Research on the impact of artificial intelligence on the the manufacturing industry chain resilience

Abstract

Strengthening the robustness and the stability of industrial chains is not only a necessary pathway to achieving the strategic goals of China’s new development paradigm but also a fundamental pillar in building a modern economic system with Chinese characteristics. Drawing on panel data from 30 Chinese provinces spanning 2011 to 2022, this study constructs index systems for artificial intelligence (AI) and manufacturing industry chain resilience (Mir) using the entropy weight method. It systematically examines the effects of AI on Mir and its underlying mechanisms. The key findings are as follows: AI significantly enhances Mir, a result confirmed through multiple robustness tests; Mechanism analysis reveals that AI improves Mir primarily through technological innovation effects and digital empowerment effects; Threshold effects indicate that AI’s impact on Mir exhibits nonlinear characteristics under varying levels of AI development and digital infrastructure; Heterogeneity analysis shows a pronounced structural divergence in AI’s resilience-enhancing effects, with stronger technological empowerment observed in high-resilience cluster, dimensions of adaptive recovery and innovative reconfiguration, and export-oriented industry chains. This study offers a novel theoretical perspective and methodological framework for assessing Mir within the new development paradigm.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-18956-1.

Introduction

The international landscape is undergoing profound transformation. Geopolitical uncertainties are intensifying, and the frequent emergence of “black swan” and “gray rhino” events has led to significant disruptions in the global political and economic order. As major power rivalries escalate and international cooperation mechanisms undergo fundamental shifts, global industrial, supply, and value chains are now entering an unprecedented period of structural reconstruction—comprehensive in scope and profound in impact. The global industrial ecosystem is currently facing severe challenges. On one hand, the wave of “de-globalization” continues to gain momentum, while the COVID-19 pandemic and other global crises have exposed the fragility of global supply chains, amplifying the risks of chain bottlenecks, breaks, and disconnections. On the other hand, the traditional paradigm of “efficiency first” has become increasingly incompatible with the evolving global context. Countries must now explore and establish industrial chain systems that emphasize both operational efficiency and security. Consequently, there is an urgent need across major economies to develop new governance models and strategic approaches that enhance resilience while maintaining global competitiveness. Industrial chain resilience has become a strategic high ground in global power competition. The 20th National Congress of the Communist Party of China explicitly identified “enhancing the resilience and security of industrial and supply chains” as a fundamental strategy for industrial development in the new era¹. Furthermore, General Secretary Xi Jinping has introduced the dialectical development strategy of “strengthening weak links and consolidating strong ones,” and the National Security Strategy (2021–2025) has incorporated industrial shock resistance into the broader national security framework. These initiatives present a uniquely Chinese approach to building a secure, efficient, and resilient modern industrial system. Therefore, exploring the mechanisms and pathways for enhancing industrial chain resilience is not only a strategic response to global restructuring, but also a necessary step toward achieving high-quality economic development and safeguarding national industrial security.

Over the past four decades of reform and opening-up, China’s manufacturing sector has undergone a transformation through the reconfiguration of its dynamic comparative advantages, achieving a leap from low-end embedding in the global value chain to becoming a manufacturing powerhouse. The structure of industrial chains has evolved from a factor endowment–driven, extensive division of labor to a networked system characterized by economies of scale and synergistic effects². However, as global competition increasingly centers on the control of industrial chains, China’s manufacturing industry now faces dual internal and external challenges³: internally, it is constrained by path dependence in technological innovation and the lock-in effect at the low end of the value chain; externally, it is confronted with rising technological barriers and the reshaping of international trade rules. At this critical juncture, the diffusion of AI is driving structural changes in production functions. AI-powered algorithms optimize resource allocation, machine learning enhances industrial chain resilience, and industrial internet technologies facilitate cross-link coordination, providing a new technological paradigm for strengthening industrial chain resilience. In July 2022, the Ministry of Science and Technology in China, in collaboration with five other government agencies, introduced the Guiding Opinions on Accelerating Scenario Innovation to Promote High-Level AI Applications for High-Quality Economic Development. This policy framework marked a new phase of deep integration between China’s AI innovation and industrial development. By the end of 2023, the penetration rate of generative AI among Chinese enterprises had reached 15.2%, contributing to a market size exceeding 14.4 trillion yuan⁴. This underscores AI’s transformative impact on the macroeconomic structure⁵. In advanced manufacturing, AI adoption exhibits clear industry-specific heterogeneity. The new energy vehicle (NEV) sector, for instance, has widely integrated digital twin technology and intelligent decision-making systems, significantly reducing R&D cycles and improving fault prediction accuracy⁶. By strengthening the robustness of the supply chain and reducing external threats, AI offers new possibilities for the modernization and transformation of China’s industrial sector. Yet, a critical question remains: Can AI truly serve as the “breakthrough key” in forging Mir? If so, what are its underlying mechanisms? Addressing these questions is essential not only for understanding how technological dividends translate into industrial competitiveness but also for ensuring self-sufficiency and strategic control of industrial chains under the new development paradigm.

The main limitations of the existing research can be summarized in three aspects. Firstly, while previous literature has explored Mir and the application of AI separately^7–9, there is still a lack of a systematic theoretical framework that links the two, especially with respect to how AI can be embedded into the mechanisms of building supply chain resilience. Secondly, existing studies mainly focus on macro-level resilience measurements while neglecting the heterogeneous responses of different parts of the supply chain (e.g., upstream supply and downstream market) to AI technology. Thirdly, most of the current literature is based on linear assumptions about the impact of technology, without sufficiently considering the moderating effects of external factors such as digital infrastructure, leading to an inadequate understanding of the boundary conditions for AI’s empowerment of resilience. Based on these gaps, this paper addresses the following core questions. Firstly, what is the pathway through which AI affects Mir? This paper proposes a dual-path mechanism of “technological innovation effect” and “digital empowerment effect,” where the former emphasizes how AI enhances the supply chain’s ability to cope with shocks by optimizing the R&D process and fostering new technologies, while the latter focuses on how AI-driven data integration and intelligent decision-making enhance the supply chain’s dynamic adaptability. Secondly, does the impact of AI on supply chain resilience exhibit non-linear characteristics? This study introduces AI level and digital infrastructure level as key threshold variables to explore the differences in AI’s empowerment effects under varying digitalization conditions. Thirdly, does the effect of AI on resilience vary across different supply chain resilience dimensions, internal supply chain segments, and industries with different export dependence? This paper dissects supply chain segments and industry types to uncover the heterogeneous patterns of AI’s impact. This research constructs a multi-dimensional systemic analysis framework that not only effectively addresses the shortcomings in existing theories regarding mechanisms, non-linear relationships, and heterogeneity but also provides a solid theoretical foundation for formulating differentiated policies to enhance industrial chain resilience in the digital economy era. Moreover, it opens new theoretical perspectives for future research and lays the groundwork for a more comprehensive analytical foundation.

This is how the remaining content of the paper is structured. Firstly, the relevant literature is reviewed comprehensively. Secondly, the theoretical framework is delineated and the research hypotheses guiding the study are formulated. Thirdly, the empirical model, methodology, and data are described. Fourthly, the empirical estimation results are presented and their implications are analyzed in detail. Fifthly, mechanism pathways, threshold effects, and heterogeneity analyses are examined. Finally, the study concludes by summarizing the key findings and presenting actionable policy recommendations derived from the empirical analysis.

Literature review

The literature closely related to this study can be categorized into three main areas: research on Mir, research on AI, and research on the relationship between AI and the development of the manufacturing sector.

Firstly, research on Mir primarily focuses on two core directions. The first is the interdisciplinary evolution of the concept of “resilience.” From an etymological perspective, the term originates from the Latin word “resilire”, and was initially introduced in the fields of physics and engineering to describe the ability of a material to recover and adapt under external stress^10,11. In physics, the concept of resilience was first applied in 1973, referring to a system’s resistance to compression following an impact—also known as engineering elasticity¹². As the term migrated into the social sciences, Wink¹³ employed a neoclassical equilibrium framework to define economic resilience as the endogenous capacity of a regional economic system to maintain equilibrium in the face of external shocks. Consequently, related concepts such as social resilience¹⁴, ecological resilience¹⁵, engineering resilience¹⁶, organizational resilience^17,18, and supply chain resilience^19,20 have been successively proposed. The second direction concerns the construction of resilience measurement systems. Currently, there is no consensus within the academic community regarding a unified standard for measuring resilience. Scholars typically select either single or multiple indicators based on the specific object of study and research purpose. Single-dimensional measurement approaches rely on proxy variable theory, using key representative indicators to reflect system resilience. For instance, Markman and Venzin²¹ proposed a method based on volatility and asset return (Volare) to comprehensively assess an organization’s market volatility and long-term profitability, thereby evaluating organizational resilience. Desjardine et al.²² developed a dynamic evaluation system based on loss magnitude and recovery speed to capture a firm’s risk resistance and recovery mechanisms. Brakman et al.²³ constructed a recovery index based on the neoclassical growth model, incorporating an unemployment adjustment factor. In contrast, multi-dimensional approaches follow a composite index framework, applying techniques such as principal component analysis and entropy weighting to integrate multiple indicators. Examples include economic resilience indices^24,25, agricultural resilience indices²⁶, supply chain resilience indices^27,28, and urban resilience indices²⁹.

Secondly, research on AI. As the core driver of a new wave of technological revolution, AI has become a key engine of industrial development^30,31. The rapid advancement of AI not only enhances innovation efficiency but also provides new opportunities for firms to break through bottlenecks, promote industrial upgrading, and improve overall quality and efficiency^32–34. A review of the literature on AI’s effects reveals three main analytical perspectives. The first is from the perspective of labor market restructuring. Scholars are divided into two main viewpoints. On one hand, some argue that the widespread adoption of automation technologies displaces low-skilled labor, leading to decreased labor demand and rising unemployment³⁵. On the other hand, others emphasize that AI creates new employment opportunities by fostering emerging industries and job roles³⁶. The second is from the perspective of economic growth drivers. As a general-purpose technology, AI plays a profound role in driving macroeconomic development. It not only alleviates the labor supply pressure caused by population aging but also fosters the formation of cross-sector collaborative innovation networks, promoting deep integration and diffusion across technologies. At the same time, AI optimizes resource allocation and enhances production efficiency, thereby reshaping the traditional path of total factor productivity growth and emerging as a key engine of sustained economic growth³⁷. The third is from the perspective of global value chain restructuring. Existing studies show that AI, through the co-evolution of technology and institutions, enhances firms’ technological complexity at the micro level and optimizes national positioning in global value chains at the macro level³⁸. This process of digital empowerment restructures global production networks and facilitates the formation of a new international division of labor.

Thirdly, research on the relationship between AI and manufacturing development. Eder et al.³⁹, using data from small and medium-sized enterprises between 2008 and 2015, found that firms employing industrial robots had double the sales of those without and achieved a 5% increase in labor productivity. Moreover, AI significantly promotes technological innovation and produces complementary effects³⁷. Liu et al.⁴⁰, using Panel data from 14 Chinese manufacturing sectors between 2008 and 2017, demonstrated that AI fosters technological innovation, with a more pronounced impact on low-tech sectors. Furthermore, Yang et al.⁴¹ further demonstrate that AI significantly enhances supply chain resilience by promoting diversification, improving efficiency, and reducing inefficient investments, with stronger effects observed in non-state-owned enterprises and when executives have overseas experience. Ma et al.⁴² found that the development of regional AI significantly improves the supply chain’s resilience and recovery capacity, especially through mechanisms that optimize supply-demand matching and improve supply quality, thereby enhancing the stability and ability to cope with external shocks. At the same time, Gupta et al.⁴³ emphasize the important role of AI in sensing the business environment and promoting blockchain deployment. Particularly under the dynamic adjustments in the environment, the synergy between AI and blockchain technology significantly improves the financial resilience of the supply chain, further boosting supply chain resilience. Together, these studies highlight the critical role of AI in enhancing Mir, especially in optimizing production efficiency, driving technological innovation, and improving enterprises’ ability to adapt to external changes. How, then, does AI development specifically influence Mir? From a technological innovation perspective, AI optimizes the allocation of production factors to alleviate labor shortages, enhances the ability to respond to sudden risks via dynamic monitoring and early warning systems, and accelerates technological iteration to overcome bottlenecks in key links, thereby boosting supply chain autonomy and shock resistance. From a data empowerment perspective, AI improves cross-segment collaboration efficiency through big data analytics, reduces coordination costs caused by information asymmetries, enables flexible production adjustments through accurate demand tracking, and enhances resource allocation via intelligent decision-making systems—thus fostering the emergence of resilient supply chain networks that are adaptive to market fluctuations.

Theoretical analysis and research hypotheses

The direct impact of artificial intelligence on manufacturing industry chain resilience

As a foundational technology of Industry 4.0, empowered by the Internet of Things (IoT), AI is deeply embedded in manufacturing supply chains. It not only enhances the intelligence and process efficiency of manufacturing systems but also optimizes industrial structures through “supply chain supplementation”, “supply chain strengthening”, and “supply chain extension”. These processes address weaknesses, expand advantages, and develop emerging sectors, ultimately reinforcing the self-sufficiency and controllability of modern industrial systems^44–46. In terms of complementing the industry chain, AI enhances the substitution and restoration capacity of critical segments by optimizing the labor-capital input structure, effectively addressing weaknesses in the industrial chain and improving its stability and resilience against external shocks. As a new general-purpose technology, AI mitigates factor endowment constraints caused by demographic shifts and reduces the transmission probability of sudden shocks through automated production networks. This effect is particularly evident during public health crises, where intelligent logistics systems and digital twin technologies enhance collaboration among upstream and downstream industrial firms, thereby lowering the risk of supply chain disruptions⁴⁷. Additionally, the superior data expansion and connectivity capabilities of AI significantly reduce factor mobility costs, promoting the large-scale flow of innovation resources across industries and regions. In terms of strengthening the industry chain, AI enhances key segments of the industrial chain by improving production efficiency and technological capabilities. Leveraging industrial big data and deep learning technologies, AI transforms traditionally experience-based tacit knowledge into replicable and applicable digital resources, thereby promoting systematic knowledge accumulation and sharing. Intelligent systems continuously monitor production conditions, dynamically adjust algorithms, and optimize processes, enabling ongoing improvements in manufacturing operations. This continuous optimization drives a shift from experience-driven to algorithm-driven production models, significantly enhancing the overall competitiveness and resilience of the industrial chain. In terms of extending the industry chain, AI facilitates flexible combination and reconfiguration of industrial modules, breaking the limitations imposed by asset specificity in traditional supply chains. With the deep integration of manufacturing and services, AI not only improves value creation efficiency but also fosters broader network-level synergies. The interaction between technology and markets becomes more integrated, promoting both vertical extension and horizontal expansion of industrial chains. This process helps build cross-industry innovation networks, significantly improving the adaptability of industrial chains to uncertainty and continuously strengthening their overall resilience. Consequently, this article posits hypothesis H1:

H1: AI directly enhances Mir.

In accordance with Metcalfe’s Law, the worth of a network grows exponentially with technological progress⁴⁸. This implies that the influence of AI on Mir is not a linear increase but rather a progressive enhancement driven by technological breakthroughs, organizational optimization, and improvements in the external environment. This process exhibits a distinct “network effect”. Firstly, from the perspective of technology diffusion, the initial application of AI often faces high marginal costs, long technology iteration cycles, and uncertain returns. These factors create sunk cost constraints that make firms hesitant to adopt AI, resulting in diminishing marginal effects on supply chain resilience. However, with the coordinated development of related technologies such as big data, machine learning, and the industrial Internet of Things, AI adoption gradually reaches a tipping point, triggering significant network externalities. This markedly improves the alignment between technology and industry, thereby accelerating the enhancement of supply chain resilience. Once firms surpass the inflection point of the S-curve in technology diffusion, their adaptability and risk resistance undergo a fundamental transformation, facilitating a transition toward more technology-intensive and resilient industrial chains. Secondly, from the perspective of organizational adjustment, the early application of AI is often constrained by limited capacity for data utilization and rigid organizational structures, which hinders the realization of collaborative innovation. As data governance frameworks improve and intelligent decision-support systems are implemented, a higher degree of synergy between organizational and digital resources is achieved, producing notable compounding effects. The formation of modular and collaborative organizational networks helps firms overcome path dependence, enhances their responsiveness to risks and environmental changes, and significantly strengthens the resilience of industrial chains. Thirdly, from the perspective of environmental adaptability, in situations where external shocks are frequent and internal technology-organization integration is insufficient, the effectiveness of AI may be limited and even face the risk of technological lock-in. However, when the institutional environment stabilizes and technological capabilities accumulate, firms can enhance their resilience by building appropriate organizational redundancy and storing technological potential. In this context, the benefits of AI are released in tandem with improvements in total factor productivity, driving the industrial chain toward a higher level of resilience and enabling a more sustainable development trajectory. Consequently, this article posits hypothesis H2:

H2: The impact of AI on Mir exhibits a threshold effect. Once the threshold is surpassed, its positive influence on Mir progressively increases.

The indirect impact of artificial intelligence on manufacturing industry chain resilience

In accordance with prior academic findings, this article further explores the transmission mechanism of AI in enhancing Mir from the perspectives of technological innovation effects and digital empowerment effects (shown in Fig. 1).

Fig. 1.

Mechanism diagram of the role of artificial intelligence in the manufacturing industry chain.

Technological innovation effect

Technological innovation is widely regarded as one of the core drivers of value creation⁴⁹. As the core force driving the new round of technological revolution, AI is profoundly reshaping the innovation collaboration structure between upstream and downstream in supply chains, promoting the deep integration and collaborative optimization of the supply chain and innovation chain, thereby significantly enhancing Mir. In terms of complementing the industry chain, AI breaks the constraints of traditional factor endowments, reshaping the elasticity of factor substitution, especially in contexts where labor supply is gradually constrained due to aging, significantly improving the marginal technological substitution rate between labor and capital⁵⁰. Specifically, AI’s penetration features allow enterprises to leverage intelligent sensor networks and big data analysis to comprehensively monitor environmental impacts during the production process, thus optimizing production processes and reducing resource waste⁵¹, thereby promoting technological innovation⁵². This real-time data analysis and feedback adjustment mechanism helps enterprises rapidly respond to changing market environments, enhancing the adaptability and flexibility of the supply chain. In terms of strengthening the industry chain, AI drives the transformation and development of enterprises in key industries, demonstrating significant multiplier effects, especially in overcoming geographical limitations and reducing learning and communication costs between enterprises⁵³. Through technological innovation, enterprises accelerate the integration of cutting-edge knowledge and organizational efficiency, significantly enhancing the collaborative effects between upstream and downstream in the supply chain⁵⁴. Moreover, AI promotes the formation of industrial clusters, triggering economies of scale and knowledge spillovers, providing enterprises with more opportunities for collaboration and innovation. The rise of AI has fostered new industries and business models, greatly improving the operational efficiency of the supply chain. Through forward-looking layouts and active exploration of emerging business models, enterprises can quickly seize future industry development opportunities, gaining a competitive advantage in new markets, thereby driving efficient supply chain operations and enhancing market competitiveness. In terms of extending the industry chain, AI enables enterprises to significantly improve production efficiency through automation and intelligent production, promoting product deep processing and increasing added value, thus extending the supply chain vertically⁵⁵. At the same time, by analyzing large volumes of user data, AI optimizes supply-demand matching, creating a flexible production model, and horizontally expanding the supply chain at the enterprise level, creating new competitive advantages. Based on this, we propose hypothesis H3::

H3: AI indirectly enhances Mir by driving technological innovation.

Digital empowerment effect

Digital elements, as important innovative resources for the restructuring and upgrading of the supply chain, with their independent knowledge creation and intelligent characteristics, are gradually becoming the key driving force for industry transformation and economic vitality. AI applications, based on extensive computing resources⁵⁶, have the ability to collect, process, analyze, and synthesize large amounts of data from various sources, providing precise decision support for enterprises⁵⁷. Intelligent big data processing and precise decision-making improve the collaborative efficiency of supply chain segments, facilitate the digital transformation of the supply chain, and enhance innovation capabilities, thereby improving supply chain adaptability and risk resilience. In terms of complementing the industry chain, enterprises efficiently integrate data with AI, reshaping internal collaboration mechanisms, optimizing production efficiency and resource allocation. Digital technologies accelerate the integration of production factors, improving enterprises’ production efficiency and driving entrepreneurial activities and economic model transformation in urban areas⁵⁸. By deeply mining production data, enterprises optimize processes, not only improving decision-making precision but also enhancing the stability of the supply chain through resource reallocation, thus improving its resilience and adaptability. In terms of strengthening the industry chain, AI, leveraging the network collaboration effect of the industrial internet platform, reconstructs the Nash equilibrium between upstream and downstream in the supply chain. The introduction of digital elements makes the flow of information within the supply chain more efficient and real-time, significantly reducing the cost of coordination across segments. In this process, intelligent decision-making not only drives continuous optimization of production processes but also promotes the collaborative development of industrial clusters, improving the overall competitiveness of the supply chain. In terms of extending the industry chain, enterprises utilize AI to mine demand-side big data, triggering the long-tail effect, driving both horizontal expansion and vertical deepening of the supply chain. During the digital transformation process, digital technologies deeply embed into key areas such as enterprise R&D, product architecture, production operations, sales services, and management systems, facilitating the optimal allocation of resources through efficient data circulation and sharing, while significantly improving enterprise production efficiency and operational effectiveness⁵⁹. The digital production model enables enterprises to more accurately respond to market changes, enhancing their ability to adapt to market fluctuations through the construction of an innovation ecosystem, ultimately improving the supply chain’s coverage and resilience. Based on this, we propose hypothesis H4:

H4: AI indirectly enhances Mir by leveraging digital empowerment effect.

The nonlinear impact of artificial intelligence on manufacturing industry chain resilience

The development and optimization of digital infrastructure are essential for enhancing supply chain resilience. They serve as a critical foundation for ensuring stable operations under unconventional shocks and risk environments⁵⁴. When digital infrastructure remains at a low equilibrium state, the ability of AI to strengthen Mir is significantly constrained. Due to poor data circulation and insufficient computing infrastructure, it is difficult to generate effective network effects, which limits the full potential of AI in optimizing production efficiency, improving resource allocation, and enhancing upstream-downstream coordination. At the same time, enterprise digital transformation is often hindered by the “digital divide” and organizational inertia, leading to a limited application of AI that remains at the basic level of automation without achieving intelligent upgrading. This shallow level of technological adoption not only intensifies supply chain fluctuations but also weakens the ability to respond to diverse market demands, potentially resulting in an industrial chain that is “highly connected but fragile and unstable.”

Once digital infrastructure surpasses a critical threshold, the enabling effect of AI on Mir exhibits increasing marginal returns. As a vital channel for information dissemination, digital infrastructure drives the metamorphosis of conventional industries toward digitalization, informatization, and intelligentization^60,61. Via profound interaction with supply chains, it enhances production transparency, optimizes resource allocation, and significantly strengthens supply chain resilience^62,63. According to Romer’s endogenous growth theory, the advancement of digital infrastructure—particularly in computing power (cloud computing), connectivity (5G IoT), and data systems (industrial internet)—creates a synergistic effect. By improving cross-domain data mobility, reducing information entropy, and strengthening network externalities, it effectively addresses the Solow productivity paradox. At this stage, AI optimizes production function parameters through deep learning algorithms, enables Pareto improvements via digital twin-based dynamic supply chain simulations, and achieves Nash equilibrium through blockchain-based smart contracts. Notably, AI facilitates the integration of traditional and emerging production models, accelerating the transition toward advanced manufacturing. The development of digital infrastructure removes geographical constraints, expands the boundaries of supply chain collaboration, and supports the upgrading of manufacturing industries within global value chains. At this stage, AI’s impact on supply chain resilience follows a pattern of increasing marginal returns. As digital infrastructure continues to improve, its enabling role strengthens, fostering supply chain optimization and sustainable growth. Consequently, this article posits hypothesis H5:

H5: The impact of AI on Mir is influenced by digital infrastructure, exhibiting a threshold effect.

Model construction and variable selection

Model construction

The paper builds the benchmark regression model below to provide light on how AI affects Mir:

1

To further investigate the pathway through which AI influences Mir, this study establishes the following mechanism effect model:

2

3

4

To investigate whether the impact of AI development on Mir demonstrates nonlinear characteristics contingent on different threshold variables, this study formulates the following threshold effect model:.

5

In the formula: represents manufacturing industry chain resilience; represents artificial intelligence; is control variables that affect Mir; is a mechanism variable (including innovation activity and level of digital elementization); is the threshold variable. I() represents the indicator function, where the condition in parentheses is 1, otherwise it is 0. In subsequent empirical studies, the econometric testing model can be extended to multiple thresholds based on actual conditions.To eliminate the influence of individual characteristics and time trends of provinces on Mir, the model introduces individual and time effects of provinces, where denotes the province-specific fixed effects, captures the year-specific fixed effects, and represents the random error term, accounting for unobserved heterogeneity and temporal variations in the model.

Variable selection

Explained variables

This study adopts manufacturing industry resilience (Mir) as the dependent variable. Following relevant literature^64–66, we construct a resilience index system from four key dimensions: impact resistance capability, adaptive recovery capability, innovative reconfiguration capability, transformative development capability. The entropy weight method is applied for indicator weighting (shown in Table 1). Firstly, shock resistance reflects the industry’s ability to withstand external pressures at the onset of a crisis—this constitutes the first line of defense for resilience. Secondly, recovery and adaptation capability measures how quickly and effectively the industrial chain returns to normal operation after disruption—this is the core manifestation of resilience. Thirdly, innovation and reorganization ability assess the industrial chain’s capacity to upgrade through technological innovation and resource reallocation—this determines the sustainability of resilience. Fourthly, transformation and development potential evaluates the ability of the industrial chain to break through existing models and achieve higher-level development—this is critical for long-term resilience. These four dimensions jointly cover the full cycle from short-term response to long-term development and form a progressive structure from passive defense to proactive upgrading. They allow for a comprehensive and dynamic assessment of industrial chain resilience. The calculation results are shown in Table 1.

Table 1.

Resilience index system for manufacturing industry chain.

Indicator system	First level indicator	Second level indicator	Third level indicator	Specific measurement	Direction	Weight
Manufacturing industry chain resilience	Impact resistance capability	Industrial stability	Total profit of manufacturing industry	Total profit of high-tech manufacturing industry	+	0.0006
			Profit margin of manufacturing enterprises	Combined profits of major industrial enterprises/Main operational costs of major industrial enterprises	+	0.0022
		Stable employment	Employment in manufacturing industry	Workforce participation in urban manufacturing industries	+	0.1202
			Rate of unemployment	Officially recorded urban unemployed individuals/Resident population count at the close of the year	-	0.0037
	Adaptive recovery capability	Production recovery	Growth rate of manufacturing industry	Annual industrial value-added for this year/Annual industrial value-added for the previous year	+	0.0119
			Labor productivity in manufacturing industry	Industrial production value-added/Average workforce size in the manufacturing sector	+	0.0726
			Advanced structure of manufacturing industry	Operating revenue of the high-tech sector/Operating earnings of industrial firms exceeding the size threshold	+	0.2213
			Proportion of added value in the tertiary industry	Value added of the tertiary industry/Value-added generated by the secondary industry	+	0.0592
		Capital guarantee	Financial development level	Total deposits and loans held by financial institutions	+	0.0879
			Social investment level	Expenditure on fixed assets	+	0.0676
			Government assistance level	Budgetary outlays of local governments	+	0.0443
	Innovative reconfiguration capability	R&D investment	R&D investment intensity	R&D outlays by industrial firms exceeding the designated size/Primary operational revenue of large-scale industrial firms	+	0.0447
		Innovation output	Per capita number of invention patents	Total invention patents granted/R&D personnel tally in large-scale industrial firms	+	0.0409
		Talent support	R&D personnel investment intensity	Number of R&D personnel in industrial enterprises above a certain scale/Workforce count in the manufacturing sector	+	0.1519
	Transformative development capability	Green development	Unit value-added energy consumption	Coal consumption volume/Industrial added value	-	0.0118
			Unit value-added sulfur dioxide emissions	Sulfur dioxide emissions/Industrial added value	-	0.0050
			Unit value-added industrial solid waste generation	Comprehensive utilization of industrial solid waste/Industrial added value	-	0.0036
			Chemical oxygen content of wastewater with unit added value	Chemical oxygen content of wastewater/Industrial added value	-	0.0074
		Consumption potential	Urbanization rate	Urban population/Year-end permanent resident population	+	0.0255
		Consumption level	Engel’s coefficient	Urban households’ Engel coefficient	+	0.0176

Firstly, eliminate the dimensional interference of the impact resistance capability, adaptive recovery capability, innovative reconfiguration capability and transformative development capability, and transformational development capacity, and standardize them using the following formula:

6

Among them, represents the initial indicator value, i and j are the region and indicator respectively, and n represents the number of indicators, indicating that the indicator value has been standardized.

Secondly, based on the standardized data, calculate the normalized weight of the j-th indicator in the i-th sample:

7

Then calculate the information entropy of the j-th indicator:

8

Among them, k is the normalization coefficient used to ensure that the entropy value is between 0 and 1. The larger the entropy value, the more evenly distributed the information of the indicator, and its contribution to the uncertainty of the system is smaller, and vice versa.

Thirdly, calculate the entropy weight of each indicator based on information entropy:

9

Fourthly, by combining the impact resistance index, recovery and adaptation index, innovation and restructuring index, and transformation and development index, the comprehensive index is calculated to obtain the resilience evaluation values of the manufacturing industry chain in different regions. The calculation formula is as follows:

10

Based on the entropy weight method (EWM) evaluation framework, we further observed that from 2011 to 2022, the resilience of the manufacturing industry chain and its four key dimensions (impact resistance capability, adaptive recovery capability, innovative reconfiguration capability, transformative development capability) showed a steady upward trend (as shown in Fig. 2). Specifically, impact resistance capability has strengthened in response to improvements in global supply chain governance, enhanced autonomy in critical industrial segments, and precise industrial policy interventions, demonstrating a heightened capacity to withstand external disruptions. Adaptive recovery capability has significantly improved under the impetus of digital empowerment, the widespread adoption of smart manufacturing technologies, and enhanced supply chain coordination, reflecting the industry’s ability to self-recover and dynamically adjust in response to market fluctuations. Innovative reconfiguration capability has been bolstered by sustained increases in R&D investment, the refinement of collaborative innovation networks across upstream and downstream industries, and breakthrough applications of emerging technologies, driving structural optimization and reinforcing long-term competitiveness. Transformative development capability has been continuously strengthened by the transition toward greener, smarter, and more advanced industrial paradigms, underscoring the manufacturing sector’s intrinsic momentum toward breaking developmental path dependence and advancing high-quality growth. Notably, the varying growth rates and periodic fluctuations across these resilience dimensions indicate significant heterogeneity, suggesting that the manufacturing sector exhibits distinct structural adaptation characteristics when responding to different types of shocks. In this context, our study further explores the underlying mechanisms shaping manufacturing resilience and its asymmetric evolutionary patterns, offering theoretical insights and policy recommendations for building a more shock-resistant, adaptable, and innovation-driven manufacturing ecosystem.

Fig. 2.

The resilience evolution and dimension growth trend of China’s manufacturing industry chain from 2011 to 2022.

Core explanatory variables

This study selects the level of AI development as the core independent variable. Drawing on relevant literature⁶⁷, we measure AI development across provinces using the entropy method, based on three dimensions: infrastructure, production application, and social benefits (shown in Table 2). Infrastructure focuses on assessing the technological foundation and hardware facilities that support AI development. Production application evaluates the practical impact of AI in economic activities, reflecting its role in enhancing industrial performance and efficiency. Social benefits capture AI’s contribution to social welfare and sustainable development. Together, these three dimensions comprehensively reveal the overall influence of AI on regional development, highlighting its multifaceted role in driving technological progress, economic transformation, and societal well-being.

Table 2.

Index system for the development level of artificial intelligence.

Indicator system	First level indicators	Second level indicators	Specific measurement	Direction
Artificial intelligence	Basic construction	Internet infrastructure investment	Fiber optic line length/Provincial area	+
		Intelligent funding investment	R&D funds for high-tech enterprises	+
		R&D talent investment	R&D personnel in high-tech enterprises	+
		Investment in smart devices	Fixed asset expenditure in the software and information technology services industry	+
	Production application	Software development and application	Software product revenue/Essential business revenue of industrial enterprises	+
		Intelligent product development	Embedded system business revenue/Principal operating income of industrial firms	+
		Development of Intelligent Enterprises	Main business income of high-tech enterprises/Main business income of industrial enterprises	+
		Application of intelligent technology	Sales of new products in high-tech industries/Core business-related income of industrial enterprises	+
	Social effect	Innovation ability	Number of authorized national patent applications/R&D personnel at that time	+
		Market profit	Overall profitability of high-tech manufacturing enterprises	+
		Economic effect	Quantify the asset contribution rate and cost utilization rate in each province	+
		Social effect	Evaluate the energy consumption per unit of GDP across provinces (electricity and coal as two types of energy)	+

Control variables

This study includes five control variables. The level of economic development (Lngdp) is measured by the natural logarithm of regional per capita GDP⁶⁸, reflecting the overall economic performance of a region. A higher economic development level often correlates with stronger market demand, better infrastructure, and richer innovation resources, thereby supporting the growth and competitiveness of the manufacturing sector. Government support intensity (Gov) is measured by the ratio of local general budget expenditure to regional GDP³⁸.  Stronger public service provision enhances administrative efficiency and resource allocation, boosting manufacturing productivity and competitiveness. The level of openness (Ope) is assessed by the proportion of total imports and exports in regional GDP. While greater openness can stimulate competition, it may also expose regions to external market fluctuations. Overreliance on foreign markets weakens supply chain autonomy, limiting resilience under global shocks⁶⁹. Environmental regulation intensity (Envir) is measured by the ratio of investment in industrial pollution control to industrial value added. Stronger regulation may raise costs due to investment in cleaner technologies but also encourages innovation and efficiency, supporting a green transformation. Infrastructure level (Instra) is proxied by per capita road area⁷⁰, indicating the development of transport and logistics systems. Better infrastructure reduces logistics costs and enhances manufacturing efficiency and market competitiveness.

Mechanism variables

This study includes two mechanism variables. Innovation activity (Innov) is measured by the ratio of total patent applications to the year-end urban resident population, reflecting the intensity of regional innovation activities. A higher level of innovation activity contributes to increased frequency and quality of technological innovation, thereby promoting technological progress and enhancing the competitiveness of the manufacturing sector^71,72. Digital factorization level (Digfac) is proxied by software business revenue, representing the scale and development level of the regional digital industry. A higher level of digital factorization facilitates the widespread application of information technologies⁷³, supporting the digital transformation and overall efficiency of the manufacturing industry.

Threshold variables

This study selects two threshold variables. The first is the level of AI, with its measurement system detailed in Table 2. Based on relevant literature⁵⁴, the second threshold variable—Digital New Infrastructure (Digfra)—is evaluated from three dimensions: information infrastructure, integrated infrastructure, and innovation infrastructure, using the entropy weight method to construct the indicator system (shown in Table 3). Information infrastructure provides efficient support for data collection, transmission, and processing, thereby facilitating the deep application of intelligent manufacturing and the industrial internet, and enhancing real-time monitoring and refined management in production processes. Integrated infrastructure improves resource allocation efficiency and strengthens coordination across production stages through highly interconnected technologies and platforms. Innovation infrastructure offers stronger support for enterprise R&D activities and stimulates the application of emerging technologies. Together, these factors improve production efficiency, reduce operational costs, and enhance the adaptability and risk resistance of manufacturing enterprises in global competition.

Table 3.

Index system for digital new infrastructure construction.

Indicator system	First level indicators	Second level indicators	Direction
Construction of new digital infrastructure	Information infrastructure	Length of optical cable line per square kilometer	+
		Per capita internet port	+
		Mobile phone base station per square kilometer	+
		Mobile internet penetration	+
		Proportion of internet access users	+
		Per capita number of domain names	+
	Converged infrastructure	Per capita operating length of public trams	+
		Per capita railway operating mileage	+
		Per capita road length	+
		Per capita highway mileage	+
		Proportion of e-commerce enterprises	+
		Proportion of personnel in the information industry	+
		Per capita e-commerce sales revenue	+
		Per capita software business revenue	+
	Innovation infrastructure	Proportion of R&D personnel	+
		Proportion of scientific and technological expenditures	+
		R&D expenditure intensity	+
		Per capita number of patent applications	+

Data sources

This study employs panel data from 30 Chinese provinces (excluding Tibet) spanning 2011–2022 to analyze the impact of AI on Mir. Data sources include the National Bureau of Statistics database, provincial statistical yearbooks (covering autonomous regions and centrally administered municipalities), and various national statistical yearbooks such as the China Industrial Statistical Yearbook, China Science and Technology Statistical Yearbook, and China Energy Statistical Yearbook. To maintain data continuity and completeness, missing values are imputed using linear interpolation. Table 4 presents the descriptive statistics for the key variables included in our analysis, summarizing their central tendencies, dispersion, and distributional properties to provide a comprehensive overview of the dataset.

Table 4.

Descriptive statistics.

Variable	Sample size	Mean	Standard deviation	Minimum value	Maximum value
Mir	360	0.192	0.0960	0.0598	0.649
AI	360	11.21	10.49	0.337	61.85
Lngdp	360	10.87	0.461	9.682	12.15
Gov	360	0.259	0.111	0.105	0.758
Envir	360	0.00281	0.00289	6.24e-05	0.0280
Ope	360	0.272	0.281	0.00763	1.464
Instra	360	16.55	5.058	4.040	28
Digfra	360	0.194	0.151	0.0189	1
Innov	360	14.18	16.63	0.860	92.82
Digfac	360	0.188	0.327	2.68e-05	2.250

Empirical analyses

Benchmark regression

Based on the Hausman test results, this paper uses a fixed-effects model for baseline regression analysis to control for estimation bias caused by individual heterogeneity. The empirical results presented in Table 5 indicate that the effect of AI on Mir is statistically significant in both the simple model (Column 2) and the extended model that includes control variables (Column 4). This finding echoes the conclusions of Li et al. (2025)⁷⁴ regarding the positive impact of digital transformation on enhancing supply chain resilience, providing strong support for the validity of hypothesis H1.

Table 5.

Benchmark regression results.

	(1)	(2)	(3)	(4)
Variable	Mir	Mir	Mir	Mir
AI	0.0092***	0.0081***	0.0080***	0.0075***
	(23.43)	(12.89)	(11.46)	(10.51)
Lngdp			0.0387***	−0.0040
			(3.57)	(−0.14)
Gov			−0.0098	0.0528
			(−0.23)	(0.70)
Ope			0.0009	−0.0095
			(0.04)	(−0.49)
Envir			0.3256	0.3979
			(0.82)	(1.00)
Instra			−0.0020**	−0.0021**
			(−2.27)	(−2.14)
Observations	360	360	360	360
Fixed urban effect	YES	YES	YES	YES
Fixed year effect	NO	YES	NO	YES
R2	0.863	0.884	0.872	0.893

AI directly enhances industrial chain resilience through its data processing capabilities, resource optimization, and improved decision-making efficiency. Specifically, AI technology, through high-frequency data analysis and intelligent decision support systems, allows businesses to monitor and predict potential risks in the production process in real time, enabling quick adjustments and responses. When facing market demand fluctuations, technological changes, or unforeseen events, AI can optimize production scheduling and resource allocation through adaptive algorithms, increasing production efficiency and reducing the negative impact of supply chain disruptions or external shocks. This flexible management mechanism provides the manufacturing supply chain with enhanced adaptability, significantly strengthening its resilience. Furthermore, AI plays a crucial role in driving the digital transformation of the supply chain. As globalization accelerates and market demands and technology continuously evolve, traditional manufacturing industries face increasingly complex competitive pressures and technological challenges. AI, by deeply mining and utilizing big data, helps companies quickly identify market changes and technological trends, promoting innovation and technological advancement, thereby enhancing the technical adaptability and flexibility of the supply chain. This technological innovation not only improves the collaborative effects among supply chain segments but also provides strong support for the long-term competitiveness of the manufacturing sector.

Endogeneity discussion

This study may face potential endogeneity issues from two main sources. Firstly, there may be a bidirectional causal relationship between AI and Mir. On one hand, AI technologies can significantly enhance supply chain resilience through intelligent forecasting to reduce risks, flexible resource allocation enabled by the industrial internet, and inventory optimization via machine learning⁷⁵. On the other hand, a resilient industrial chain often features more advanced data infrastructure, a richer accumulation of industrial data, and more efficient innovation collaboration mechanisms, all of which, in turn, provide essential support for the iterative upgrading of AI technologies. Secondly, the model may suffer from omitted variable bias—failing to include key factors that affect both AI applications and industrial chain resilience, such as industrial policy environments or regional innovation capacity. This omission could result in biased causal inference. To ensure the robustness and reliability of our findings, we employ several methods to systematically test for potential endogeneity, including instrumental variable (IV) approaches and lagged variable analysis.

Firstly, existing literature commonly adopts the lag of the endogenous variable as an instrument⁷⁶. This choice is based on two considerations: AI development generally follows a time trend, and current AI levels are strongly correlated with future developments; meanwhile, future AI levels are unlikely to directly affect past levels of industrial chain resilience. Therefore, we use the one-period lag of AI as an instrumental variable. Secondly, following the approach of Acemoglu and Restrepo⁷⁷, we use the density of industrial robot installations as an instrument for AI, and conduct two-stage least squares (2SLS) regression to address potential endogeneity bias. Thirdly, we re-estimate the model by lagging all explanatory variables by one period⁷⁸. As shown in Table 6, the results from these endogeneity tests demonstrate that, after addressing the potential bias, the main findings remain robust and unaffected by endogeneity concerns.

Table 6.

Endogenous test results.

Variable	2SLS		2SLS		Explanatory variable lags behind by one period
IV	1.0487***		0.0002***
	(85.92)		(26.06)
AI		0.0086***		0.0093***
		(32.48)		(30.49)
L.AI					0.0075***
					(9.79)
Observations	330	330	360	360	330
Controlled variable	Yes	Yes	Yes	Yes	Yes
Fixed urban effect	Yes	Yes	Yes	Yes	Yes
Fixed year effect	Yes	Yes	Yes	Yes	Yes
R2	0.984	0.899	0.866	0.897	0.875

Robustness test

Firstly, sample data selection. To minimize possible prejudice and interference from outliers in the sample, this study applies 1% and 99% winsorization to the data, minimizing the distortion and systematic bias in the regression results. Table 7, Column (1) exhibits the regression findings. Secondly, exclusion of municipalities. Given that municipalities may receive more policy dividends due to their unique geographic, economic, and policy positions, potentially causing significantly different effects on Mir compared to other provinces, we exclude these municipalities from the sample to avoid potential interference or bias in the analysis. Table 7, Column (2) exhibits the regression findings. Thirdly, modification of the time window. Considering that the COVID-19 Pandemic in 2020 may have impacted Mir, we exclude the sample for this year and reset the time window for the regression analysis. Table 7, Column (3) exhibits the regression findings. Fourthly, change in estimation method. In the baseline regression analysis, we first apply a fixed-effects panel model for estimation and conduct robustness checks by replacing it with a random-effects panel model to ensure the reliability and consistency of the results. Table 7, Column (4) exhibits the regression findings. Fifthly, substitution of core explanatory variables. To strengthen the robustness and reliability of our empirical results, we utilize principal component analysis (PCA) to develop a comprehensive composite index that captures the multifaceted dimensions of AI development, replacing the entropy method used previously. Table 7, Column (5) exhibits the regression findings. The outcomes of these robustness checks consistently show that AI has a noteworthy favorable effect on Mir, demonstrating the validity of the analysis’s findings.

Table 7.

Robustness test results.

Variable	(1)	(2)	(3)	(4)	(5)
	Sample data screening	Exclude municipalities directly under the central government	Change time window	Change estimation method	Replace core explanatory variables
AI	0.0073***	0.0080***	0.0073***	0.0079***	0.0754***
	(11.70)	(10.45)	(10.11)	(14.42)	(4.20)
Observations	344	312	330	360	360
Controlled variable	Yes	Yes	Yes	Yes	Yes
Fixed urban effect	Yes	Yes	Yes	Yes	Yes
Fixed year effect	Yes	Yes	Yes	Yes	Yes
R2	0.884	0.899	0.895	0.885	0.812

Further analyses

Quantile regression

This study employs the quantile regression method to systematically examine the impact of AI on Mir at different quantiles (20%, 40%, 60%, 80%). The empirical results in Table 8 show that the regression coefficients are significantly positive at all quantiles, providing strong evidence that artificial intelligence consistently strengthens the resilience of manufacturing supply chains. Whether at the early stage of resilience cultivation in industry chain segments (20% quantile), at the key transformation and upgrading period in mid-stream industries (40%−60% quantile), or at the leading industries with strong competitive advantages (80% quantile), AI demonstrates a significant promoting effect. This consistent impact across different stages of development strongly validates the universality of its empowering effect. Specifically, in the relatively weak segments of the industry, AI drives resilience through dual mechanisms of technological penetration and digital transformation, effectively strengthening the foundation for risk resistance. In the mid-stream sector during the industrial upgrading process, its value mainly lies in process reengineering and system collaboration. For high-end industries, its prominent advantages are seen in leading technological innovation and optimizing intelligent decision-making. By innovatively using the quantile regression method, this study not only confirms the widespread existence of the empowering effect of AI but also delves into the differences in its strength across different stages of development. This provides a more precise theoretical basis and practical guidance for formulating differentiated industrial digital transformation policies.

Table 8.

Quantile regression results.

Variable	20%	40%	60%	80%
AI	0.0085***	0.0084***	0.0085***	0.0080***
	(57.57)	(22.17)	(30.66)	(34.51)
Lngdp	−0.0143	−0.0025	0.0040	0.0288***
	(−1.58)	(−0.36)	(0.40)	(5.87)
Gov	−0.0336	−0.0624**	−0.0713***	−0.0657***
	(−0.87)	(−2.25)	(−2.81)	(−2.96)
Ope	−0.0183	−0.0263**	−0.0216**	−0.0159
	(−1.49)	(−1.98)	(−2.22)	(−1.43)
Envir	−0.9288	−1.1872	−0.6282	−0.2568
	(−1.50)	(−1.55)	(−0.64)	(−0.37)
Instra	0.0006	0.0008	0.0008	0.0010
	(0.80)	(1.30)	(0.94)	(1.63)
Observations	360	360	360	360
Pseudo R2	0.5276	0.6022	0.6724	0.7563

Mechanism test

As discussed in the previous theoretical framework, AI may influence Mir through innovation-driven effects and the digital factorization effect. Based on the control of various control variables, the direct and indirect impacts of AI on Mir, along with the corresponding test results, are presented in Table 9.

Table 9.

Mechanism test results.

Variable	(1)	(2)	(3)	(4)	(5)
	Mir	Innov	Mir	Digfac	Mir
AI	0.0075***	0.8536***	0.0067***	0.0369***	0.0054***
	(14.14)	(6.50)	(11.20)	(10.27)	(7.28)
Innov			0.0009***
			(3.53)
Digfac					0.0555***
					(3.07)
Observations	360	360	360	360	360
Controlled variable	Yes	Yes	Yes	Yes	Yes
Fixed urban effect	Yes	Yes	Yes	Yes	Yes
Fixed year effect	Yes	Yes	Yes	Yes	Yes
R2	0.978	0.945	0.979	0.934	0.980
Bootstrap verification (1000 times)		ind_eff[0.0017548, 0.0038725] dir_eff[0.0003074, 0.0013393]		ind_eff[0.1030996, 0.1644467] dir_eff[0.0539509, 0.1338946]

Columns (2) and (3) in Table 9 explore the transmission mechanism of AI’s impact on Mir, using innovation activity as a mediating variable. In Column (2), the resilience indicator for Mir is significantly positive, indicating that AI has a significant promoting effect on Mir. In Column (3), the estimated coefficients for both explanatory variables are positive and statistically significant. Upon further relaxation of the assumptions and performing a bootstrap test, the confidence interval for the indirect effect is [0.0017548, 0.0038725], and the confidence interval for the direct effect is [0.0003074, 0.0013393], neither of which includes 0, confirming the existence of a mediating effect. Table 10 presents the Sobel test results for the mechanism effect. In the Sobel test, the indirect effect of AI on Mir through the innovation-driven effect is 0.002814, the direct effect is 0.000823, and the proportion of the mechanism effect is 77.36%. The Sobel test’s Z-value is 9.836, indicating that the innovation-driven effect is indeed the underlying mechanism through which AI enhances Mir, thereby validating hypothesis H3.

Table 10.

Sobel test results.

Mechanism variable	Indirect effect	Direct effect	Total effect	Mechanism effect proportion	Z statistics
Innov	0.002814***	0.000823***	0.003637***	77.36%	Z = 9.836(Effective mechanism)
	(0.000286)	(0.000201)	(0.000321)
Digfac	0.133773***	0.093923***	0.227696***	58.75%	Z = 13.31(Effective mechanism)
	(0.010052)	(0.010645)	(0.008321)

Columns (4) and (5) in Table 9 explore the transmission mechanism of AI’s impact on Mir, using the level of digital factorization as the mediating variable. In Column (4), the resilience indicator for Mir is significantly positive, indicating that AI significantly promotes Mir. In Column (5), the coefficients of both explanatory variables are significantly positive. Upon relaxing the assumptions and performing a bootstrap test, the confidence interval for the indirect effect is [0.1030996, 0.1644467], and the confidence interval for the direct effect is [0.0539509, 0.1338946], neither of which includes 0, confirming the existence of the mediating effect. As shown in Table 10, in the Sobel test, the indirect effect of AI on Mir through the digital empowerment effect is 0.133773, the direct effect is 0.093923, and the proportion of the mechanism effect is 58.75%. The Sobel test’s Z-value is 13.31, indicating that the digital empowerment effect is indeed the underlying mechanism through which AI enhances Mir, thereby validating hypothesis H4.

Threshold effect test

This study first employs the Bootstrap method to conduct a significance test for threshold effects in the panel data, using 500 bootstrap replications to test the existence and specific threshold values of the threshold effects. The test results are shown in Table 11. Specifically, the level of AI passes the 1% confidence level test under a single threshold condition, with a threshold value of 25.6815. Similarly, digital new infrastructure also passes the 1% confidence level test under a single threshold condition, with a threshold value of 0.0331. These results indicate the presence of significant threshold effects for both AI and digital new infrastructure, with their respective threshold values having passed rigorous statistical tests, thus providing a reliable empirical foundation for further analysis of the nonlinear relationship between these factors and Mir.

Table 11.

Threshold effect test results.

Threshold variable	Inspection type	F value	P value	Critical value			Threshold value	95% confidence interval
				10%	5%	1%
AI	Single threshold	39.75	0.0000	18.5040	21.2163	30.4601	25.6815	[24.1038, 26.1155]
Digfra	Single threshold	52.42	0.0000	16.6399	19.2657	28.2486	0.3539	[0.3420, 0.3746]

Separate single-threshold regression models were established for both AI levels and digital new infrastructure, with the panel threshold regression results presented in Table 12. Specifically, Column (1) shows the regression results using the level of AI as the threshold variable. When the level of AI is within the range of Th1 ≤ 25.6815, its effect on enhancing Mir is 0.0062958. When the level of AI exceeds Th1 > 25.6815, the effect further intensifies to 0.0076141, with both results being statistically significant at the 1% level. These empirical findings suggest that higher levels of AI significantly strengthen its positive effect on Mir, providing further support for the validity of hypothesis H2. Column (2) presents the regression results with digital new infrastructure as the threshold variable. When the level of digital new infrastructure is within the range of Th2 ≤ 0.3539, the effect of AI on enhancing Mir is 0.0067583. When the level of digital new infrastructure exceeds Th2 > 0.3539, the effect increases to 0.007821, with both results being statistically significant at the 1% level. These empirical results indicate that higher levels of digital new infrastructure significantly enhance the positive effect of AI on Mir, adding to the evidence that hypothesis H5 has been verified.

Table 12.

Threshold effect regression results.

Threshold variable	(1)	(2)
	Th1 = AI	Th2 = Digfra
Th1 ≤ 25.6815	0.0062958***
	(0.0008492)
Th1 > 25.6815	0.0076141***
	(0.0007729)
Th2 ≤ 0.3539		0.0067583***
		(0.000557)
Th2 > 0.35391		0.007821***
		(0.0004917)
Observations	360	360
Controlled variable	Yes	Yes
R2	0.8612	0.8402

Heterogeneity test

Heterogeneity of resilience dimensions in the industrial chain

Based on the hypothesis of supply chain resilience threshold heterogeneity, this study divides the sample regions into high-resilience cluster and low-resilience cluster according to Mir and performs group regression, as shown in Table 13. The empirical results indicate that AI has a statistically significant positive effect only in the high-resilience cluster. The underlying reason lies in the fact that high-resilience supply chains demonstrate greater adaptability and flexibility, enabling them to quickly adjust and optimize resource allocation when faced with external changes, thereby facilitating the rapid absorption and efficient application of technology. Through the collaborative configuration of resources and the continuous accumulation of technology, high-resilience supply chains not only enhance the overall efficiency of resource allocation but also provide ample room for AI to fully realize its potential in improving production process optimization and boosting industrial efficiency. In contrast, low-resilience supply chains, when confronted with issues such as uneven resource allocation, technological path dependency, and insufficient technological absorption capacity, experience a significant slowdown in the diffusion of technological innovation. As a result, the potential benefits of AI are not effectively realized, leading to diminishing marginal returns and ultimately reducing its impact on improving supply chain resilience.

Table 13.

Heterogeneity test results.

Variable	Heterogeneity of resilience dimensions in the industrial chain		Heterogeneity of resilience dimensions in the industrial chain				Heterogeneity of industrial chain export dependence in the industrial chain
	High resilience cluster	Low resilience cluster	Dkl	Hfl	Czl	Fzl	Extraverted type	Introverted type
AI	0.0068***	0.0013	0.0003	0.0052***	0.0020***	−0.0000	0.0074***	0.0060***
	(7.04)	(1.35)	(1.02)	(14.23)	(5.62)	(−0.48)	(9.31)	(5.53)
Observations	180	180	360	360	360	360	180	180
Controlled variable	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Fixed urban effect	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Fixed year effect	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
R2	0.928	0.763	0.258	0.906	0.511	0.710	0.915	0.870

Heterogeneity of internal resilience in the industrial chain

This study, based on the supply chain resilience deconstruction framework, decomposes the resilience index into four dimensions: recovery adaptability, innovation reorganization ability, shock resistance ability, and transformation development ability, and constructs a group regression model. The regression results are presented in Table 13. The empirical results show that the impact of AI on supply chain resilience exhibits significant dimensional heterogeneity. Particularly, the driving effects of AI were statistically significant at the 1% level in terms of recovery adaptability and innovation reorganization ability, while its impact on shock resistance and transformation development ability did not reach statistical significance. This suggests that AI demonstrates a “selective enhancement” characteristic when improving supply chain resilience, focusing particularly on improving dynamic adaptability and innovation reconstruction ability when responding to external changes. On one hand, in the dimensions of recovery adaptability and innovation reorganization ability, AI significantly enhances the response capacity and innovation speed of supply chains in rapidly changing environments through big data analysis and intelligent decision support systems, facilitating resource allocation optimization and organizational efficiency improvement. On the other hand, shock resistance ability is more reliant on the stability of the supply chain’s infrastructure and physical assets. While AI can improve efficiency, it cannot compensate for the limitations imposed by physical rigidity in the short term. Transformation development ability is influenced by the institutional environment and the complementarity between technology and institutions. Despite AI driving technological innovation, it still faces challenges such as delayed institutional changes and mismatched technology-institution interactions, resulting in its impact on this dimension not being significantly pronounced.

Heterogeneity of industrial chain export dependence in the industrial chain

This study, based on the heterogeneity hypothesis of global value chain (GVC) embedding, divides the manufacturing supply chain into outward-oriented (high GVC integration) and inward-oriented (low GVC integration) groups based on export dependence median and performs group regression. The regression results are shown in Table 13. The empirical results indicate that the regression coefficients for AI in both types of supply chains pass the significance test, but there is a significant coefficient difference, with AI having a greater impact on the outward-oriented supply chains. Outward-oriented supply chains, deeply integrated into the global production network, are strongly influenced by international market demand, technological exchange, and capital flow. AI accelerates the spillover of knowledge and the cross-national diffusion of innovation, effectively enhancing production efficiency and global competitiveness. At the same time, outward-oriented supply chains face numerous complex challenges from the international market, such as technological barriers, exchange rate fluctuations, and trade policy uncertainties. AI, through intelligent decision support and optimization of resource allocation, helps outward-oriented supply chains better cope with these external shocks, improving their adaptability and resilience. In contrast, inward-oriented supply chains, which mainly rely on the domestic market, are constrained by a relatively closed technological diffusion environment and market size. Their technological innovation and resource allocation optimization efficiency are relatively low. The application of AI in inward-oriented supply chains, constrained by the scale effects of the domestic market and technological path dependency, while capable of improving production efficiency to some extent, has a more limited empowering effect.

Conclusions and recommendations

Research conclusions

The panel data utilized by this research was obtained from 30 Chinese province areas between 2011 and 2022. It systematically analyzes the impact pathways and underlying mechanisms of AI on Mir using econometric methods, including fixed effects models, mechanism effect models, and threshold effect models. The key findings are as follows. Firstly, AI can directly improve Mir. Secondly, AI can indirectly enhance Mir through innovation-driven effects and digital empowerment effects. Thirdly, the consequences of AI on Mir exhibits non-linear characteristics under different levels of AI development and digital infrastructure. Specifically, when AI surpasses a specific limit, its positive impact on Mir increases; similarly, when the level of digital infrastructure exceeds a certain threshold, the impact of AI on Mir gradually strengthens. Fourthly, the enhancement repercussions of AI on Mir varies across dimensions of industrial resilience, within the supply chain itself, and in relation to the export dependence of the supply chain.

Policy recommendations

Regarding the situation of a globally uncertain economy and deep industrial restructuring, improving industrial supply chains’ security and resilience has emerged as a major concern for both governments and academic institutions. This not only involves proactively mitigating external shocks and reducing systemic risks during the restructuring of global trade patterns but also directly determines a country’s strategic positioning and long-term competitive advantage during the new international rivalry and employment divisions. Being a key force behind the recent manufacturing revolution and technological advancements, AI is accelerating the reshaping of global industrial competition through its profound impact on technological innovation, digital empowerment, and collaborative efficiency in supply chains, thereby providing strategic support for national economic security and supply chain resilience enhancement.

Firstly, implement a comprehensive strategy for the integration of AI across all production factors. To alleviate the financial constraints faced by small and medium-sized enterprises (SMEs) in adopting new technologies, a combination of fiscal and tax incentives—such as accelerated depreciation, innovation subsidies, and dedicated bonds—can be introduced to lower the marginal cost of technology diffusion. This will help enhance the resilience of industrial chains and achieve systemic improvements through the accumulation of technological capacity.

Secondly, restructure the innovation ecosystem around data as a key production factor. Efforts should focus on improving the efficiency of data utilization, establishing robust mechanisms for data ownership, trading, and cross-border flow, and building industrial knowledge graphs and digital twin systems to optimize resource allocation. By promoting the unification of technical standards and strengthening digital governance capabilities, a synergistic innovation system can be cultivated, facilitating the transformation of industrial chains toward more technology-intensive paradigms.

Thirdly, optimize the spatial distribution of new digital infrastructure. Using spatial econometric models to identify the marginal return differences across regions, a regionally differentiated investment strategy should be implemented. Priority should be given to building industrial internet platforms that align closely with local industrial needs to reduce the digital divide between regions. Furthermore, cross-regional computing power sharing mechanisms and digital skill training systems should be established to improve the technological absorptive capacity of less-developed areas and ensure widespread sharing of the digital dividend.

Fourthly, develop a differentiated policy response mechanism tailored to specific industries and regions. Based on the industrial chain’s position within the global value chain and regional factor endowments, dynamic thresholds for policy intervention should be designed. For inward-oriented industrial chains, focus should be placed on overcoming organizational inertia and enhancing innovation capabilities. For outward-oriented chains, efforts should be directed at aligning with international digital trade rules to reduce institutional transaction costs. By establishing a dynamic mapping between industrial chain resilience evaluation metrics and policy instruments, a shift from “one-size-fits-all” approaches to precision governance can be achieved, ultimately fostering a synergistic effect between technological empowerment and institutional innovation.

Restrictive discussion

This study primarily investigates the relationship between AI and Mir. However, it has certain limitations. First, the analysis is focused on China, and future research should expand the scope to an international context for broader insights. Second, due to data constraints, this study relies on provincial-level data. Future research will incorporate prefecture-level data to enhance the robustness of the findings and provide a more granular perspective for related studies.

CRediT authorship contribution statement

Qiujie He: Writing-review & editing, Conceptualisation, Project administration. Xiangling He: Writing-original draft, Formal analysis, Data curation, Investigation, Writing-review & editing, Software. Guoqing Chen: Software, Resources, Supervision. Piyapatr Busababodhin: Validation, Methodology. Wenjing Li: Obtaining funds, Visualisation.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

This article was completed with the support of the Guangxi Philosophy and Social Science Program “Study on the Performance Enhancement of Ethnic InterembeddedCommunity Governance in Guangxi under the Consciousness of Building Chinese National Community (24SHC002)”; Humanities and SocialScience Fund of Ministry of Education of China “Research on the Practice Mode, Influencing Factors and Path of Party Building Leading Urban CommunityGovernance Community (23YJC840032)”;"Source City Development Research Center Project Research on the Mechanism and Path of Improving the FineGovernance Level of Sichuan Resource based Cities Driven by Digitalization (ZYZX-YB-2404)";"Tuojiang River Basin High Quality DevelopmentResearch Center Project (TJGZL2025-01,TJGZL2023-12)".

Data availability

All data generated or analyzed during this study are included in this published article (and its supplementary information files).

Declarations

Competing interests

The authors declare no competing interests.