AI’s Cross-Domain Evolution Unveiled Through Patent Network Analysis
In the rapidly evolving landscape of artificial intelligence (AI), understanding how technological innovation emerges and spreads across disciplines has become a critical task for researchers, policymakers, and industry leaders. A groundbreaking study conducted by Chen Yufen and Wang Keping from the School of Statistics and Mathematics at Zhejiang Gongshang University offers a comprehensive analysis of AI’s cross-domain convergence patterns using a novel framework rooted in patent data. Published in the Journal of Intelligence, the research leverages advanced network analysis techniques to uncover the structural dynamics behind AI’s integration with diverse technological fields, shedding light on the mechanisms that drive its rapid advancement.
The study, titled Cross Domain Pattern Recognition of Artificial Intelligence Based on Patent Technology Co-Occurrence Network, draws on a vast dataset from the Derwent Innovations Index, encompassing over 112,000 AI-related patents filed between 1995 and 2018. By applying technology co-occurrence network analysis and brokerage role identification, the researchers map the evolution of AI’s technological ecosystem over four distinct developmental phases. Their findings not only confirm AI’s inherently interdisciplinary nature but also reveal the key mediating technologies that facilitate knowledge transfer across domains.
At the heart of the methodology is the use of International Patent Classification (IPC) codes, specifically the four-digit level (IPC4), to identify and categorize technological fields. Each patent that contains multiple IPC4 codes is considered a point of technological convergence, indicating that the invention integrates knowledge from different domains. By constructing co-occurrence networks where nodes represent IPC4 categories and links represent their joint appearance in patents, the authors trace how AI has progressively absorbed and recombined knowledge from fields such as electrical engineering, mechanical systems, chemistry, and information management.
One of the most significant contributions of this research is its application of social network theory—particularly the concept of brokerage—to the domain of technological innovation. In network science, a broker is an actor that connects otherwise disconnected groups, facilitating the flow of information and resources. Translating this concept into the context of patent networks, the researchers identify specific IPC4 categories that act as technological intermediaries, bridging disparate knowledge domains and enabling cross-fertilization of ideas.
The analysis reveals that AI’s development has followed a clear trajectory: beginning with foundational advancements in core technical areas before expanding into broader application domains. During the initial phase (1995–2000), AI innovation was concentrated in a few key areas, most notably computer technology (IPC4: G06F), control systems (G05B), and machine tools (B23Q). These domains formed the technological backbone upon which subsequent innovations were built. As time progressed, the scope of AI applications broadened significantly, with increasing integration into sectors such as transportation, healthcare, pharmaceuticals, and consumer electronics.
A pivotal moment in AI’s evolution appears to have occurred around 2007, coinciding with major breakthroughs in deep learning and the rise of big data. The study notes a sharp increase in patenting activity during this period, followed by a stabilization and then a surge in the 2013–2018 phase, driven by advances in cloud computing, the Internet of Things (IoT), and mobile technologies. This acceleration reflects not only improved computational capabilities but also a growing recognition of AI’s potential to transform industries far beyond its original technological base.
What distinguishes this research from prior studies is its focus on dynamic patterns of convergence rather than static snapshots of technological overlap. While earlier work has often measured the degree of technology fusion at a given point in time, Chen and Wang’s approach captures how these fusion patterns evolve over decades. Their framework identifies three primary modes of cross-domain interaction: independent development, unidirectional convergence, and bidirectional convergence.
Independent development refers to technological progress within a single domain without significant external influence. While present in some sectors—particularly chemistry and mechanical engineering—this mode plays a relatively minor role in AI’s overall innovation landscape. Instead, the dominant patterns are unidirectional and bidirectional convergence.
Unidirectional convergence occurs when a technology absorbs knowledge from an external domain and applies it internally, or when internal innovations are transferred outward to other fields. This pattern is exemplified by roles such as the “gatekeeper” and “representative” in brokerage theory—entities that control the flow of information between a group and its environment. In the context of AI, unidirectional convergence often manifests in the adaptation of machine learning algorithms to specialized industrial processes or the deployment of AI-driven analytics in medical diagnostics.
Bidirectional convergence, on the other hand, involves a two-way exchange of knowledge between domains. The study distinguishes between two subtypes: advisory-type and liaison-type convergence. Advisory-type convergence occurs when a technology is refined or enhanced by an external domain and then reintroduced to its original context. This reflects the “consultant” role in brokerage theory, where an external expert provides insights that improve internal operations. An example might be the use of biological models to enhance neural network architectures, which are then applied back to computational problems.
Liaison-type convergence, represented by the “liaison” broker role, involves the transfer of knowledge from one domain to another, creating new hybrid applications. This is the most integrative form of technological fusion and requires high levels of knowledge integration capability. It is particularly evident in AI’s expansion into fields like robotics, autonomous vehicles, and smart manufacturing, where advances in sensing, control, and decision-making systems are combined with domain-specific knowledge.
The research finds that liaison-type bidirectional convergence has consistently been the most prevalent mode in AI’s development, followed closely by unidirectional patterns. This suggests that AI does not merely borrow from other fields but actively recombines and repurposes knowledge in transformative ways. The stability of these convergence patterns across all four phases indicates that cross-domain integration has been a defining characteristic of AI since its early stages.
Perhaps the most actionable insight from the study lies in the identification of key technological intermediaries—specific IPC4 categories that repeatedly serve as bridges between domains. Among the most prominent are G06F (electronic data processing), G05B (control systems), G05D (regulation systems), G06T (image data processing), and G06K (data recognition). These categories appear consistently across all convergence modes and throughout the entire time period, underscoring their foundational role in AI’s architecture.
Notably, the study observes the emergence of new intermediaries in later phases, reflecting shifts in the technological landscape. For instance, H01L (semiconductor devices) begins to play a significant role as a liaison broker starting in the 2007–2012 period, highlighting the growing importance of hardware advancements in enabling AI capabilities. Similarly, G06Q (data processing systems for administrative, commercial, or financial data processing) emerges as a key intermediary in the final phase, signaling the increasing integration of AI into business intelligence, fintech, and enterprise resource planning systems.
These findings carry important implications for both corporate strategy and public policy. For firms engaged in AI research and development, the study suggests that focusing on core mediating technologies can yield higher returns on innovation investment. By positioning themselves at the intersection of multiple knowledge domains, organizations can act as brokers themselves, facilitating the flow of ideas and capturing value from cross-sector applications.
Moreover, the persistence of certain technological intermediaries over time indicates path dependence in AI’s evolution—once a particular technology becomes embedded in the innovation network, it tends to remain influential. This implies that early investments in foundational AI technologies can create long-term competitive advantages, reinforcing the importance of sustained R&D efforts in core areas.
From a policy perspective, the research underscores the need for strategic support of cross-disciplinary research initiatives. Governments seeking to accelerate AI adoption should not only fund basic research but also create institutional mechanisms that encourage collaboration between traditionally separate scientific and engineering communities. Programs that promote joint projects between computer scientists, engineers, biologists, and social scientists could help replicate the organic convergence patterns observed in the patent data.
The study also highlights the value of patent data as a real-time indicator of technological change. Unlike academic publications, which often reflect theoretical advances, patents capture applied innovations that are closer to market implementation. By analyzing co-occurrence patterns in patent classifications, researchers can detect emerging convergence trends before they become widely recognized, providing valuable foresight for strategic planning.
One limitation acknowledged by the authors is the inherent bias in patent data toward certain types of innovation—particularly those that are commercially viable and legally protectable. Some forms of AI advancement, especially in open-source software or algorithmic design, may be underrepresented in the dataset. Nevertheless, the sheer volume and global coverage of the Derwent database provide a robust foundation for identifying macro-level trends.
Another consideration is the evolving nature of AI itself. As the field matures, the boundaries between AI and other technologies may become increasingly blurred. What is classified today as an AI patent might tomorrow be seen as a standard component of software engineering or data analytics. This conceptual fluidity challenges traditional classification systems and suggests that future research may need to adopt more flexible, semantic-based approaches to categorizing technological innovation.
Despite these challenges, the study by Chen Yufen and Wang Keping represents a significant step forward in our understanding of how complex technologies evolve through interdisciplinary interaction. By combining rigorous network analysis with a deep engagement with patent data, the authors provide a nuanced picture of AI’s developmental trajectory—one that emphasizes connectivity, integration, and systemic interdependence.
Their work also contributes to a broader conversation about the nature of innovation in the 21st century. In an era where breakthroughs increasingly occur at the boundaries between disciplines, the ability to navigate and exploit cross-domain knowledge becomes a critical success factor. The brokerage roles identified in the study—coordinator, consultant, gatekeeper, representative, and liaison—are not just abstract network positions but real-world functions performed by individuals, teams, and organizations that thrive on boundary-spanning activities.
For businesses, this means cultivating cultures of openness and collaboration, encouraging employees to engage with external knowledge sources, and designing organizational structures that facilitate information flow across departments. For educational institutions, it suggests a need to move beyond siloed curricula and develop interdisciplinary programs that prepare students to operate in hybrid knowledge environments.
Ultimately, the research demonstrates that AI’s power does not reside solely in its algorithms or computational speed but in its capacity to connect, integrate, and transform knowledge from diverse domains. As AI continues to permeate every aspect of modern life—from healthcare and finance to transportation and entertainment—understanding the structural underpinnings of its innovation ecosystem will be essential for harnessing its full potential.
The insights generated by this study are likely to inform not only future academic research but also practical decision-making in both the public and private sectors. By revealing the hidden architecture of AI’s cross-domain convergence, Chen and Wang have provided a valuable roadmap for navigating the complex, interconnected world of technological innovation.
Chen Yufen, Wang Keping, School of Statistics and Mathematics, Zhejiang Gongshang University, Journal of Intelligence, DOI:10.3969/j.issn.1002-1965.2021.07.002