 0.008 and 0.0123 in Brier scores, and 0.3901 and 0.2593 in AUPRC for node and edge prediction, respectively. Overall, this study provides a novel analytical perspective for understanding technological evolution and offers valuable insights for technology foresight and innovation breakthroughs.

ABSTRACT. The efficient transformation of scientific knowledge into technological innovation represents a critical challenge in the national innovation system. However, existing citation-based and text-mining approaches inadequately capture tacit knowledge connections between science and technology. This study proposes a novel framework integrating semantic representation and network structure to identify potential knowledge diffusion pathways. We develop a Dual-Channel Heterogeneous Graph Transformer (DC-HGT) model combined with cross-domain interaction mechanisms to align scientific and technological nodes within a unified semantic space. Subsequently, a link prediction model is employed to identify latent S-T associations that are not revealed by explicit citations or co-occurrence. A multidimensional index comprising topic relevance, tacit knowledge relevance, and transformation probability is established for the systematic evaluation of potential diffusion paths. An empirical analysis in the stem cell domain identified 25 scientific topics, 29 technological topics, and 14 high-potential knowledge diffusion paths. The results reveal diversified cross-disciplinary knowledge flow patterns. Notably, technological topic 26 (stem cell-mediated anti-tumor therapy) receives knowledge contributions from 11 distinct scientific topics, while scientific topic 7 (mechanisms of microenvironment-regulated cell behavior) radiates to 4 different technological topics, demonstrating both the broad radiating capacity of core basic research and the knowledge convergence effect in key application areas. This framework can provide actionable decision-making support for research funding agencies and innovation policymakers.

ABSTRACT. Main Path Analysis (MPA) is a widely used method for tracing knowledge flows in citation networks. Conventional MPA approaches treat documents as vertices and overlook the substantive content within the documents, which restricts a deeper understanding of knowledge evolution and reduces interpretability. To overcome this limitation, we propose a deep learning–augmented, entity-centered MPA framework that supports entity-based path discovery and enhances interpretability. Our method follows a four-step pipeline: (1) data preprocessing to structure the citation network; (2) knowledge entity extraction using a BERT–BiLSTM–CRF model; (3) extraction of multiple main paths via a semantic-aware main path method; and (4) identification of strongly associated entity pairs between citing and cited documents using an attention model with a three-level masking mechanism, which filters out irrelevant entity pairs and enables drilling down from document-level to entity-level representations, thereby generating fine-grained main paths. We validate the proposed approach through extensive experiments on a patent dataset from the thin-film head domain in computer hardware. Results demonstrate that our method reveals finer-grained knowledge flows across key subfields and improves the interpretability of candidate paths

ABSTRACT. Understanding and anticipating scientific change requires models that distinguish between endogenous consolidation and exogenous diffusion of scientific concepts. Using the quantum computing subtree of concepts in OpenAlex, we construct a temporally resolved concept co-occurrence network and track each concept pair through its upstream citation lineage and downstream diffusion. We train LightGBM models on distributional and diversity-aware features to predict four outcomes: endogenous reinforcement, exogenous diffusion, their ratio, and diffusion entropy. After controlling for overall publication growth of the scientific body, endogenous reinforcement proves largely unpredictable. In contrast, exogenous diffusion and entropy are strongly predictable (R2 up to 0.78) and are driven by upstream heterogeneity, citation breadth, and distributional dispersion, as shown by SHAP analyses. Case studies reveal that sharp entropy increases coincide with the opening of new conceptual frontiers, while entropy collapses signal technological convergence or paradigm displacement. These results demonstrate that conceptual diffusion is governed by stable structural regularities embedded in semantic and citation environments. By identifying early diversity-based signals of cross-domain uptake, the approach provides a scalable foundation for anticipatory scientometrics, technology foresight, and innovation-oriented policy analysis in rapidly evolving research fields.

ABSTRACT. To address challenges in Chinese scientific literature, such as severe term nesting, fuzzy boundaries, and long-tail uneven class distributions, this paper proposes a Multi-Teacher Knowledge Distillation Framework (MT-KDF) for entity recognition, integrating multi-scale feature enhancement and Low-Rank Adaptation (LoRA) fine-tuning. To overcome the bottleneck of single models in feature extraction, we first construct a specialized teacher architecture integrating "Character-Lexicon" dual embedding and an Adaptive Temporal Convolutional Network (ATCN). This architecture is designed to capture differentiated local dependency features—ranging from extremely short abbreviations to long descriptive terms—through parallel channels, thereby strengthening the perception of domain terminology. On this basis, a "Single-Entity Expert" strategy is employed to independently train multiple teacher networks, which generate consistent global supervision signals through multi-channel knowledge aggregation to guide student model training. Finally, LoRA technology and a hybrid loss function are introduced in the student model phase to achieve efficient transfer of domain knowledge while freezing the majority of backbone parameters. Experimental results on the SciCN and CMeEE datasets demonstrate that the proposed method achieves highly competitive performance in complex contexts via expert knowledge aggregation. Notably, MT-KDF surpasses mainstream pre-trained baselines and Large Language Models (LLMs) in key metrics while maintaining a superior trade-off between accuracy and computational efficiency through lightweight fine-tuning.

ABSTRACT. This study presents a multidimensional framework for tracking interdisciplinary knowledge evolution by integrating Large Language Model (LLM)-augmented semantic BERTopic modeling with advanced temporal modeling techniques. While conventional approaches, such as bibliometrics and tech mining, often struggle to capture the dynamic and evolving nature of interdisciplinary relationships due to their inability to handle large-scale data and complex semantic structures, our framework systematically models evolution through three replicable phases: (1) comprehensive data collection and structured information extraction; (2) LLM-augmented semantic representation and hierarchical organization of research topics through automated keyword generation, topic interpretation, and distance-based clustering; (3) extension of temporal pattern analysis with topic strength calculations and network metrics. By leveraging LLM to enhance semantic granularity, we reduce manual effort while improving the detection of interdisciplinary knowledge diffusion. Applying this multidimensional framework to Embodied AI (31,618 publications, 2000–2024), we identified eight interconnected research directions and quantified a four-stage evolutionary trajectory characterized by increasing interdisciplinary collaboration. These findings reveal the evolving nature of Embodied AI research, highlighting the growing importance of interdisciplinary collaboration in advancing this field. This work provides actionable insights for researchers and funding agencies, equipping them with a powerful new lens to visualize, understand, and navigate intricate interdisciplinary research landscapes.

ABSTRACT. Accurately quantifying the distribution of patents across specific innovation stages is crucial for evaluating industrial upgrading and refining guiding policies. However, traditional patentometric methods based on IPC codes lack the semantic granularity required to identify the functional attributes of patents, often obscuring the underlying technological structure. This study proposes a novel integration of Large Language Models (LLMs) with causal inference to construct a novel framework for identifying urban technological structures. Taking China's New Energy Vehicle (NEV) industry as a case study, we evaluate the impact of national pilot policies on structural upgrading at the city level. Methodologically, we adopt a "New Technology/New Process/New Product" trichotomy. A fine-tuned LLM (PatentBERT) was employed to automatically annotate approximately 100,000 NEV patent abstracts (2000-2023), with reliability ensured through confidence scoring and manual validation. Subsequently, a multi-period Difference-in-Differences (DID) model was utilized to analyze the policy-induced changes across these three categories. The results indicate that: (1) Since 2010, the technological structure of China's NEV industry has undergone a significant structural shift from downstream product integration to mid-and-upstream R&D; (2) The pilot policies not only boosted the total patent volume but also exerted a disproportionately positive effect on New Technology and New Process categories, suggesting that policies effectively drove qualitative structural upgrading. This research provides micro-evidence for the precise evaluation of industrial policy targeting efficacy and offers a reproducible automated paradigm for large-scale patent semantic mining by bridging economic innovation chain theory with advanced AI technology.

ABSTRACT. This paper proposes an automated approach for technology theme identification based on patent semantic analysis and large language models (LLMs). Using patent texts as the research object, the method systematically extracts Subject–Action–Object (SAO) structures to represent technical knowledge in a structured form, enabling the automatic identification of technology themes. The approach consists of three key stages: first, patent texts are preprocessed and semantically parsed using LLMs to generate SAO triplets, capturing technical entities, functional actions, and target objects; second, the extracted SAOs are vectorized and clustered using unsupervised methods to identify semantically and functionally related technical units, abstracting from microscopic SAOs to mesoscopic technology themes; finally, LLMs are employed to summarize the semantics of each cluster and generate interpretable technology labels. This end-to-end workflow achieves structured representation of technical knowledge and intelligent clustering analysis. Experimental results demonstrate that the proposed method achieves high precision, coverage, and semantic interpretability, providing an effective tool for technology intelligence mining and technological evolution analysis.

ABSTRACT. Technology trees serve as essential tools for systematically structuring technical knowledge and visualizing hierarchical relationships, identifying development gaps, and supporting strategic R&D decision-making. However, as technological systems grow increasingly complex, traditional expert-driven methods for constructing these trees suffer from inefficiency, subjectivity, and a lack of timely updates. To address these challenges, this study proposes an adaptive framework that integrates large language models (LLMs) with graph mining techniques. The proposed methodology proceeds in three main steps. First, the study employs LLM-based few-shot prompting to accurately extract technical entities and hierarchical relations from unstructured patent texts. Second, it utilizes SBERT embeddings to compute semantic similarities, constructing a network that captures latent associations among entities. Third, an improved community clustering algorithm combined with constrained Depth-First Search transforms this network into a structured, layered technology tree. An empirical case study on Electric Vehicle (EV) charging demonstrates the feasibility of our approach. The proposed automated method for technology tree construction significantly enhances objectivity and optimizes efficiency, and provides a novel perspective for technology forecasting and strategic planning.

ABSTRACT. Patent citations embody the flow of technical knowledge and reflect the value and influence of patented technologies. This study constructs a three-layer network of keywords, IPCs, and inventors based on three perspectives: textual semantics, technical structure, and social cooperation. Combining co-occurrence and citation networks, a method for measuring and extracting potential citation relationships is designed. Co-occurrence, citation, and potential citation relationships are extracted for each layer, resulting in a multi-layer heterogeneous network of patent technologies. The R-GCN model is introduced to achieve patent representation learning based on multi-layer heterogeneous relationships. An attention mechanism is used to select and fuse important relationships, identifying key factors in patent citations to improve patent citation prediction capabilities.

ABSTRACT. Science, Technology, and Innovation (STI) policies are central to national and international competitiveness, yet their complexity makes systematic mapping and continuous monitoring a persistent challenge. This study draws on one of the largest initiatives in the field, the OECD STIP Compass survey, which collects and organizes data on STI policy from OECD countries and has historically relied on extensive manual survey efforts to ensure global consistency. Large Language Models (LLMs) are redefining representation learning in NLP, enabling them to process and internalize knowledge from long unstructured documents. This paper presents a novel application of LLMs for structured information extraction and generation from STI policy documents, focusing on OECD data across six sample countries. We develop a data extraction pipeline based on long-context in-context learning to encode task-specific schemas that allow learning of survey taxonomy labels from public URLs referencing policy initiatives. The pipeline integrates validation steps using a secondary LLM for relevance and evidence scoring, and comparison with survey responses completed by human respondents. For evaluation, we apply multiple overlap measures, including overlap ratios, agreement scores between human-generated and LLM-generated policy indicators, and K-fold cross-validation for AI-generated labels. Our findings indicate that LLMs can achieve high overlap with human respondents for policy indicators (84-95%). Qualitative analysis reveals that the model tends to provide more detailed descriptions, complementing human-written content. Our approach points to the potential of an AI-assisted framework for STI policy monitoring, enhancing both efficiency and quality in international policy intelligence.

ABSTRACT. This research investigates whether large language models (LLMs) can provide a content-based complement to existing journal evaluation systems in biological sciences. Using UK REF 2021 submissions in UoA 5 (Biological Sciences), we construct a dataset of journal articles with departmental REF quality profiles, citation data, and cleaned English abstracts. An LLM is prompted as a “virtual REF reviewer” to perform comparative quality judgements on pairs of titles and abstracts, producing both binary preferences and 1*–4* REF-style ratings. From these outputs, we derive LLM-based quality indicators (LLM_REF and BT_score), which are benchmarked against REF-derived departmental quality indicators (Dept REF mean) and citation-based impact indicators (MNLCS and SJR). Analyses examine citation patterns across star levels, overlap in highly ranked journals, novelty preferences, and rank correlations among indicators. The project aims to assess how far LLM-based content judgements can support, rather than replace, expert-led journal evaluation.

ABSTRACT. Limitations and future work sections are a core component of scientific papers, explicitly inviting authors to reflect on constraints encountered during the research process and to articulate directions for further inquiry informed by this experience. Despite their central role in scholarly communication, these sections have received limited systematic attention as an object of analysis in their own right. This research-in-progress paper presents an exploratory, methodological study of limitations and future research statements in recent Digital Libraries research, with the aim of establishing an initial analytical framework for studying such reflective sections. We focus on papers published between 2020 and 2025 that are classified under the cs.DL (Digital Libraries) subcategory in arXiv. As a proof-of-concept, we draw on a random subsample of 50 papers, from which 20 contain a dedicated section or subsection explicitly addressing limitations and/or future work. Using arXiv metadata and full-text PDFs, we extract and analyze these sections. To support the analysis, we explore the use of large language models (LLMs) to assist with the extraction, summarization and classification of limitations and future work statements. Based on an initial inspection of the extracted statements, we propose a preliminary typology consisting of ten broad categories that can be applied consistently to both limitations and future work. Applying this typology to the proof-of-concept subsample indicates that data- and method-related considerations are most frequently foregrounded, while aspects such as usability, computational constraints, and scope limitations are less commonly articulated. These observations are indicative rather than definitive and primarily serve to demonstrate the analytical potential of the proposed framework. By framing limitations and future work sections as a systematic object of analysis, this paper contributes to meta-research on scholarly communication and provides a preliminary analytical framework that can be extended to larger corpora, other research fields, and longitudinal studies of reflective scientific discourse.

ABSTRACT. This study examines the impact of Large Language Models (LLMs) on the research practices and productivity of young researchers. Through an analysis of PhD students at French universities who defended their theses between 2019 and 2025, we investigate how the emergence of LLMs has been integrated into research activities and the changes it has brought to academic practices. Using a matching strategy that compares students sharing the same supervisor before and after ChatGPT's release, combined with a placebo design, we provide evidence on the dimensions and effects of LLM adoption. Our quantitative analysis of nearly 6.000 STEM theses is complemented by a survey of 265 doctoral graduates, shedding light on the modalities and concerns related to LLM implementation among young scholars.

ABSTRACT. Scientific novelty plays a crucial role in expanding the frontiers of knowledge, driving innovation, and responding to urgent sustainability challenges. This research explores how a nation’s scientific relatedness, defined as the proximity of its existing knowledge base to a specific Sustainable Development Goal (SDG) domain, relates to the further production of novel scientific discoveries. Analyzing more than 4 million SDG-related publications from 165 countries between 2000 and 2023, we reveal a dual effect of scientific relatedness on future knowledge production. While it strongly promotes subsequent research productivity in a domain, it also suppresses scientific novelty. This “scientific relatedness penalty” is especially marked in the Global South. These findings make contributions by clarifying how nations can balance the innovation paradox, the inherent tension between leveraging existing competencies and pursuing novel directions. For policy makers, our outcomes highlight the need to enhance global scientific connectivity to support transformative sustainability science.

ABSTRACT. Novelty is a core value in scientific research, and its measurement has been of wide scholarly and practical interest. Numerous bibliometric indicators for novelty have been proposed, some shared in open repositories, which have facilitated empirical investigation into scientific novelty. However, there remains a fundamental limitation. That is, we have insufficient knowledge as to what these indicators truly measure because of limited efforts for examining measurement validity. This study addresses this gap by evaluating a range of novelty indicators, employing various operationalisation strategies, against self-reported novelty assessments obtained from our originally designed questionnaires covering multiple novelty dimensions. Our analyses examining the correlation between the self-assessed scores and bibliometric indicators offer several insights. First, while most indicators detect some aspects of novelty, a single indicator may not sufficiently capture all forms of novelty. Second, a cross-disciplinary comparison reveals that indicators’ performance varies across disciplines – some indicators demonstrate consistent correlations across disciplines while others show correlations only in limited disciplines. Third, as employing language models in novelty evaluation has become common, we compare the static language model and the contextual large language model, finding that indicators based on the latter outperforms those on the former. Fourth, we examine ex-post indicators, which require post-publication data (e.g., forward citation), and find that they offer no clear advantages over ex-ante indicators in detecting novelty. These findings highlight both the potential and limitations of existing indicators and offer implications for the future development and application of novelty indicators.

ABSTRACT. We introduce an interpretable machine-learning algorithmic approach that forecasts emerging links between research concepts by modelling how OpenAlex concept networks in quantum computing evolved from 1990 to 2023. Using 59 semantic and topological features, a two-stage LightGBM model predicts both the formation and growth of concept pairs (AUC ≈ 0.95). Its regression performance remains stable: RMSLE increases from 0.45 at one year to 0.6 at five years, meaning that predicted link strengths stay within roughly a factor of 2 despite exponential growth. Feature attribution shows that structural factors, particularly Adamic–Adar similarity and degree-based Hadamard measures, consistently drive forecasting accuracy. These patterns suggest that breakthroughs tend to emerge in tightly connected sub-networks where ideas recombine rapidly. Two expert-validated examples, quantum annealing and AI-enabled quantum architectures, illustrate how the model captures technological convergence as anticipated by experts. Building on these findings, we outline a three-layer decision architecture that connects automated detection, expert translation, and institutional integration, to support evidence-based research strategy and policy. The framework offers a reproducible foundation for transforming large-scale knowledge data into actionable intelligence for science and technology governance.

ABSTRACT. "Standing on the shoulders of giants" has long been recognized as a fundamental pathway for scientific innovation. However, a systematic analysis of 1.64 million papers quantifies the true cost of citing classic literature, revealing a pervasive Novelty–Impact Paradox: papers that cite highly innovative classics exhibit significantly higher linguistic novelty, whereas their explicit impact (measured by citation counts) suffers a significant innovation penalty (a decline of approximately 15.9%).This paradox is subject to triple moderation by disciplinary context, the type of cited classics, and the stage of knowledge evolution. We find that the paradox is more pronounced in the natural sciences or in the same-discipline citation scenarios. The paradox onset point—the time at which the paradox first manifests—appears earlier for practical classics than for theoretical classics. Furthermore, traditional citation metrics primarily capture impact during the knowledge creation phase, creating a time blind spot regarding knowledge flows during the diffusion and maturity phases, thereby exacerbating the paradox. By introducing "novelty reuse" as a proxy for implicit impact, our study demonstrates that while citing classics may reduce explicit impact, it significantly fosters the implicit dissemination of knowledge and conceptual inheritance, thus partially mitigating the conflict between novelty and impact.These findings challenge monolithic and myopic impact evaluation paradigms. They provide a theoretical and empirical basis for constructing a more comprehensive and time-sensitive scientific evaluation system, while also offering strategic insights for researchers on how to critically inherit classics to achieve sustainable innovation in practice.

ABSTRACT. Citations constitute the core mechanism of scientific knowledge dissemination. However, existing research predominantly focuses on rhetorical functions or authorial intent, overlooking the substantive knowledge contributions of cited works. More critically, prior studies often analyze isolated contribution statements rather than authentic citation contexts, resulting in findings divorced from real-world scenarios. Consequently, accurately identifying the knowledge contributions of citations has become a pivotal challenge in understanding scientific knowledge production and evolution. To address these challenges, we propose the Scientific Knowledge Contribution (SKC) classification framework, which for the first time categorizes citations from an objective contribution perspective into five classes: Method & Technology, Resource & Tool, Theory, Empirical Finding, and Background. Through two-stage sampling, we construct a high-quality annotated dataset and employ six representative models for systematic evaluation, establishing performance benchmarks for this task. This study achieves a methodological shift in citation analysis from subjective intent to objective contribution. The theoretical framework reveals knowledge dissemination characteristics in the NLP domain while uncovering performance limitations of existing models. This research provides essential infrastructure for fine-grained knowledge contribution identification, establishing foundations for large-scale scientific knowledge flow analysis and disciplinary evolution studies.

ABSTRACT. Building on the context of big science facilities, this study provides a more micro method to identify the scientific mobility procedure, named short-term scientific mobility hereafter, and associates with ability to yield novel and disruptive knowledge. We classify external users of big science facilities into two types (travelers and locals) by measuring the number of facilities the focal author used in one-year, represented by the number of publications supported by the corresponding facilities. We also include these authors’ other publications without supports from big science facilities. Scientific performance is measured by novelty score and the ten-year disruption index. Results show that locals associated with more disruption while travelers perform better in novel knowledge production. Diverse variables related to travel validate the robustness of our preliminary results. This study contributes to understanding the performance evaluation and science policy in the context of big science facilities and enriches the research in scientific mobility. The discoveries also could be a reference for those short periods scientific activity related to mobility but without visible information to quantify.

ABSTRACT. This study draws on the guideline module of the Chinese Medical Journal Full-text Database to systematically analyze the characteristics of academic references cited in Chinese clinical guidelines during the COVID-19 pandemic and to explore how they reflect the societal impact of research. Based on an analysis of 49 guidelines containing 1,617 references, the study reveals the following: (1) Journal articles constituted the predominant type of references (82.7%), indicating the strong reliance of guidelines on scholarly publications as evidence; (2) the temporal distribution exhibited pandemic-driven, stage-specific features, with a marked citation peak in 2020, reflecting the amplified demand for evidence during a public health emergency; (3) the country distribution was dominated by journals from the United States and the United Kingdom (58.9% combined), while papers published in Chinese-language journals (17.7%) played a distinctive role in providing contextualized evidence; (4) the journal distribution showed that leading international medical journals and core domestic journals jointly formed the evidence base; and (5) case analysis demonstrated that the evidence base comprised both domestic epidemiological and policy studies and international high-quality clinical and pathological studies, jointly supporting the rapid response and evidence-based decision-making of the guidelines during the pandemic. The findings suggest that evidence adoption in Chinese clinical guidelines followed a dual-track model of rapid domestic response combined with authoritative international support. This approach enhanced both the scientific rigor and contextual adaptability of the guidelines, while also providing a China-specific empirical pathway for evaluating the societal impact of research.

ABSTRACT. Scientific knowledge serves as the theoretical foundation for technological development and innovation, and the scientific knowledge network, with scholarly literature as its knowledge carrier, plays a crucial role in full-cycle technology forecasting. While current technology forecasting research has made significant algorithmic progress in handling large-scale data using black-box AI models, such as deep learning, it provides limited support for domain experts’ judgment and lacks explainable insights and theoretical foundations for the complex interactions among scientific and technological entities. Building upon existing technology forecasting theoretical models, this study proposes a new theoretical model of influencing factors for technology forecasting, thoroughly analyzing the mechanism by which semantic content and citation structure features in scientific knowledge networks enhance forecasting effectiveness. To validate the theoretical model, we designed a human-AI collaborative forecasting framework based on matrix decomposition and trend extrapolation. Using computer science literature data, we constructed a multiplex graph and employed frequency, semantic and citation integrated feature to predict the frequency of keywords representing sci-tech research topics. Feature ablation experiments and temporal performance analysis across different network scales demonstrate that semantic and structure features in scientific knowledge networks significantly improve forecasting performance. Moreover, the results reveal a widespread teleconnection phenomenon of multi-dimensional fused features across the network. This study offers new theoretical foundations and methodological guidance for technology forecasting in theory, practice, and policy making.

ABSTRACT. Understanding how the amount of experimental resources used in scientific experiments relates to research innovativeness is an important issue in the scientometric analysis of large-scale research infrastructures. While existing studies have extensively examined the aggregate productivity and citation-based performance of major facilities, relatively limited attention has been paid to whether variations in experimental resource use at the paper level are systematically associated with differences in research innovativeness. This study addresses this issue by examining the relationship between experimental procedural complexity and paper-level innovativeness in the context of synchrotron radiation research.

Using publication data from the European Synchrotron Radiation Facility (ESRF), experimental procedural complexity is operationalized as the number of beamline stations employed per paper. This measure reflects the amount of experimental resources required to conduct a study and serves as a parsimonious proxy for the procedural complexity of experimental execution, insofar as experiments involving multiple beamlines typically entail more diverse measurement techniques and experimental configurations. Research innovativeness is measured using the Relative Independence Index (IND), a citation-network-based indicator capturing the degree to which a publication positions itself independently from its referenced literature. Unlike citation volume or diffusion-based metrics, IND reflects a structural aspect of innovativeness related to deviations from established citation pathways.

Methodologically, the study adopts a generalized propensity score matching (GPSM) framework to estimate the dose–response relationship between experimental procedural complexity and research innovativeness. Treating beamline usage as a continuous treatment variable and adjusting for a set of publication-level covariates, GPSM allows for a flexible assessment of how expected innovativeness and its marginal changes vary across different levels of experimental procedural complexity.

The empirical results indicate that publications based on experiments involving higher procedural complexity tend to exhibit greater innovative relative independence. The estimated dose–response function shows an overall increasing pattern, suggesting a positive association between the number of beamline stations used and research innovativeness. At the same time, marginal effect estimates reveal diminishing returns, with the positive association being more pronounced at lower levels of beamline usage and gradually weakening as additional beamline stations are incorporated. These findings point to a non-linear relationship between experimental procedural complexity and innovativeness rather than a simple linear scaling effect.

By demonstrating how scientometric methods can be applied to quantify the innovative implications of experimental procedural complexity, this study provides empirical evidence relevant to the assessment of experimental designs in large-scale research infrastructures. The results offer a methodological basis for informing discussions on experimental resource utilization at synchrotron radiation facilities, while remaining agnostic about the underlying organizational or managerial mechanism.

ABSTRACT. The science of science has emerged as a major interdisciplinary field for analyzing the structure, dynamics, and social functions of scientific research. However, its intellectual organization and patterns of knowledge integration across subfields remain insufficiently systematized. This study provides a quantitative and comparative mapping of the major subfields of the science of science based on journal-level publication and citation data. Using articles collected from representative international journals, we apply bibliometric and scientometric techniques combined with knowledge graph visualization to analyze research topics, contributing actors, and citation linkages. Topic co-occurrence networks and inter-journal citation networks are constructed to identify the thematic structure of each subfield and to trace knowledge flows within and between them. The results show that all subfields exhibit a heterogeneous and evolving topic structure, in which established research themes coexist with rapidly emerging ones and form distinct evolutionary trajectories. Moreover, knowledge flows are highly concentrated among conceptually proximate subfields, while cross-domain exchanges remain comparatively limited, resulting in a modular but interconnected knowledge system. These findings provide a systematic empirical basis for understanding the interdisciplinary configuration and developmental dynamics of the science of science.

ABSTRACT. This study aims to address the issue of current high-value patent identification models overly relying on historical statistical data and lacking causal explanatory power, in order to enhance the effectiveness and interpretability of high-value patent identification. Therefore, a high-value patent recognition model based on counterfactual data augmentation is proposed. This model constructs heterogeneous graphs containing multiple types of nodes, captures relationships between nodes through ensemble graph attention networks (GAT) and graph convolutional networks (GCN), and uses adversarial neural networks to generate counterfactual samples to perturb edge features. It combines incremental learning strategies to evaluate patent value and reveal causal mechanisms. The experimental results show that the model achieves accuracy, precision, recall, and F1 score of 84.27%, 84.51%, 85.17%, and 84.84%, respectively, on patent datasets in the field of artificial intelligence, which is superior to the baseline model. This validates its effectiveness and provides a new research perspective for high-value patent recognition tasks.

ABSTRACT. In technology-intensive fields, the “instant burst” and “delayed recognition” of patent citations constitute two distinct pathways for value realization. To decode the drivers of this divergence, this study constructs a comparative framework based on deep semantic understanding. We employ a multi-dimensional screening algorithm to isolate Sleeping Beauties (SBs) and Early Highly Cited (EHC) patents, utilizing a hybrid vector model—fusing Qwen3 embeddings with TF-IDF features—integrated with BERTopic to map evolutionary paths accurately. Empirical results reveal that SBs emerge on average 3.08 years prior to topic explosion, confirming their nature as "premature" innovations. Organizationally, SBs predominantly originate from equipment manufacturers securing next-generation reserves with broad claims, whereas EHC patents are driven by manufacturing giants addressing immediate bottlenecks. Ultimately, we identify the temporal misalignment between technological supply and the industrial ecosystem as the fundamental mechanism inducing dormancy, offering a micro-semantic perspective on non-linear value realization in high-tech industries.

ABSTRACT. Identifying high-value patents is of great significance for optimizing innovation resource allocation. To address the limitations in existing patent valuation methods—such as highly subjective indicator systems, insufficient utilization of textual features, and lack of feature integration—this study proposes a novel approach for identifying high-value patents by integrating semantic information with quantitative indicators. By constructing a patent value assessment indicator system covering four dimensions (technical, strategic, legal, and economic), combined with modern-BERT text feature extraction and UMAP dimensionality reduction, this study innovatively integrates quantitative patent indicators with textual semantic features and compares various models from traditional machine learning and deep learning. The results show that deep learning models perform better, with the proposed CNN-BiLSTM fusion model significantly outperforming traditional methods in terms of accuracy, F1 score, and AUC score. Ablation experiments further validate the effectiveness of the multi-feature fusion strategy, revealing the differentiated contributions of quantitative indicators and textual features to model performance. This study provides a scientific and effective method for high-value patent identification, which contributes to improving patent commercialization efficiency and promoting a win-win situation between technological innovation and economic benefits.

ABSTRACT. Abstract: In the rapidly evolving field of smart elderly care health monitoring, a large number of patents lapse due to non-technical factors. However, the technical schemes contained within these patents still possess potential value for reuse. Traditional patent valuation methods often lack applicability to lapsed patents, and existing recommendation methods struggle to integrate patent value with enterprise technical preferences. To address these issues, this paper proposes a method for the quality assessment and precise recommendation of lapsed patents based on a Heterogeneous Graph Attention Network.This study focuses on the domain of smart elderly care health monitoring. It constructs a structured patent dataset and achieves technical topic identification based on deep semantic representation and topic modeling. On this basis, the study introduces indices for technological mutation and technological outlierness from the complementary perspectives of technological evolution and distribution. These indices are used to quantitatively assess the intrinsic technical value of lapsed patents, screening them to form a candidate set of high-value lapsed patents. Furthermore, the paper integrates enterprises, patents, and technical topics into a unified modeling framework to construct a Value-Aware Heterogeneous Graph Attention Network (VA-HGAT). A weighted loss mechanism is employed to guide the model to focus on high-value patent features, enabling the effective learning of technical matching scores between enterprises and lapsed patents.In the recommendation stage, a multi-factor fusion recommendation score is constructed by combining the technical matching degree, intrinsic patent value, and time decay factors. The recommendation results are then structurally optimized through technical topic quadrant analysis. Case study results indicate that this method can stably identify high-value lapsed patents that are highly compatible with the target enterprise's technological layout. The recommendation results demonstrate strong performance in terms of both technical relevance and value rationality. This research provides an interpretable and scalable methodological path for the systematic mining and precise allocation of lapsed patents as low-cost technical resources, offering valuable reference for enterprise technology decision-making and technology transfer practices.

ABSTRACT. International collaboration in science and innovation is crucial for scientific progress and the efficient use of scientific knowledge. Studies indicate that internationally co-authored papers are cited more frequently than national or non-collaborative publications. This phenomenon can be attributed to a larger readership and the higher scientific relevance of international projects.

The distribution of international co-publications varies significantly among countries: larger nations tend to have lower shares, while smaller countries often achieve higher proportions. This discrepancy arises because larger countries are more likely to find national partners with complementary competencies, whereas smaller nations typically maintain more specialized science and innovation systems. Recent evidence, however, shows that the US has a lower share of international co-publications compared to many other countries.

This study aims to investigate the expected and optimal levels of international co-publications for various countries, based on their size, profile, and research orientation. It assesses how effectively countries utilize their scientific resources to achieve visibility and excellence.

Three methodological approaches are employed: a fixed-effects panel regression model to estimate expected co-publications, a shift-share analysis to differentiate global and national trends, and a Data Development Analysis (DEA) to evaluate the efficiency of countries in using their resources for scientific output.

Results from the regression analysis indicate that despite controlling for size, excellence, and research orientation, countries such as the US, China, South Korea, and Japan exhibit low levels of international co-publication activity, whereas countries like Singapore, Austria, France, and the UK are highly engaged in international collaboration.

The shift-share analysis reveals negative trends for the US even after accounting for global effects. Finally, the DEA assesses the efficiency of research systems using internationally comparable indicators, such as the number of researchers and R&D expenditures. This part of the analysis is still in progress, and results are not yet available.

Overall, the paper aims to enhance the understanding of the dynamics of international scientific collaboration and evaluate how different countries optimize their resource utilization to maximize scientific output performance.

ABSTRACT. Research impact is increasingly recognized as a multidimensional phenomenon encompassing academic, technological, and policy dimensions, yet their dynamic interrelationships remain insufficiently explored. This study examines the interactions among these three types of impact using longitudinal data from the bioinformatics field. Based on 3,921 publications that generated academic, technological, and policy impacts, we construct an unbalanced panel dataset linking annual paper citations, patent citations, and policy document citations from 1991 to 2019. A Panel Vector Autoregression (PVAR) model estimated via system Generalized Method of Moments is employed, complemented by Granger causality tests, impulse response analysis, and variance decomposition. The results reveal strong self-reinforcing effects within each impact dimension and a bidirectional causal relationship between academic and policy impacts, indicating mutual reinforcement over time. In contrast, technological impact follows a largely independent trajectory. These findings highlight the importance of interaction-oriented approaches to research impact assessment.

ABSTRACT. Open access (OA) is widely recognized for enhancing scholarly visibility, yet whether or how it facilitates the integration of highly influential research into technological systems remains insufficiently understood. This study investigates the pathways through which OA enables highly cited AI papers to achieve technological use, focusing on both direct and indirect mechanisms linking OA to patent citations. Drawing on an integrated dataset of 2,082 highly cited AI papers from Web of Science, InCites, and SciSciNet, we employ OLS regression and mediation analysis to examine the direct and indirect pathways through which OA influences patent citations, and to assess how these effects vary across Green, Free-to-Read, Gold-Hybrid, and Gold OA. The results reveal a dual-pathway mechanism linking OA to technological impact. OA significantly increases both paper and patent citations, with nearly half (49.11%) of its technological advantage operating indirectly through enhanced academic impact, confirming a substantial mediation effect. Additionally, OA retains a significant direct effect on patent citations, indicating that inventors benefit from OA beyond its role in scholarly dissemination. Among four OA types, Green OA exhibits the strongest effects, enhancing technological impact both directly and indirectly. Gold-Hybrid OA contributes to technological use primarily through increased scholarly visibility, while Free-to-Read and Gold OA show no significant effects on either academic or technological impact. This study advances understanding of how OA shapes science-technology linkages and highlights the role of specific OA types in promoting the technological value of highly cited AI papers.

ABSTRACT. Technology convergence is a key driving force for promoting breakthrough innovations and discovering technological opportunities. This study, based on the core innovation ideas within the field, identifies potential technology convergence opportunities and constructs an analysis framework that integrates multi-source information, combining core innovation ideas with sub-domains, IPC classifications, and text features. On this basis, using a technology integration prediction method based on graph convolutional networks and bidirectional long short-term memory networks, it provides a new perspective for related research. The model is verified using patent data from the medical robot field. The results show that the development of this field is highly dependent on the deep integration of robot technology, computer network control technology, medical surgical technology, and artificial intelligence technology, demonstrating a significant interdisciplinary technology convergence diffusion mechanism. This study fills the gap in existing research on technology convergence by considering cross-domain technology convergence networks while ignoring the technology convergence mechanisms within the domain, providing theoretical references for innovation practice.

ABSTRACT. Technology fusion is a primary driver of modern innovation. Forecasting its emergence is strategically vital for identifying nascent technological opportunities and anticipating industrial transformations. However, capturing the higher-order dependencies and complex temporal evolution inherent in this process—which fundamentally differ from dyadic combinations—presents a unique challenge. To address this challenge, we construct a framework based on temporal hypergraph link prediction (THLP). Fine-grained technology domains are modeled as nodes, and their fusion events—the co-occurrence of multiple technologies in a single patent—are modeled as time-evolving hyperedges. We systematically designed and extracted a multi-dimensional feature framework. Next, a temporal hypergraph deep neural network (THG-DNN) model is then proposed to predict the formation of true and potential future fusions. Then, we validate our framework against multiple baselines using patent data from a field of critical and high-frequency areas of emerging technology fusion. The superior predictive performance confirms the necessity of modeling higher-order dependencies within technology combinations. Beyond prediction, our framework reveals stable key drivers extracted from a hypergraph for technology fusion, including strong evidence for path dependency and cognitive proximity. Furthermore, it also uncovers U-shaped effects of recombination mode and inverted U-shaped effects of cooperation network structure, providing deeper insights into the fusion process. This research offers a powerful analytical tool for managers, policymakers, and investors to identify emerging technological opportunities and formulate effective and verifiable innovation strategies.

ABSTRACT. In the context of accelerating global technological and industrial transformation, this study aims to systematically reveal the multifaceted mechanisms through which technology convergence drives industrial innovation. From the perspective of technology convergence, this study develops a set of conditions shaping industrial innovation drawing on three theoretical lenses: micro-level technology recombination theory, meso-level co-evolution theory, and macro-level convergence chain theory. Following the innovation ecosystem framework, the analysis first employs the SHAP (SHapley Additive exPlanations) model to identify core variables. Subsequently, fsQCA (fuzzy-set Qualitative Comparative Analysis) is applied to derive configuration pathways, and the XGBoost (eXtreme Gradient Boosting) model is used to construct decision-tree pathways incorporating all variables. Finally, the two approaches are integrated to identify specific patterns through which technology convergence promotes industrial innovation. An empirical analysis of the biopharmaceutical industry demonstrates that the configuration pathways derived from fsQCA are largely consistent with the decision-tree results produced by XGBoost. Four pathways are identified as enhancing industrial innovation performance: institutional coordination, endogenous network evolution, industry-driven innovation, and demand-oriented convergence. These findings indicate that industrial innovation is jointly driven by technology–industry coupling as the dominant factor, collaboration among innovation actors under varying external conditions, and the combined effects of structural and diffusion convergence within and across networks.

ABSTRACT. The field of AI for Science (AI4S) has garnered increasing global attention, making the analysis of its technological landscape crucial for strategic decision-making and innovation. A distinguishing feature of AI4S is the deep convergence between general AI technologies and specific scientific scenarios. Traditional single-mode analysis often fails to capture this dualistic coupling. To address this, this paper proposes a novel framework constructing a "Technology-Scenario" Two-Mode Network to systematically analyze the landscape of technology convergence in AI4S. Methodologically, the study first employs Large Language Models (LLMs) to identify AI4S patents from the database and extract fine-grained technical entities and application scenarios from titles and abstracts of patents, constructing a high-quality dataset. Subsequently, a quantitative evaluation framework is established to address three core objectives: (1) investigating the differential adaptation of diverse AI technologies to distinct scientific problems; (2) identifying the critical nexus bridging heterogeneous disciplines amidst diverse technological convergence; and (3) characterizing the dominant fusion paradigms emerging within the AI4S domain through Louvain community detection algorithms, which cluster tightly coupled technology-scenario combinations into functional patterns. The empirical results reveal a hierarchical "multi-path" convergence pattern in the current AI4S landscape. On one hand, general technologies such as GNN and CNN serve as broad infrastructure, enabling diverse disciplines like biomedicine and materials science. On the other hand, specialized technologies incorporating domain knowledge, such as Physics-Informed Neural Networks (PINN), demonstrate high adaptation in solving distinct hard problems in physics and mathematics. These findings provide a granular view of how AI technologies differentially penetrate various scientific paradigms. This study offers a new methodological tool for visualizing the complex interactions in cross-disciplinary innovation. It provides empirical evidence for research institutions and enterprises to identify high-potential "Technology-Scenario" combinations and optimize their R&D layouts. Future work will extend this framework to include dynamic trend analysis, further elucidating the evolutionary trajectory of AI4S.

ABSTRACT. Technology convergence is widely recognized as a key mechanism of technological innovation, yet existing studies often rely on static indicators and macro-level classifications, overlooking the dynamic evolution and heterogeneity of specific technology combinations. This study develops a dynamic, two-dimensional framework to characterize technology convergence from the perspectives of structural tightness and outcome balance. Based on patent co-classification data at the IPC main group level, we construct technology combinations as micro-level units of convergence. We measure structural tightness using co-occurrence intensity and structural adhesion, and assess outcome balance through semantic balance and citation-based knowledge inflows. To trace temporal dynamics, we incorporate a time-decay weighted index and apply polynomial trend fitting to identify convergence trajectories. An empirical analysis of additive manufacturing patents from 1986 to 2024 reveals substantial heterogeneity in convergence dynamics and identifies four convergence types: Loose–Balanced, Tight–Balanced, Loose–Unbalanced, and Tight–Unbalanced. Overall, this study provides a fine-grained perspective on technology convergence and offers methodological support for technology foresight and innovation policy.

ABSTRACT. In corporate innovation decisions, the selection of foundational knowledge often involves a critical trade-off: prioritize highly cited patents widely recognized in the industry, or invest in disruptive technologies with breakthrough potential? However, existing studies tend to treat highly cited patents as a homogeneous group, ignoring how their disruptive traits affect technology dissemination differently. To fill this gap, this study employs a knowledge flow perspective to examine how the subsequent technology diffusion performance of patents differs after absorbing different types of foundational knowledge. Based on 8,032,090 utility patents granted by the United States Patent and Trademark Office (USPTO) from 1980 to 2024, we employ the PatentSBERTa patent text embedding model to assess semantic similarity between patents, define highly cited patents using five-year forward citations, and identify disruptive patents via the disruption index (DI). The study draws three main conclusions: First, the semantic similarity between patents and the technologies they cite has steadily declined, which confirms that technological innovation is shifting from a single knowledge base to diversified knowledge integration. Second, subsequent patents citing highly cited patents exhibit a long-term and stable advantage in technological diffusion. Third, disruptive patents and the subsequent patents citing them have lower semantic similarity and hinder the technological diffusion of subsequent patents. Supported by large-scale data, this study offers a new empirical view on patent technology diffusion mechanisms. It also provides key practical guidance for enterprises to balance innovation certainty and breakthrough potential, and improve their knowledge base choices.

ABSTRACT. To reveal the evolutionary mechanisms of innovation pathways in complex technological systems, this study introduces a fitness landscape approach to construct a model of technological innovation space. Using a patent-based NK modeling framework, the research identifies core technological elements and their interactions to reconstruct a performance “topography” of the design space. A local hill-climbing algorithm is applied to simulate the evolutionary process of technology combinations, analyzing issues of path dependence and local optima. Results indicate that the technological innovation space exhibits a distinct multi-peaked and highly rugged structure, and system complexity strongly influences the direction and efficiency of innovation searches. The findings provide a computable analytical tool for innovation forecasting, technology mapping, and intelligence monitoring, offering decision-making insights for complex industrial systems.

ABSTRACT. This study investigates whether Generative AI (GenAI) can overcome the limitations of traditional bibliometrics to identify high-potential, non-consensus scientific research early. We address the critical dilemma of whether GenAI acts as a visionary “scout” for disruptive ideas or a conservative “guardian” that amplifies existing biases. Employing a novel counterfactual historical prediction framework, we simulate an early assessment point for AI publications, using only contemporaneous data to task GenAI models with forecasting long-term impact, which is validated against patent citations and “Sleeping Beauty” papers. Expected results indicate a nuanced role for GenAI: it demonstrates conditional superiority in predicting technological impact but struggles to identify paradigm-shifting academic work and exhibits systematic biases favoring prestigious institutions. This suggests GenAI’s utility is currently better suited for spotting applied innovations than scientific revolutions, underscoring the necessity of ethical governance and human oversight in its integration into research evaluation.

ABSTRACT. This study proposes a comprehensive text-based framework for cross-industry technology opportunity analysis to transform technological restrictions into opportunities. Compared with existing research on technology opportunity analysis, we improve upon the limitations of traditional patent-based approaches that predominantly confine analysis within single technological domains and fail to capture cross-industry application potential under restriction conditions. In this study, we introduce an integrated analytical framework that systematically processes patent data and policy texts, extracts and expands technical vocabularies from CCL restrictions and SEI classifications through semantic modeling, and applies weak signal detection methodology combining text dissimilarity analysis with KIM-KEM composite charting for weak signal terms. We then demonstrate how this framework works by examining China's new energy vehicle industry under U.S. export controls. Through cross-industry technology mapping, we identified systematic correlations between restricted technologies and industry applications. Using weak signal detection, we derived key terms including "airflow," "green," and "cool control" from outlier patent documents. Through SAOX semantic analysis, we extracted multiple semantic structures revealing cross-industry applications such as intelligent thermal management systems, green energy ecosystem integration, and intelligent thermal control systems. We built a comprehensive opportunity transformation framework based on analysis results to explore promising cross-industry opportunities for organizations facing technological restrictions. We contributed to the related research field by enabling organizations to leverage existing technological foundations to systematically identify cross-industry opportunities beyond traditional single-domain analytical constraints.