The knowledge economy necessitates robust scientific infrastructures to handle increasingly complex research challenges. Large-scale research infrastructures (LSRIs), such as China's National Supercomputing Centers (NSCs), are crucial for advancing science and technology, and contributing to societal well-being. While existing research has explored the definition, types, and distribution of LSRIs, and their theoretical scientific effects, systematic evaluations based on causal inference frameworks remain limited, particularly concerning their regional-level impact on innovation. This study addresses this gap by examining the impact and mechanisms of China's NSCs on knowledge innovation (KI). The NSCs provide a suitable case study due to their recent establishment in various Chinese cities (not just major metropolitan areas), allowing for a quasi-natural experiment to assess their causal effects on local development. The study also investigates NSCs' role as an extension of digital infrastructure and a practice of place-based innovation policy, examining their contribution to regional knowledge production in the context of China's unique institutional and development setting. The study aims to provide empirical evidence for the causal relationship between LSRIs and innovation performance, map scientific effects to tangible and intangible innovation resources based on the resource-based view, and classify the impact mechanisms into basic, network, and technology effects.

Existing literature explores the definition, types, and distribution of LSRIs, and their theoretical scientific effects (Michalowski, 2014; Qiao et al., 2016; D'ippolito and Rüling, 2019). Studies have categorized LSRIs' roles across scientific, technological, economic, educational, and social aspects (Marcelli, 2014; Caliari et al., 2020; Beck and Charitos, 2021), examining scientific achievements, impacts of construction and operation, personnel training, scientific cooperation, technological innovation, and education (Michalowski, 2014). However, systematic evaluations based on causal inference frameworks and identifying LSRIs' roles in regional-level innovation growth are limited (Bollen et al., 2011; Caliari et al., 2020; Beck and Charitos, 2021). This study draws upon the resource-based view, integrating it with social network theory and innovation efficiency research to create a framework that links LSRIs' scientific effects to regional innovation resources. This builds upon prior work that analyzes the impacts of digital infrastructure on economic growth, urban innovation, and social development (Cardona et al., 2013; Balcerzak and Bernard, 2017; Zhou et al., 2021; Zhang et al., 2022; Tang and Zhao, 2023), extending the analysis to the specific role of computing infrastructure (like NSCs) in scientific knowledge production. The study also contributes to the literature on place-based innovation policy, evaluating the impact of NSCs as a geographically-targeted intervention to foster regional innovation.

This study utilizes a quasi-natural experiment approach, leveraging the staggered rollout of NSCs across various Chinese cities since 2009. The dependent variable is per capita number of SCI-indexed publications, reflecting knowledge innovation (KI). The independent variable is a dummy variable indicating NSC presence. Mechanism variables capture the basic effect (R&D expenditure, S&T talent, digital infrastructure index), network effect (degree and closeness centrality in a city-level scientific research network), and technology effect (innovation efficiency measured using stochastic frontier analysis (SFA)). Control variables include economic development, industrial structure, population growth, financial sector development, human capital, and transport volume. Panel data from 283 Chinese cities (2000-2020) are used. The main estimation method is a spatial Durbin model (SDM), accounting for spatial autocorrelation in KI. A time-varying difference-in-differences (DID) approach is employed to account for temporal heterogeneity. Mechanism analysis uses Alesina and Zhuravskaya's (2011) method, and policy spillover is examined using spatial weight matrices based on geographical, cooperation, and digitization proximity. Innovation convergence is assessed using a β-convergence model. Robustness checks include OLS and spatial econometric models, instrumental variables (IV) estimation, propensity score matching-DID (PSM-DID), placebo tests, and alternative model specifications (Tobit and negative binomial models). Heterogeneity analysis involves grouping and moderating effect tests, and synthetic difference-in-differences (SDID).

The spatial Durbin model results show that NSC construction significantly and positively impacts knowledge innovation in both the treated cities and their surrounding areas. The estimated coefficients suggest that NSC construction substantially increases KI, emphasizing the importance of accounting for policy spillover effects. Mechanism analysis confirms the significant positive contribution of the basic effect (increased R&D expenditure, S&T talent, and digital infrastructure). The network effect is also significant, as NSC presence enhances urban centrality in the national scientific research network. The technology effect demonstrates that NSC improves innovation efficiency, both computationally and allocatively. The policy spillover is confirmed through geographic, cooperation, and digitization proximity, although the impact on regional innovation convergence is not significant. Heterogeneity tests reveal a stronger positive effect of NSC in application innovation-oriented cities like Shenzhen, compared to cities focused on basic research (Mathematics). Robustness checks across different model specifications, instrumental variable approaches (addressing endogeneity), placebo tests, and PSM-SDID confirm the positive and robust impact of NSC on KI.

The findings confirm the positive impact of large-scale research infrastructure, specifically NSCs, on regional knowledge innovation in China. The results support the conceptual framework linking LSRIs' scientific effects to innovation resources. The three identified mechanisms—basic effect, network effect, and technology effect—provide a comprehensive understanding of how NSCs stimulate KI. The spatial spillover effects highlight the importance of considering the geographic and digital interconnectedness of regions. The limited impact on regional convergence suggests potential avenues for policy refinement. The observed heterogeneity underlines the need to consider local knowledge orientation and innovation ecosystems when planning and implementing LSRIs. These findings have implications for understanding the impact of LSRIs globally, the role of digital infrastructure in promoting innovation, and the effectiveness of place-based innovation policies. The study's results provide empirical evidence to support policy decisions regarding investment in LSRIs and the design of targeted strategies for fostering regional innovation.

This study demonstrates the positive but limited impact of NSCs on regional knowledge innovation in China. The findings highlight the importance of considering the basic, network, and technology effects of LSRIs, as well as the role of geographical, cooperation, and digitization proximity in knowledge spillover. While NSCs promote local and regional innovation, their impact on innovation convergence remains limited, suggesting a need for improved inter-regional coordination and more comprehensive infrastructural development. Future research could explore the long-term impacts of NSCs, investigate the role of specific disciplines in benefiting from supercomputing resources, and examine the effectiveness of NSCs in promoting innovation in less developed regions.

This study's limitations include potential omitted variable bias, despite the use of control variables and instrumental variables. The focus on NSCs in China limits the generalizability of the findings to other countries with different institutional contexts. The chosen measure of knowledge innovation (SCI publications) may not fully capture all aspects of KI. Finally, the data collection period may not fully reflect the long-term impacts of NSCs.