Impact of Intelligent Transformation on Industrial Carbon Emission Efficiency and Its Spatial Spillover Effect: Evidence from 284 Chinese Cities

Abstract

This study explores the impact of industrial intelligent transformation on industrial carbon emission efficiency and its spatial spillover effect, which is closely related to industrial sustainability. Based on panel data of 284 cities in China from 2011 to 2023, we find that intelligent transformation significantly improves urban industrial carbon emission efficiency, and reducing energy consumption intensity and promoting green technological innovation are two critical mediating channels. Moreover, both marketization level and environmental regulation stringency strengthen the promoting role of intelligent transformation on industrial carbon emission efficiency. Heterogeneity analysis demonstrates that the promotional effect of intelligent transformation on industrial carbon emission efficiency is strongest in Eastern China, followed by Central China, and weakest in Western China. In addition, this effect is significant in non-resource-based cities but insignificant in resource-based cities. Furthermore, intelligent transformation exerts a negative “competitive spillover effect” on industrial carbon emission efficiency of geographically adjacent cities, while generating a positive “demonstration spillover effect” on cities with similar economic development levels.

1. Introduction

Global climate change is a shared challenge for humanity’s sustainable development, with carbon emission reduction and climate governance now a core international consensus. As the world’s largest developing country, China has explicitly articulated its “dual carbon” strategic goals and is committed to the comprehensive green transition of economic and social development, which is a crucial practice to promote global sustainability [1,2]. As a core pillar of China’s economy, the industrial sector accounts for approximately 65% of the country’s total energy consumption and over 70% of total carbon emissions, thus emerging as a key focus for carbon reduction [3,4]. Improving the carbon emission efficiency of urban industry is directly tied to progress in achieving China’s “dual carbon” goals and promoting high-quality development. However, traditional industry’s path dependence on high input, high consumption, and high emissions remains unresolved, and the tension between industrial growth and carbon emission constraints has become increasingly prominent [5,6]. Addressing this dilemma through technological innovation has thus become an urgent and critical task.

With the rapid advancement of digital technologies and artificial intelligence (AI), industrial intelligent transformation has become a core trend in global industrial upgrading [7], thereby injecting new momentum into the green and low-carbon development of industry. Centered on intelligent technologies and supported by smart equipment and systems, industrial intelligent transformation restructures production processes. It also optimizes factor allocation and innovates governance models, driving industrial production to shift from “scale priority” to “efficiency priority” and offering new solutions to balance industrial growth with carbon emissions. Existing studies have preliminarily explored the relationship between artificial intelligence and carbon emissions [8,9,10,11], which confirms the potential carbon reduction effects of digital technologies. Other studies have focused on the influencing factors of carbon emission efficiency [12,13,14], laying an important foundation for research on green transformation. However, cities do not exist in isolation amid increasingly close regional economic ties [8,15], and carbon emissions are spatially correlated across cities. Meanwhile, technological diffusion and industrial synergy induced by industrial intelligent transformation often go beyond regional boundaries, producing cross-regional spillover effects on the industrial carbon emission efficiency of other cities. Although existing studies have achieved progress on the relationship between artificial intelligence and carbon emissions, they still exhibit clear shortcomings in several key aspects, including a lack of targeted analysis on the industrial sector, limited exploration of internal mechanisms, inadequate identification of spatial effects, and insufficient examination of the moderating role of institutional factors. Thus, they cannot sufficiently offer precise theoretical and practical support for industrial low-carbon transformation in the intelligent era.

Building on this, this study employs a suite of methods including the two-way fixed effects model, multi-period DID, IV-2SLS, System Generalized Method of Moments (SYSGMM), and the spatial Durbin model (SDM) to systematically examine the impact of industrial intelligent transformation on urban industrial carbon emission efficiency and its spatial spillover effects. As defined in this paper, industrial carbon emission efficiency is a comprehensive efficiency indicator measured by the super-efficiency SBM model with undesirable outputs. Under the input constraints of capital, labor, energy, and other factors, it reflects the overall operational efficiency of urban industrial production that pursues both economic output and carbon emission reduction, taking industrial output as the desirable output and industrial carbon emissions as the undesirable output. The spatial spillover effect examined in this paper refers to the cross-regional transmission and linkage effects exerted on a city’s industrial carbon emission efficiency by relevant factors in other cities, which captures the interconnection and interaction of low-carbon development among cities. Drawing on panel data from 284 Chinese cities from 2011 to 2023, the analysis focuses on the following key questions: (1) Does industrial intelligent transformation exert a significant impact on urban industrial carbon emission efficiency? (2) What are the underlying mechanisms through which industrial intelligent transformation affects urban industrial carbon emission efficiency? (3) Does this impact exhibit spatial spillover effects, and do these effects differ between geographically and economically proximate cities? (4) Do marketization level and environmental regulation stringency play moderating roles in the relationship between the two? (5) Are there heterogeneous impacts across cities in different regions? Answering these questions deepens theoretical understanding of the link between intelligent transformation and industrial carbon emission efficiency, while providing targeted guidance for advancing industrial intelligent transformation and regional coordinated low-carbon development. It also offers a new theoretical perspective and empirical insights for interdisciplinary research on global industrial low-carbon transformation, as well as a useful reference for countries at different development stages to formulate differentiated intelligent carbon reduction strategies.

This study contributes to the existing literature in five key respects. In contrast to most existing studies that explore the impact of intelligence on carbon emissions from an aggregate perspective, it focuses on the high-carbon industrial sector and conducts an in-depth analysis of how intelligent transformation impacts urban industrial carbon emission efficiency. This expands the research context and segmented dimensions in the field of industrial sustainability, making conclusions more industry-specific and practically relevant. Drawing on theories including technological innovation, ecological modernization, and factor reallocation, this paper further unpacks the enabling mechanisms through which intelligent transformation enhances industrial carbon emission efficiency, and identifies the dual mediating paths of reducing energy consumption intensity and promoting green technological innovation. This enriches the theoretical framework of intelligent low-carbon transition. Beyond direct and mediating impacts, this paper explores the spatial spillover effects of intelligent transformation and differentiates between spillover patterns under geographical proximity and economic proximity, which provides targeted implications for regional coordinated carbon abatement. Additionally, it examines the moderating roles of marketization level and environmental regulation stringency, emphasizing the importance of the institutional environment as a crucial external factor and identifying the institutional arrangements that enable intelligent transformation to boost industrial carbon reduction. Finally, heterogeneous analysis is conducted across regions, which delivers both theoretical insights and practical evidence to support policymakers in designing differentiated and targeted strategies.

This paper adopts a structured analytical framework to achieve its research objectives. After theoretically analyzing the impact and mechanisms of industrial intelligent transformation on industrial carbon emission efficiency and proposing relevant hypotheses, we present the research design and empirical results, including baseline regressions, influence mechanisms, moderating effects, and heterogeneous analysis. We further examine the spatial spillover effects. Finally, the paper concludes with key findings and targeted policy suggestions.

2. Theoretical Analysis and Research Hypotheses

The core of industrial carbon emission efficiency lies in achieving the coordinated optimization of industrial output and carbon emissions, which essentially reflects the dynamic adaptation between production factor allocation efficiency and environmental regulation constraints. The positive enabling effect of industrial intelligent transformation on industrial carbon emission efficiency can be supported by three theoretical perspectives: technological innovation theory, ecological modernization theory, and factor restructuring theory [12,16,17], and is mainly realized through two core pathways: reducing energy consumption intensity and promoting green technological innovation.

From the viewpoint of technological innovation theory, as a general-purpose technological change, industrial intelligent transformation relies on technologies such as intelligent sensing, digital twins, and industrial robots to achieve precise control of production processes, replacement of manual operations, and refined energy management. By the real-time capture of key information such as equipment energy consumption and process parameters, it improves the energy utilization efficiency and reduces carbon emissions at the micro level [5,7,18]. Meanwhile, it shifts industrial production from large-scale mass production to customized flexible production, realizing the precise matching between production scale and environmental load [13,14]. According to ecological modernization theory, intelligent transformation breaks the dual opposition between economic growth and environmental governance in traditional industries, and constructs a positive cycle where technological progress, environmental performance, and economic growth promote each other [1,19]. Digital and intelligent technologies enable real-time feedback of interactive data between production and environmental systems, helping enterprises balance production efficiency and environmental costs [9,20]. Through real-time monitoring and dynamic adjustment of industrial energy use [1,4], as well as raising the share of clean energy, the carbon emissions per unit of output can be directly reduced [19,21], thereby further improving industrial carbon emission efficiency. Under the analytical framework of factor restructuring theory, the integration of data factors with traditional factors breaks the temporal and spatial constraints on factor allocation and improves the accuracy of factor utilization, thereby enhancing carbon emission efficiency [16,22]. Through in-depth data mining and analysis, intelligent technologies identify inefficient production links, optimize factor allocation ratios, and reduce energy waste and carbon emissions caused by factor misallocation [8,23]. In addition, the integration of data and technological factors accelerates the iteration of production technologies, promotes the transformation of high-energy-consuming technologies to low-carbon and energy-saving ones [17], and facilitates cross-enterprise and cross-regional factor flow to achieve factor complementarity and resource sharing, thereby reducing regional industrial carbon emission redundancy and improving overall industrial carbon emission efficiency.

In terms of mechanism analysis, industrial intelligent transformation promotes the improvement of industrial carbon emission efficiency mainly through two core pathways: reducing energy consumption intensity and promoting green technological innovation. Firstly, energy economics theory identifies technological progress as the core driving force of declining energy consumption intensity [4,20]. Industrial intelligent transformation can reduce energy consumption intensity through three dimensions: technological substitution, management optimization, and structural upgrading, thereby improving industrial carbon emission efficiency [24]. Specifically, technological substitution involves replacing energy-intensive traditional equipment with energy-saving intelligent ones, real-time energy monitoring via intelligent technologies [18], and process optimization through technologies such as waste heat recovery [5]. Management optimization relies on digital technologies for full-process energy data integration, AI for identifying energy waste nodes [1], and industrial internet platforms for rapid energy deployment [22]. Structural upgrading accelerates industrial green restructuring, reduces the overall industrial energy consumption intensity by transforming high-energy-consuming industries, and promotes the agglomeration of low-energy-consuming intelligent manufacturing industries [14]. Secondly, green technological innovation, as a technological R&D and transformation activity aimed at pollution reduction and energy conservation, constitutes the core driver for improving industrial carbon emission efficiency [25]. During the process of industrial intelligent transformation, digital and intelligent technologies provide comprehensive support for green innovation. In terms of innovation factor allocation, digital technologies break the temporal and spatial barriers of R&D resources, while intelligent technologies leverage industrial internet platforms to facilitate industry-university-research collaboration to break through core technologies [22]. In terms of reducing innovation costs, technologies such as digital twins restructure the R&D model, reducing experimental consumption and shortening the R&D cycle [17]. In terms of expanding innovation scenarios, intelligent transformation fosters new application scenarios, providing practical carriers for green technological innovation [26,27]; Wang et al., 2025. Ultimately, the implementation and application of green technological innovation drive industrial low-carbon transformation and further enhance industrial carbon emission efficiency [28].

In summary, industrial intelligent transformation is supported by three theories and significantly improves industrial carbon emission efficiency through two core mechanisms: reducing energy consumption intensity and promoting green technological innovation. Based on this, the following research hypothesis is proposed:

Hypothesis 1.

Industrial intelligent transformation exerts a significantly positive impact on industrial carbon emission efficiency.

In addition, considering the interactive characteristics among regions, industrial intelligent transformation may exert heterogeneous spatial spillover effects on industrial carbon emission efficiency from different perspectives of spatial proximity. Specifically, we focus on two scenarios: geographical proximity and economic proximity. From the perspective of geographical proximity, we argue that intelligent transformation generates a negative “competitive spillover effect”, meaning that intelligent transformation in adjacent cities mutually inhibit local industrial carbon emission efficiency. The core reason lies in three types of competition: First, competition for scarce factors. Geographically adjacent cities are located within the same factor market circle [4,14]. Local intelligent transformation will seize these scarce resources, such as low-carbon talents and intelligent equipment, through price premiums and policy support, thereby inhibiting the emission reduction efforts of surrounding cities. Second, homogeneous industrial competition. Cities with geographical proximity tend to have similar industrial structures. Competition for similar low-emission industries compresses the space for industrial upgrading in surrounding cities [20]. Third, competition for policy dividends. The local supportive policies for intelligent transformation will trigger imitation in surrounding cities, weakening the incentive for transformation and even leading to a “race to the bottom” [6]. In contrast, from the perspective of economic proximity, intelligent transformation brings a positive “demonstration spillover effect”, i.e., cities with close economic links mutually promote industrial carbon emission efficiency. The logic lies in a three-dimensional demonstration and diffusion mechanism: First, adaptive learning of transformation experience. Cities with similar economic development levels face homogeneous pain points in low-carbon transformation [17,29], so the local experience in low-carbon technologies and government-enterprise collaboration can be directly referenced by surrounding cities, reducing their trial-and-error costs. Second, benchmarking of factor allocation. Cities with similar economic levels have comparable factor endowments [13,19], and the local factor allocation model can serve as a reference for surrounding cities to avoid the risk of factor misallocation. Third, policy coordination. Owing to their shared policy objectives [10,21], local policies serve as a reference for other cities, facilitating the formation of regional policy synergy [3,30]. Based on the above theoretical analysis, we propose the following hypothesis:

Hypothesis 2.

The impact of industrial intelligent transformation on industrial carbon emission efficiency exhibits a spatial spillover effect, and this effect is heterogeneous in direction across different perspectives of spatial proximity.

3. Methodology

3.1. Variable Selection

First, drawing on existing studies [12,29], we employ the super-efficiency Slack-Based Measure (SBM) model with undesirable outputs to quantify industrial carbon emission efficiency (ICEF) across 284 Chinese cities. As an extended non-radial and non-oriented efficiency evaluation model built upon the traditional SBM framework, the super-efficiency SBM model effectively addresses the input-output slack problem. Its super-efficiency feature enables efficiency values of efficient decision-making units (DMUs) to exceed 1, facilitating more accurate differentiation of efficiency differences among efficient units. In selecting input and output indicators, we designate urban industrial carbon emissions as the undesirable output, gross urban industrial output value as the desirable output, and capital, labor, and energy as the core input factors.

Assuming the system comprises n DMUs, each with $m$ input factors, $l_1$ desirable outputs, and $l_2$ undesirable outputs. The matrices can be defined as follows:

$$\left\{\begin{array}{c}x=\left[x_1,\dots ,x_n\right]\in R_{m\times n}\\ y^d=\left[y_1^d,\dots ,y_n^d\right]\in R_{l_1\times n}\\ y^u=\left[y_1^u,\dots ,y_n^u\right]\in R_{l_2\times n}\end{array}\right.$$

(1)

Therefore, the model can be expressed as:

$${\rho }_{kt}=\min \rho ={\displaystyle \frac{1-\frac{1}{m}{\displaystyle {\sum }_{i=1}^m\frac{w_i^{-}}{x_{ik}}}}{1+\frac{1}{l_1+l_2}\times \left({\displaystyle {\sum }_{s=1}^{r_1}{w}_s^d/y_{sk}^d+{\displaystyle {\sum }_{q=1}^{r_2}{w}_q^u/y_{qk}^u}}\right)}}$$

(2)

The constraints are as follows:

$$s.t.\left\{\begin{array}{c}{x}_{ik}={\displaystyle {\sum }_{j=1}^n{x}_{ij}{\lambda }_j+w_i^{-}}\\ y_{sk}^d={\displaystyle {\sum }_{j=1}^n{y}_{sj}^d{\lambda }_j+w_s^d}\\ y_{qk}^u={\displaystyle {\sum }_{j=1}^n{y}_{qj}^u{\lambda }_j+w_q^u}\end{array}\right.$$

(3)

Herein, $w_i^{-}, w_s^d, w_q^u, {\lambda }_j\ge 0$; $j=1\cdots n$; $i=1\cdots m$; $s=1\cdots r_1$; $q=1\cdots r_2$. $x_{ik},y_{sk}^d,y_{qk}^u$, and ${\lambda }_j$ denote the i-th input, s-th desirable output, q-th undesirable output, and j-th DMU linear combination coefficients for the k-th DMU, respectively. $w_i^{-},w_s^d$, and $w_q^u$ denote slack variables, stemming, respectively, from the i-th input, s-th desirable output, and q-th undesirable output. ${\rho }_{kt}$ denotes the efficiency value of the k-th DMU. The k-th DMU is SBM-efficient if and only if ${\rho }_{kt}$ = 1. Therefore, the super-efficient SBM model can be formulated as:

$${\theta }_{kt}=\min \theta ={\displaystyle \frac{\frac{1}{m}{\displaystyle {\sum }_{i=1}^m\frac{\overline{x}}{x_{ik}}}}{\frac{1}{l_1+l_2}\times \left({\displaystyle {\sum }_{s=1}^{r_1}{\overline{y}}^d/y_{sk}^d+{\displaystyle {\sum }_{q=1}^{r_2}{\overline{y}}^u/y_{qk}^u}}\right)}}$$

(4)

The constraints are as follows:

$$s.t.\left\{\begin{array}{c}\overline{x}\ge {\displaystyle {\sum }_{j=1,j\ne k}^n{x}_{ij}{\lambda }_j}\\ {\overline{y}}^d\le {\displaystyle {\sum }_{j=1,j\ne k}^n{y}_{sj}^d{\lambda }_j}\\ {\overline{y}}^u\ge {\displaystyle {\sum }_{j=1,j\ne k}^n{y}_{qj}^u{\lambda }_j}\\ {\displaystyle {\sum }_{j=1,j\ne k}^n{\lambda }_j=1}\end{array}\right.$$

(5)

Herein, $\overline{x}\ge x_{ik}$; ${\overline{y}}^d\le y_{sk}^d$; ${\overline{y}}^u\ge y_{qk}^u$; ${\lambda }_j\ge 0$; $j=1\cdots n(j\ne k)$; $i=1\cdots m$; $s=1\cdots l_1$; $q=1\cdots l_2$. ${\theta }_{kt}$ represents the urban industrial carbon emission efficiency (ICEF).

Second, drawing on the research design of relevant literature [7], this study quantifies the extent of industrial intelligent transformation (INT) by employing city-level industrial robot installation density. The raw data are obtained from the industrial robot installation reports published by the International Federation of Robotics (IFR). As the IFR exclusively reports industry-level stocks of industrial robots, we compile city-level data through industry classification mapping and an instrumental variable (IV) strategy. In the first step, we map the IFR industry classification system precisely to China’s National Economic Industry Classification to ensure classificatory consistency and quantify industry-specific industrial robot installations in China. Next, building on the core logic of the Bartik instrumental variable approach, we combine city-level industry employment shares with the corresponding industry-level robot installation densities to rigorously compute city-level industrial robot installation density. The exact formula is presented below:

$$IN{T}_{it}={\displaystyle {\sum }_{j=1}^J\frac{em{p}_{ijt}}{em{p}_{it}}\ast \frac{robo{t}_{jt}}{em{p}_{jt}}}$$

(6)

Herein, i stands for city, t for year, and j for industry; robot refers to industrial robot installation density, and emp denotes employment number.

Third, we identify energy consumption intensity (ECI) and green technological innovation (GTI) as mediating variables. Among them, ECI is measured as the ratio of energy consumption to GDP, while GTI is measured by the natural logarithm of the count of authorized green patents in each city [25,30]. The moderating variables of this study primarily consist of regional marketization level (MAR) and environmental regulation stringency (ERS). Specifically, we adopt the National Economic Research Institute (NERI) Index of Marketization of China’s Provinces as the proxy for MAR [19]. Since this index is only calculated at the provincial level, we utilize the provincial-level marketization index corresponding to each city’s administrative jurisdiction as the measure of MAR. For city-level ERS, we identify 15 core keywords including environmental protection, pollution, green, low-carbon, ecology, and carbon emissions; we extract relevant expressions from the annual government work reports of each city through text mining, count the frequencies of the target keywords, and then compute their proportion relative to the total word count of the entire reports to measure ERS [15].

Finally, based on relevant literature [23,31,32], this study incorporates the following control variables: industrial structure (ISTR), operationalized as the ratio of value-added of the tertiary industry to that of the secondary industry; population density (POP), calculated as the ratio of total urban population to urban built-up area; digital infrastructure level (DIGI), proxied by mobile phone penetration rate; education level (EDU), gauged as the ratio of local education expenditure to urban GDP; urbanization level (URB), defined as the share of urban population in the total population; fiscal self-sufficiency rate (FIS), measured as the ratio of local fiscal revenue to local fiscal expenditure; financial development level (FIN), quantified as the ratio of total deposits and loans of local financial institutions to GDP; and economic development level (ECO), captured by the local GDP growth rate.

The definitions and calculation methods of the variables are presented in

Table 1. Variable definitions and calculation methods.

Variables	Definitions	Calculation Methods
ICEF	Industrial Carbon Emission Efficiency	Super-efficiency SBM model with undesirable outputs
INT	Industrial Intelligent Transformation	Industrial robot installation density
ISTR	Industrial Structure	Added value of the tertiary industry/Added value of the secondary industry
POP	Population Density	Natural logarithm of (Total urban population/Urban area)
DIGI	Digital Infrastructure Level	Mobile phone penetration rate
EDU	Education Level	Education expenditure/Urban GDP
URB	Urbanization Level	Proportion of urban population in total population
FIS	Fiscal Self-Sufficiency Rate	Fiscal revenue/Fiscal expenditure
FIN	Financial Development Level	Total deposits and loans of financial institutions/GDP
ECO	Economic Development Level	Urban GDP growth rate
ECI	Energy Consumption Intensity	Energy consumption/GDP
GTI	Green Technological Innovation	Natural logarithm of the number of authorized green patents
MAR	Marketization Level	Fan Gang’s marketization index
ERS	Environmental Regulation Stringency	Text mining of government work reports

.

3.2. Model Construction

To investigate the impact of industrial intelligent transformation on urban industrial carbon emission efficiency, this study first adopts the Ordinary Least Squares (OLS) estimation method for empirical analysis and constructs the following regression model:

$${ICEF}_{it}=\alpha +{\beta }_1{INT}_{it}+{\beta }_2{Controls}_{it}+{\epsilon }_{it}$$

(7)

In this specification, i denotes the city and t denotes the year. ICEF represents urban industrial carbon emission efficiency, INT denotes the level of industrial intelligent transformation, Controls represents a set of control variables, and ${\epsilon }_{it}$ is the random error term.

Furthermore, accounting for potential city-specific individual heterogeneity and time-specific trend effects, this study employs a static panel two-way fixed effects model controlling for both individual and time fixed effects for regression estimation:

$${ICEF}_{it}=\alpha +{\beta }_1{INT}_{it}+{\beta }_2{Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(8)

Among them, ${\mu }_i$ denotes the individual fixed effect and ${\theta }_t$ denotes the time fixed effect, with other variables consistent with those in Equation (7).

Meanwhile, this study specifies the following mediating effect models to examine the underlying transmission mechanism through which industrial intelligent transformation affects urban industrial carbon emission efficiency:

$${Medium}_{it}=\alpha +{\beta }_1{INT}_{it}+{\beta }_2{Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(9)

Herein, Medium denotes the mediating variable, including energy consumption intensity (ECI) and green technological innovation (GTI).

In addition, to further examine the moderating role of the external institutional environmental factors in the relationship between industrial intelligent transformation and urban industrial carbon emission efficiency, this study specifies the following moderating effect models for regression estimation:

$${ICFE}_{it}=\alpha +{\beta }_1{INT}_{it}+{\beta }_2{INT}_{it}\ast {MOD}_{it}+{\beta }_3{MOD}_{it}+\gamma {Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(10)

Herein, MOD denotes the moderating variables, namely regional marketization level (MAR) and environmental regulation stringency (ERS). INT∗MOD denotes the interaction term between industrial intelligent transformation (INT) and MOD.

Moreover, we employ several methods to perform endogeneity tests. The model for the multi-period difference-in-differences (Multi-period DID) is specified as follows:

$${ICFE}_{it}=\alpha +\beta {Treat}_{it}+\gamma {Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(11)

Herein, Treat is a group dummy variable that equals 1 for the treatment group and 0 for the control group.

The first-stage model of the instrumental variable two-stage least squares (IV-2SLS) is specified as follows:

$${INT}_{it}={\alpha }_0+{\alpha }_1{IV}_{it}+{\alpha }_2{Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(12)

Herein, IV denotes the instrumental variable.

The second-stage model is specified as follows:

$${ICEF}_{it}=\alpha +{\beta }_1{I\widehat{N}T}_{it}+{\beta }_2{Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(13)

Herein, $I\widehat{N}T$ represents the fitted value estimated from the first stage.

The two-step system generalized method of moments (SYS-GMM) model for dynamic panels is specified as follows:

$${ICFE}_{it}=\alpha +{\beta }_1{ICFE}_{it-1}+{\beta }_2{INT}_{it}+\gamma {Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(14)

Herein, ICEF_it−1 is the first-order lag of ICEF_it.

Finally, to examine the spatial spillover effect of industrial intelligent transformation on urban industrial carbon emission efficiency, this study employs the Spatial Durbin Model (SDM) with fixed effects to perform regression estimation. The specific model specification is as follows:

$${ICEF}_{it}=\rho {WICEF}_{it}+{\beta }_1{INT}_{it}+\lambda {WINT}_{it}+{\beta }_2{Control}_{it}+\gamma W{Control}_{it}+{\mu }_i+{\theta }_t+{\epsilon }_{it}$$

(15)

Among them, W represents the spatial weight matrix, and WICEF, WINT, and WControl, respectively, denote the spatial lag terms of the explained variable, the core explanatory variable, and the control variables.

3.3. Sample, Data, and Descriptive Statistics

To ensure the representativeness, coverage, and timeliness of the research sample, this study utilizes a panel dataset covering 284 prefecture-level and above cities in China for the period 2011–2023. Regarding data sources, industrial robot installation density data are obtained from the annual industrial robot installation reports released by the International Federation of Robotics (IFR). Other relevant data are mainly collected from the National Bureau of Statistics of China (NBSC), Wind Information Database, China Research Data Services Platform (CNRDS), and various official statistical yearbooks, namely China Statistical Yearbook, China Energy Statistical Yearbook, China Environmental Statistical Yearbook, China Urban Statistical Yearbook, and the statistical yearbooks and bulletins of individual cities. For partially undisclosed city-level data, scientific imputation and estimation are performed using corresponding provincial-level data, involving the matching of key weighted indicators (e.g., economic scale share, industrial structure similarity) and the adoption of methodological frameworks (e.g., the Bartik instrumental variable approach). In addition, for a small amount of missing data, gap-filling is implemented via the trend extrapolation method. Descriptive statistics for the relevant variables are presented in

Table 2. Descriptive statistics of variables.

Variable	Mean	S.D.	Min	Median	Max	N
ICEF	0.240	0.143	0.018	0.211	1.250	3692
INT	5.278	1.896	0.000	5.072	11.54	3692
ISTR	0.434	0.102	0.102	0.431	0.849	3692
POP	5.747	0.929	1.628	5.891	8.176	3692
DIGI	1.102	0.739	0.192	0.928	10.17	3692
EDU	0.035	0.017	0.002	0.030	0.149	3692
URB	0.400	0.215	0.075	0.341	1.000	3692
FIS	0.448	0.218	0.056	0.415	1.541	3692
FIN	2.635	1.256	0.588	2.342	21.30	3692
ECO	7.382	3.814	−20.63	7.600	22.50	3692
ECI	17.79	14.87	0.465	14.21	160.2	3692
GTI	4.652	1.608	0.693	4.585	9.667	3630
MAR	8.577	1.775	3.359	8.552	13.36	3692
ERS	0.287	0.115	0.000	0.278	1.000	3692

Note: Unit of measurement: INT is measured in robots per 1000 employees; POP is a logarithmic value and thus unitless; DIGI is measured in sets per capita; ISTR, EDU, URB, FIS, and FIN are ratios and unitless; ECO is measured in percent; ECI is measured in kg of standard coal per 10,000 yuan of GDP; ICEF, GTI, MAR and ERS are unitless.

. Notably, the sample cities exhibit significant heterogeneity across industrial carbon emission efficiency, industrial intelligent transformation level, industrial structure, population density, digital infrastructure level, education level, and urbanization level.

4. Results

4.1. Baseline Results

Table 3. Results of baseline regression.

	OLS		Fixed Effect Model
	(1)	(2)	(3)	(4)
Variables	ICEF	ICEF	ICEF	ICEF
INT	0.0349 ***	0.0403 ***	0.00822 ***	0.0233 ***
	(0.00155)	(0.00200)	(0.00214)	(0.00328)
ISTR		−0.337 ***		−0.332 ***
		(0.0357)		(0.0597)
POP		0.00558 **		0.193 ***
		(0.00274)		(0.0617)
DIGI		0.0656 ***		0.0158
		(0.00764)		(0.0125)
EDU		−0.286 **		−2.565 ***
		(0.129)		(0.584)
URB		−0.0161		−0.0264
		(0.0123)		(0.0358)
FIS		0.0561 ***		−0.0154
		(0.0148)		(0.0344)
FIN		−0.0195 ***		−0.0131 ***
		(0.00286)		(0.00432)
ECO		0.0043 ***		0.000704
		(0.00053)		(0.000533)
Constant	0.0559 ***	0.0803 ***	0.197 ***	−0.731 **
	(0.00755)	(0.0205)	(0.0113)	(0.359)
City FE	No	No	Yes	Yes
Year FE	No	No	Yes	Yes
N	3692	3692	3692	3692
R-squared	0.213	0.433	0.022	0.170

Note: *** and ** indicate significance at the 1% and 5% levels, respectively. Robust standard errors are in parentheses.

reports the results of the benchmark regression analysis. Columns (1) and (2) report the regression results using the OLS method, while Columns (3) and (4) report those from the individual-time two-way fixed effects model for static panel data. Columns (1) and (3) exclude control variables, whereas Columns (2) and (4) include them. The estimated regression coefficients of industrial intelligent transformation (INT) in Columns (1) to (4) are all significantly positive at the 1% significance level. This result indicates that industrial intelligent transformation significantly improves urban industrial carbon emission efficiency, thus providing empirical support for Hypothesis 1.

4.2. Influence Mechanism Analysis

As stated in the previous theoretical analysis section, this study regards reducing energy consumption intensity and advancing green technological innovation as two core channels through which industrial intelligent transformation enhances industrial carbon emission efficiency. This section conducts empirical verification on these mediating channels, and the findings are reported in

Table 4. Results of the influence mechanism test.

	(1)	(2)
Variables	ECI	GTI
INT	−1.484 ***	0.426 ***
	(0.351)	(0.0176)
ISTR	19.68 ***	1.487 ***
	(5.589)	(0.302)
POP	−19.80 ***	1.559 ***
	(5.633)	(0.320)
DIGI	−3.079 ***	0.232 ***
	(0.820)	(0.0714)
EDU	215.7 ***	0.784
	(46.34)	(2.170)
URB	2.514	0.157
	(3.129)	(0.235)
FIS	−14.40 ***	−0.325 *
	(2.848)	(0.173)
FIN	1.918 **	−0.0392 ***
	(0.783)	(0.0150)
ECO	0.0686	−0.0174 ***
	(0.0546)	(0.00294)
Constant	126.7 ***	−7.125 ***
	(32.24)	(1.841)
City FE	Yes	Yes
Year FE	Yes	Yes
N	3692	3630
R-squared	0.231	0.748

Note: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust standard errors are in parentheses.

. Columns (1) and (2) correspond to the results of the energy consumption intensity channel and the green technological innovation channel, respectively. The estimated regression coefficient of INT in Column (1) is significantly negative at the 1% significance level, indicating that industrial intelligent transformation significantly reduces urban industrial energy consumption intensity [1,20]. Meanwhile, the estimated regression coefficient of INT in Column (2) is significantly positive at the 1% significance level, demonstrating that industrial intelligent transformation significantly promotes urban green technological innovation [17,26]. These results are consistent with the theoretical analysis presented earlier.

4.3. Endogeneity Tests

We employ methods including the multi-period DID, IV-2SLS, and SYS-GMM to address endogeneity concerns [6,19]. First, building on the research design of quasi-natural experiments, we utilize the “National Smart City Pilot” policy as an exogenous shock to construct a multi-period DID model for mitigating endogeneity. The core rationale for selecting this policy as an instrumental variable lies in its strong exogeneity (pilot qualification acquisition is exogenous), high alignment with industrial intelligent transformation, and ability to generate an effective policy shock. We define a treatment variable (Treat): for cities listed as “National Smart City Pilots” during the sample period, we assign a value of 0 to these cities before the pilot year and 1 after the pilot year; all non-pilot cities are assigned a value of 0. The regression results of the multi-period DID model are reported in Column (1) of

Table 5. Results of endogeneity test.

	Multi-Period DID	IV-2SLS		SYS-GMM
	(1)	(2)	(3)	(4)
Variables	ICEF	INT	ICEF	ICEF
L.ICEF				0.865 ***
				(0.0625)
Treat	0.0130 **
	(0.00594)
IV		1.046 ***
		(0.00533)
INT			0.0252 ***	0.0146 ***
			(0.00239)	(0.00168)
ISTR	−0.294 ***	0.432 ***	−0.327 ***	−0.0559 *
	(0.0414)	(0.0781)	(0.0351)	(0.0296)
POP	0.232 ***	0.0394	0.179 ***	−0.00181
	(0.0350)	(0.0633)	(0.0395)	(0.00236)
DIGI	0.0239 *	0.000680	0.00608	0.00544
	(0.0132)	(0.0129)	(0.0148)	(0.00716)
EDU	−2.439 ***	−3.009 ***	−2.381 ***	−0.0712
	(0.379)	(0.597)	(0.397)	(0.123)
URB	−0.0285	0.145 **	−0.0434 **	−0.00125
	(0.0204)	(0.0683)	(0.0221)	(0.00853)
FIS	0.0120	0.0377	−0.0103	−0.0177
	(0.0210)	(0.0460)	(0.0213)	(0.0113)
FIN	−0.0145 ***	0.00869	−0.0133 ***	−0.00432 **
	(0.00395)	(0.00654)	(0.00395)	(0.00216)
ECO	−0.000718	0.00140	0.000227	0.00243 ***
	(0.000483)	(0.00118)	(0.000427)	(0.000305)
Constant	−0.861 ***	−0.319	0.000	−0.0102
	(0.209)	(0.374)	(0.000)	(0.0163)
City FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
N	3692	3408	3408	3408
R-squared	0.800	0.994	0.178
LM statistic		286.398
C-D Wald F		65000
AR2(p)				0.093
Hansen(p)				1.000

Note: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust standard errors are in parentheses. AR2(p) and Hansen(p) correspond to the p-values of the second-order serial correlation test and Hansen over-identification test, respectively.

. Second, we adopt the first-order lag term of industrial intelligent transformation (L.INT) as the instrumental variable (IV) for INT and employ the 2SLS method for estimation, and the results are reported in Columns (2) and (3) of Table 5. These two columns report the first-stage and second-stage estimation results, respectively. Finally, we employ the two-step system generalized method of moments (SYS-GMM) for dynamic panel estimation to address endogeneity, with the results reported in Column (4) of Table 5. The aforementioned results demonstrate that after addressing endogeneity concerns, the core conclusion that industrial intelligent transformation improves industrial carbon emission efficiency remains robust.

4.4. Robustness Tests

This study conducts robustness checks using methods including variable substitution, estimation method replacement, sample adjustment, and sample period truncation. First, we employ the natural logarithm of (the number of regional artificial intelligence patents + 1) (INT_pat) as the alternative proxy for INT to conduct regression estimation. Second, we employ the static panel random effects model as an alternative estimation approach for the regression analysis. Third, we exclude the four municipalities directly under the central government (Beijing, Shanghai, Tianjin, Chongqing) and perform the analysis with an adjusted sample. Finally, accounting for the potential impact of the COVID-19 pandemic, we exclude data from 2020 onwards and re-perform the regression estimation. The results are reported in

Table 6. Results of robustness tests.

	Alternative Variable	Alternative Method	Alternative Sample	Adjusted Period
	(1)	(2)	(3)	(4)
Variables	ICEF	ICEF	ICEF	ICEF
INT_pat	0.00888 ***
	(0.00326)
INT		0.0273 ***	0.0231 ***	0.0293 ***
		(0.00302)	(0.00328)	(0.00648)
ISTR	−0.170 ***	−0.353 ***	−0.325 ***	−0.483 ***
	(0.0544)	(0.0588)	(0.0598)	(0.0859)
POP	0.268 ***	0.0244 ***	0.170 ***	0.0984 **
	(0.0672)	(0.00902)	(0.0605)	(0.0467)
DIGI	0.0267 *	0.0268 ***	0.00749	0.0271 ***
	(0.0141)	(0.00799)	(0.0121)	(0.0100)
EDU	−3.653 ***	−1.910 ***	−2.470 ***	−1.648 ***
	(0.531)	(0.428)	(0.583)	(0.394)
URB	0.0451	−0.0287	−0.0278	0.00776
	(0.0343)	(0.0275)	(0.0348)	(0.0533)
FIS	−0.0365	0.0307	−0.00708	−0.0204
	(0.0355)	(0.0270)	(0.0342)	(0.0327)
FIN	−0.00415 *	−0.0133 ***	−0.0130 ***	−0.0128 **
	(0.00234)	(0.00381)	(0.00434)	(0.00502)
ECO	5.31 × 10−5	0.000961 *	0.000720	0.00238 ***
	(0.000571)	(0.000522)	(0.000523)	(0.000583)
Constant	−1.163 ***	0.171 ***	−0.602 *	−0.223
	(0.391)	(0.0631)	(0.351)	(0.269)
City FE	Yes	No	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
N	3692	3692	3640	2556
R-squared	0.127		0.170	0.166

Note: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust standard errors are in parentheses.

. It can be observed that the regression coefficients of INT_pat or INT across all columns are significantly positive at the 1% significance level, confirming the robustness of the core research findings.

4.5. Moderating Effects Analysis

The promoting effect of intelligent transformation on industrial carbon emission efficiency is constrained by urban external conditions [28,33]. As key external condition variables, marketization level and environmental regulation stringency may moderate the intensity of this promoting effect. Therefore, this section conducts a moderating effect analysis, and the regression results are reported in

Table 7. Results of the moderating effect analysis.

	(1) Marketization Level	(2) Environmental Regulation Stringency
Variables	ICEF	ICEF
INT	0.0134 ***	0.0236 ***
	(0.00338)	(0.00326)
INT∗MAR	0.00587 ***
	(0.000880)
INT∗ERS		0.0262 **
		(0.0115)
MAR/ERS	0.0106 ***	−0.0413 **
	(0.00389)	(0.0168)
ISTR	−0.258 ***	−0.321 ***
	(0.0557)	(0.0591)
POP	0.129 **	0.191 ***
	(0.0524)	(0.0609)
DIGI	0.0258 **	0.0157
	(0.0116)	(0.0125)
EDU	−2.615 ***	−2.527 ***
	(0.572)	(0.576)
URB	−0.0485	−0.0242
	(0.0329)	(0.0353)
FIS	0.0289	−0.00787
	(0.0322)	(0.0345)
FIN	−0.0143 ***	−0.0122 ***
	(0.00468)	(0.00408)
ECO	0.000284	0.000514
	(0.000467)	(0.000528)
Constant	−0.461	−0.719 **
	(0.304)	(0.353)
City FE	Yes	Yes
Year FE	Yes	Yes
N	3692	3692
R-squared	0.228	0.177

Note: *** and ** indicate significance at the 1% and 5% levels, respectively. Robust standard errors are in parentheses. To effectively avoid the potential multicollinearity problem between the interaction terms, the core explanatory variable, and the moderating variables, INT, MAR, and ERS in this table have all been centralized by subtracting their respective sample means. INT∗MAR and INT∗ERS are interaction terms constructed using these mean-centered variables.

. Column (1) reports the results of the moderating effect of urban marketization level. The coefficient of INT is significantly positive, which further validates the promoting effect of intelligent transformation on industrial carbon emission efficiency. The regression coefficient of the interaction term INT∗MAR is significantly positive, indicating that a higher marketization level further amplifies the positive impact of intelligent transformation on industrial carbon emission efficiency. Column (2) reports the results of the moderating effect of urban environmental regulation stringency. The regression coefficient of the interaction term INT∗ERS is significantly positive, indicating that stricter urban environmental regulation stringency further amplifies the positive promoting effect of intelligent transformation on industrial carbon emission efficiency.

4.6. Heterogeneity Analysis

Drawing on criteria including economic development level and geographical location, regions in China are usually divided into eastern, central, and western regions. Significant heterogeneity exists across different regions in terms of technological reserves, industrial structure, digital infrastructure, and factor agglomeration capacity, which directly influences the integration effect of intelligent technologies and industrial low-carbon production [3,31]. Based on the above analysis, this study stratifies the sample cities into three subgroups to conduct heterogeneity analysis. The regression results are reported in

Table 8. Results of heterogeneity analysis.

	(1) Eastern	(2) Central	(3) Western	(4) Resource-Based Cities	(5) Non-Resource-Based Cities
Variables	ICEF	ICEF	ICEF	ICEF	ICEF
INT	0.0300 ***	0.0182 ***	−0.00361	0.0114	0.0254 ***
	(0.00564)	(0.00338)	(0.0174)	(0.00723)	(0.00356)
ISTR	−0.232 **	−0.187 **	−0.328 ***	−0.259	−0.370 ***
	(0.0961)	(0.0927)	(0.117)	(0.161)	(0.0508)
POP	0.142	0.189 *	0.237 **	−0.0538	0.212 ***
	(0.102)	(0.114)	(0.105)	(0.0558)	(0.0715)
DIGI	0.0139	0.0123	0.0325	0.0465 **	0.0156
	(0.0168)	(0.0338)	(0.0222)	(0.0201)	(0.0139)
EDU	−3.757 ***	−2.315 ***	−2.400 *	−1.247	−2.626 ***
	(0.941)	(0.671)	(1.226)	(0.939)	(0.663)
URB	−0.0117	−0.116 *	−0.00310	−0.0724	−0.0123
	(0.0533)	(0.0639)	(0.116)	(0.0801)	(0.0394)
FIS	−0.0513	0.0650	0.110	0.124	−0.0462
	(0.0439)	(0.0620)	(0.0936)	(0.107)	(0.0303)
FIN	−0.0232 **	−0.00594 **	−0.0281 **	−0.0174	−0.0134 ***
	(0.00910)	(0.00243)	(0.0133)	(0.0161)	(0.00463)
ECO	0.000282	0.00109 ***	−0.00223	0.00133 **	0.000455
	(0.000950)	(0.000356)	(0.00157)	(0.000628)	(0.000690)
Constant	−0.496	−0.794	−0.729	0.534 *	−0.835 *
	(0.637)	(0.669)	(0.556)	(0.290)	(0.424)
City FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes
N	1300	1300	1092	624	3068
R-squared	0.238	0.172	0.219	0.155	0.196

Note: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust standard errors are in parentheses.

, where Columns (1)–(3) correspond to the regression results of cities in the eastern, central, and western regions, respectively. The coefficients of INT in Columns (1) and (2) are 0.0300 and 0.0182, respectively, and both are significantly positive at the 1% significance level, while the regression coefficient of INT in Column (3) is not statistically significant. Furthermore, we adopt Fisher’s Permutation test to examine whether statistically significant differences exist between the coefficients of INT in Columns (1) and (2). The corresponding p-value of the test result is 0.045, confirming a significant difference between these two coefficients. The above results indicate that the positive promoting effect of intelligent transformation on industrial carbon emission efficiency is the most significant in eastern cities, followed by central cities, and is negligible in western cities.

Based on the “National Sustainable Development Plan of Resource-Based Cities” issued by the State Council of China in 2013, this paper compiles a list of resource-based cities, divides all sample cities into two groups: resource-based cities and non-resource-based cities, and performs a heterogeneity analysis. The results are shown in Table 8, where Columns (4) and (5) correspond to the regression results of resource-based city samples and non-resource-based city samples, respectively. The regression coefficient of INT in Column (4) does not pass the significance test; that of INT in Column (5) is 0.0254, which is significant at the 1% level. This result indicates that the positive promoting effect of intelligent transformation on industrial carbon emission efficiency is significant in non-resource-based cities, but insignificant in resource-based cities.

5. Spatial Spillover Effect Analysis

Given the spatial correlation among cities, this study further adopts spatial econometric models to empirically test the spatial spillover effects of industrial intelligent transformation on industrial carbon emission efficiency. First, we employ Moran’s I index to test whether the explained variable has spatial autocorrelation.

Table 9. Moran’s I of urban industrial carbon emission efficiency.

Year	Moran’s I	Z-Statistics	p-Value
2011	0.199	5.226	0.000
2012	0.229	5.947	0.000
2013	0.188	4.871	0.000
2014	0.211	5.488	0.000
2015	0.255	6.563	0.000
2016	0.315	8.249	0.000
2017	0.220	5.692	0.000
2018	0.270	6.932	0.000
2019	0.298	7.658	0.000
2020	0.313	8.019	0.000
2021	0.309	7.888	0.000
2022	0.265	6.828	0.000
2023	0.285	7.296	0.000

Note: The above results are calculated based on the spatial adjacency matrix.

reports the Moran’s I indices of industrial carbon emission efficiency for sample cities during 2011–2023, along with their corresponding significance levels. All Moran’s I indices of industrial carbon emission efficiency during 2011–2023 are positive, ranging between 0.188 and 0.315. The corresponding Z-statistics range from 4.871 to 8.249, all of which are significant at the 1% level. This finding demonstrates that urban industrial carbon emission efficiency is not randomly distributed; instead, it exhibits significant positive spatial autocorrelation, thus justifying the use of spatial econometric methods.

To further intuitively illustrate the spatial distribution and correlation characteristics of industrial carbon emission efficiency, we plot the Moran’s I scatter plots of urban industrial carbon emission efficiency for selected years (2011, 2015, 2019, and 2023) in Figure 1. The scatter plots for other years exhibit similar distribution patterns. As illustrated in Figure 1, the scatter points in each year are mainly concentrated in the first and third quadrants, indicating that the number of cities with high-high (H-H) and low-low (L-L) agglomeration is significantly larger than that of cities with high-low (H-L) and low-high (L-H) agglomeration. The fitted regression line exhibits a distinctly positive slope, which clearly reflects the positive spatial agglomeration characteristics of urban industrial carbon emission efficiency.

Drawing on the results of spatial autocorrelation tests, Lagrange Multiplier (LM) tests, Likelihood Ratio (LR) tests, and Hausman tests, this study ultimately adopts the Spatial Durbin Model (SDM) for empirical analysis. We employ the two-way fixed-effects maximum likelihood estimator (MLE) for parameter estimation, and robust standard errors are clustered at the city level to mitigate heteroscedasticity and serial correlation.

Table 10. Results of spatial econometric analysis based on SDM.

	Adjacency Matrix		Geographic Distance Matrix		Economic Distance Matrix		Nested Matrix
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Variables	Main	Wx	Main	Wx	Main	Wx	Main	Wx
INT	0.0602 ***	−0.0194 ***	0.0608 ***	−0.0268	0.0546 ***	0.0187 ***	0.0557 ***	0.0132 **
	(0.00236)	(0.00406)	(0.00226)	(0.0314)	(0.00220)	(0.00581)	(0.00221)	(0.00573)
ISTR	−0.343 ***	0.190 ***	−0.350 ***	0.136	−0.296 ***	−0.304 ***	−0.281 ***	−0.170 **
	(0.0311)	(0.0523)	(0.0300)	(0.360)	(0.0285)	(0.0738)	(0.0291)	(0.0764)
POP	−0.00644 **	0.0135 ***	0.0149 ***	−0.164 ***	0.00555 **	−0.0184 ***	0.00618 **	−0.0143 ***
	(0.00327)	(0.00396)	(0.00323)	(0.0292)	(0.00235)	(0.00497)	(0.00241)	(0.00534)
DIGI	0.0493 ***	0.0196 **	0.0540 ***	0.0999 ***	0.0546 ***	−0.0105	0.0571 ***	0.00534
	(0.00494)	(0.00835)	(0.00402)	(0.0315)	(0.00323)	(0.00651)	(0.00327)	(0.00821)
EDU	−0.111	−0.869 ***	0.0450	−11.84 ***	−0.322 *	1.341 ***	−0.102	−0.521
	(0.164)	(0.252)	(0.162)	(2.013)	(0.167)	(0.430)	(0.197)	(0.416)
URB	0.00774	−0.0856 ***	−0.0200 *	−0.648 ***	−0.0386 ***	0.0920 ***	−0.0385 ***	0.0149
	(0.0116)	(0.0193)	(0.0115)	(0.122)	(0.0107)	(0.0287)	(0.0108)	(0.0289)
FIS	2.03 × 10−5	−0.0640 **	0.0149	−0.362 **	−0.0914 ***	0.143 ***	−0.0967 ***	0.0208
	(0.0180)	(0.0268)	(0.0173)	(0.164)	(0.0166)	(0.0359)	(0.0171)	(0.0371)
FIN	−0.0211 ***	0.0122 ***	−0.0244 ***	0.0440 **	−0.0210 ***	0.00356	−0.0223 ***	0.0192 ***
	(0.00226)	(0.00371)	(0.00224)	(0.0188)	(0.00199)	(0.00486)	(0.00204)	(0.00520)
ECO	5.75 × 10−5	0.000718	−0.000171	0.00719	0.000324	−8.95 × 10−5	0.000169	0.00166
	(0.000758)	(0.00112)	(0.000764)	(0.00554)	(0.00067)	(0.00180)	(0.000689)	(0.00181)
ρ	0.241 ***		0.664 ***		0.169 ***		0.180 ***
	(0.0217)		(0.0777)		(0.0283)		(0.0264)
σ2e	0.0105 ***		0.0106 ***		0.0106 ***		0.0106 ***
	(0.00025)		(0.00025)		(0.00025)		(0.00025)
N	3692	3692	3692	3692	3692	3692	3692	3692
R2	0.374	0.374	0.253	0.253	0.353	0.353	0.323	0.323

Note: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust standard errors are in parentheses; Column Main corresponds to the direct impact of each variable on local industrial carbon emission efficiency, and Column Wx represents the impact of the spatial lag term of the explanatory variable on the local dependent variable.

reports the regression results under different spatial weight matrices. Columns (1)–(2) correspond to the spatial geographic adjacency matrix, with weights set based on regional adjacency; Columns (3)–(4) represent the spatial geographic distance matrix, with weights inversely proportional to geographic distance; Columns (5)–(6) denote the spatial economic distance matrix, with weights inversely proportional to differences in economic development levels; Columns (7)–(8) represent the nested geographic-economic distance matrix, which comprehensively considers geographic proximity and economic relevance. Column Main reports the direct effect of each variable on local industrial carbon emission efficiency, while Column Wx captures the impact of the spatial lag term of the explanatory variable on the local dependent variable. The results demonstrate that the spatial autocorrelation coefficient (ρ) across all matrices is significantly positive at the 1% level, and the estimated residual variance (σ²ₑ) also passes the 1% significance test. This confirms that industrial carbon emission efficiency exhibits significant positive spatial autocorrelation, and the model has robust residual variance and a reasonable specification.

In Columns (1), (3), (5), and (7), the estimated coefficients of industrial intelligent transformation (INT) are all significantly positive at the 1% level. This indicates that after considering geographical or economic connections, local industrial intelligent transformation still exerts a significant positive effect on local industrial carbon emission efficiency. Regarding spatial spillover effects, Columns (2), (4), (6), and (8) report the impact of the spatial lag term of INT (WxINT) on local industrial carbon emission efficiency (ICEF). The coefficient of WxINT in Column (2) is statistically significant and negative at the 1% level, indicating that the industrial intelligent transformation of geographically adjacent cities exerts a significant negative spatial spillover effect on local carbon emission efficiency. Although the coefficient of WxINT in Column (4) is not statistically significant, it retains a negative sign. The coefficient of WxINT in Column (6) is 0.0187 and statistically significant and positive at the 1% level, demonstrating that the industrial intelligent transformation of cities with similar economic development levels or close economic connections can promote the improvement of local industrial carbon emission efficiency. The coefficient of WxINT in Column (8) is 0.0132 and statistically significant and positive at the 5% level, which also indicates that the industrial intelligent transformation of such cities generates a positive spatial spillover effect on local industrial carbon emission efficiency. However, its coefficient is smaller than that under the economic distance matrix, primarily because the dual weight overlap of geography and economy weakens the spillover intensity brought by a single economic connection. Thus, Hypothesis 2 is verified.

To further standardize the spatial econometric analysis, this study decomposes the spatial effects of industrial intelligent transformation on industrial carbon emission efficiency. Specifically, the direct effect reflects the net impact of local industrial intelligent transformation on its own industrial carbon emission efficiency; the indirect effect captures its spatial spillover effect on neighboring regions, and the total effect, defined as the sum of the above two effects, represents the comprehensive impact at the regional level.

Table 11. Spatial effect decomposition under geographic matrices.

	Adjacency Matrix			Geographic Distance Matrix
Variables	(1) Direct Effect	(2) Indirect Effect	(3) Total Effect	(4) Direct Effect	(5) Indirect Effect	(6) Total Effect
INT	0.0610 ***	−0.0259 ***	0.0351 ***	0.0601 ***	−0.00598	0.0541 ***
	(0.00229)	(0.00907)	(0.00104)	(0.00238)	(0.00477)	(0.00504)
ISTR	−0.353 ***	−0.344	−0.697	−0.338 ***	0.133 **	−0.205 ***
	(0.0287)	(1.138)	(1.136)	(0.0298)	(0.0624)	(0.0635)
POP	0.0135 ***	−0.492 ***	−0.478 ***	−0.00544 *	0.0146 ***	0.00920 **
	(0.00304)	(0.160)	(0.160)	(0.00305)	(0.00430)	(0.00405)
DIGI	0.0554 ***	0.427 ***	0.483 ***	0.0509 ***	0.0386 ***	0.0895 ***
	(0.00387)	(0.146)	(0.146)	(0.00454)	(0.00873)	(0.00707)
EDU	−0.0755	−37.01 ***	−37.08 ***	−0.156	−1.092 ***	−1.248 ***
	(0.159)	(11.76)	(11.79)	(0.157)	(0.291)	(0.304)
URB	−0.0264 **	−2.096 ***	−2.123 ***	0.00367	−0.103 ***	−0.0997 ***
	(0.0115)	(0.717)	(0.719)	(0.0112)	(0.0225)	(0.0228)
FIS	0.0106	−1.091 *	−1.080 *	−0.00414	−0.0784 **	−0.0825 ***
	(0.0167)	(0.575)	(0.573)	(0.0172)	(0.0315)	(0.0294)
FIN	−0.0242 ***	0.0934	0.0693	−0.0207 ***	0.00911 **	−0.0116 ***
	(0.00224)	(0.0669)	(0.0668)	(0.00224)	(0.00409)	(0.00431)
ECO	−3.94 × 10−5	0.0208	0.0208	0.000149	0.000764	0.000913
	(0.000732)	(0.0181)	(0.0180)	(0.000728)	(0.00131)	(0.00136)

Note: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust standard errors are in parentheses.

reports the decomposition results under the geographic adjacency matrix and geographic distance matrix. In the geographic adjacency matrix, the direct and total effects are both significantly positive at the 1% level, while the indirect effect is significantly negative at the 1% level. In the geographic distance matrix, the direct and total effects remain significantly positive at the 1% level, and the indirect effect is negative but statistically insignificant. These results are consistent with the negative characteristics of the Wx*INT coefficients reported earlier. The positive direct effect confirms the positive contribution of local industrial intelligent transformation to its own carbon emission efficiency. The negative indirect effect indicates that industrial intelligent transformation exerts a persistent negative impact on the carbon emission efficiency of neighboring regions through spatial transmission. The positive total effect suggests that industrial intelligent transformation generates an overall positive impact on regional industrial carbon emission efficiency.

Based on the regression results of the economic distance matrix and the geographic-economic nested distance matrix, the spatial effect decomposition is presented in

Table 12. Spatial effect decomposition under economic matrices.

	Economic Distance Matrix			Nested Matrix
Variables	(1) Direct Effect	(2) Indirect Effect	(3) Total Effect	(4) Direct Effect	(5) Indirect Effect	(6) Total Effect
INT	0.0554 ***	0.0330 ***	0.0884 ***	0.0565 ***	0.0278 ***	0.0843 ***
	(0.00225)	(0.00632)	(0.00666)	(0.00226)	(0.00648)	(0.00675)
ISTR	−0.306 ***	−0.419 ***	−0.726 ***	−0.289 ***	−0.266 ***	−0.554 ***
	(0.0278)	(0.0831)	(0.0865)	(0.0282)	(0.0867)	(0.0888)
POP	0.00532 **	−0.0202 ***	−0.0149 **	0.00604 ***	−0.0155 **	−0.00943
	(0.00229)	(0.00631)	(0.00703)	(0.00232)	(0.00659)	(0.00696)
DIGI	0.0545 ***	−0.00194	0.0526 ***	0.0575 ***	0.0181 **	0.0755 ***
	(0.00320)	(0.00714)	(0.00841)	(0.00319)	(0.00881)	(0.00910)
EDU	−0.284 *	1.533 ***	1.250 ***	−0.110	−0.623	−0.732 *
	(0.159)	(0.495)	(0.473)	(0.187)	(0.471)	(0.419)
URB	−0.0357 ***	0.101 ***	0.0648 *	−0.0376 ***	0.0101	−0.0276
	(0.0107)	(0.0331)	(0.0355)	(0.0108)	(0.0338)	(0.0352)
FIS	−0.0888 ***	0.151 ***	0.0626	−0.0972 ***	0.00559	−0.0916 **
	(0.0163)	(0.0418)	(0.0451)	(0.0166)	(0.0433)	(0.0442)
FIN	−0.0211 ***	0.000487	−0.0206 ***	−0.0219 ***	0.0186 ***	−0.00335
	(0.00202)	(0.00545)	(0.00604)	(0.00205)	(0.00581)	(0.00613)
ECO	0.000370	−0.000141	0.000229	0.000264	0.00188	0.00214
	(0.000659)	(0.00217)	(0.00234)	(0.000666)	(0.00216)	(0.00222)

Note: ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust standard errors are in parentheses.

. The results show that the direct, indirect, and total effects are all significantly positive at the 1% level, which is highly consistent with the significantly positive characteristics of the Wx*INT coefficients. The findings indicate that local intelligent transformation can not only significantly improve its own industrial carbon emission efficiency but also promote the coordinated improvement of cities with similar economic development levels via spatial spillovers, underscoring the vital role of intelligent transformation in the coordinated low-carbon development of regions.

6. Discussions

The empirical tests in this study confirm that industrial intelligent transformation significantly improves the industrial carbon emission efficiency of cities in China, which is consistent with the core conclusions of relevant studies [28,29] and further supplements the empirical evidence of the carbon emission reduction effect of intelligent transformation at the urban level. Different from the theoretical analysis that focuses on mechanism deduction alone, this study clarifies through empirical tests that reducing energy consumption intensity and promoting green technological innovation are two key mediating channels through which intelligent transformation exerts its effects, clarifies the specific transmission pathways by which intelligent transformation influences carbon emission efficiency, and makes up for the deficiency of insufficient empirical verification of relevant mechanisms in existing studies. Specifically, industrial intelligent transformation effectively reduces energy waste and carbon emission redundancy caused by factor misallocation through multi-dimensional improvements such as technological substitution and refined management, thereby enhancing carbon emission efficiency [12,16]. Moreover, its role in promoting green technological innovation is also supported by the empirical results [17,26], and the improvement of green technological innovation capacity drives the R&D of low-carbon materials, the application of energy-saving equipment, and the upgrading of environmental protection processes, which inevitably improves industrial carbon emission efficiency. By verifying the effectiveness of these two mediating paths, this study translates the carbon emission reduction mechanism of intelligent transformation from theoretical deduction to empirical verification.

The empirical results of the moderating effect reveal that both marketization level and environmental regulation intensity can significantly amplify the positive impact of industrial intelligent transformation on urban industrial carbon emission efficiency. This conclusion not only echoes existing relevant studies but also deepens the understanding of the enabling role of external condition through empirical validation. Specifically, the enabling logic of urban marketization level lies in breaking the barriers to factor flow, strengthening market competition constraints, and optimizing resource allocation efficiency, which is highly consistent with the research conclusions of existing studies [19,23] on marketization driving low-carbon transformation. Meanwhile, strict environmental regulation provides an important guarantee for the low-carbon effect of intelligent transformation through the dual mechanisms of mandatory constraints and goal guidance, which also supports the research views of existing studies [10,20] on environmental regulation guiding green transformation. The above conclusions provide a novel empirical perspective and research support for understanding the role of the external conditions in the coordination between intelligent transformation and low-carbon development.

Heterogeneity analysis shows that the impact of intelligent transformation on carbon emission efficiency presents a gradient characteristic of “eastern region > central region > western region”, which is closely related to the differences in regional development foundations and factor endowments [11,23]. Specifically, benefiting from first-mover advantages, the eastern region has well-developed digital infrastructure and agglomerated innovative resources, where intelligent technologies are deeply integrated with low-carbon processes, driving the optimization of carbon efficiency of the entire industrial chain through factor flow and technology spillover [17,29]. In the central region with a solid traditional industrial foundation, intelligent transformation has been rapidly rolled out in basic scenarios, which can effectively improve carbon efficiency [4]. However, due to the shortage of core technologies and high-end factors, the enabling effect is limited. The industrial structure in the western region is dominated by resource-based and high-energy-consuming industries [16,31], making it difficult for intelligent technologies to penetrate effectively, and the impact on industrial carbon emission efficiency is not significant. At the same time, the positive effect of intelligent transformation is significant in non-resource-based cities but not in resource-based cities. The main reason is that the industry of resource-based cities is dominated by high-energy-consuming resource extraction and primary processing, with a single industrial structure, high difficulty in emission reduction, and insufficient motivation for emission reduction. In contrast, non-resource-based cities have a more diversified industrial structure, dominated by low-energy-consuming high-end manufacturing and equipment manufacturing industries, which are more compatible with intelligent technologies. The above results are highly consistent with the realistic background that the transformation process of resource-based cities lags behind that of non-resource-based cities. This result supplements the literature on the heterogeneous impacts of intelligent transformation in different types of cities and provides targeted references for transformation practices in different regions.

In terms of spatial spillover effects, the empirical results show that the spatial spillover effect of industrial intelligent transformation on industrial carbon emission efficiency is significant and exhibits distinct heterogeneity from different dimensions of spatial proximity, which is highly consistent with theoretical expectations. From a geographical proximity perspective, there is a significant “competitive spillover effect”, that is, the improvement in the intelligent transformation level of geographically adjacent cities significantly reduces local industrial carbon emission efficiency. This aligns with the conclusions [4,14] on resource competition within factor market circles. Furthermore, this study confirms that the environmental externality from geographical proximity exacerbates the negative spillover, addressing the insufficient attention paid to environmental transmission paths in existing research. From an economic proximity perspective, the empirical results verify a significant “demonstration spillover effect”, that is, intelligent transformation in economically connected cities significantly improves local carbon emission efficiency [22,28]. This finding empirically validates the demonstration mechanism proposed at the theoretical level. Thus, we argue that similar cities, facing common transformation situations, can effectively reduce trial-and-error costs during the transformation process by learning from each other’s experiences. Additionally, the homogeneity of transformation pain points [17,29], the similarity of factor endowments [13,19], the policy coordination [10,21], and the demonstration-induced cycle [3,20] all support the conclusions of this section. In summary, the empirical findings reveal the heterogeneous spillover effects of intelligent transformation across different spatial proximity dimensions, which not only enriches the research on the relationship between intelligent transformation and carbon emission efficiency from the spatial dimension but also provides solid empirical support for advancing regional coordination between intelligence and low-carbon development and building an inter-regional transformation linkage mechanism.

It should be noted that this study still has certain limitations. First, the research perspective only focuses on the urban level, making it difficult to deeply analyze the micro-mechanism of the impact of industrial intelligent transformation on industrial carbon emission efficiency. Future research can further focus on the enterprise level to address the deficiencies of macro-level research. Second, this study only uses a single method to measure industrial carbon emission efficiency, which may lead to certain deviations in the measurement results. Future research can conduct cross-validation with multi-dimensional measurement indicators. Third, this study does not deeply explore the dynamic evolution characteristics of the impact of intelligent transformation on carbon emission efficiency and fails to reveal the heterogeneous performance of the relationship between the two at different development stages, which could serve as the focus of future research.

7. Conclusions and Policy Implications

Based on panel data of 284 Chinese cities from 2011 to 2023, this study comprehensively adopts various methods such as OLS, two-way fixed effects model, multi-period DID, IV-2SLS, SYS-GMM, and spatial SDM to examine the impact of industrial intelligent transformation on urban industrial carbon emission efficiency and its associated spatial spillover effects, which are critical to advancing industrial sustainability. The research conclusions are as follows. First and foremost, intelligent transformation significantly improves the industrial carbon emission efficiency of Chinese cities, and this conclusion remains robust to a series of endogeneity corrections and robustness checks. In addition, reducing energy consumption intensity and promoting green technological innovation are two key mediating channels. Meanwhile, the positive impact of intelligent transformation on urban industrial carbon emission efficiency is further enhanced as urban marketization level and environmental regulation stringency increase. Furthermore, the impact of intelligent transformation on industrial carbon emission efficiency exhibits significant heterogeneity: it is strongest in Eastern China, followed by Central China, and weakest in Western China; it is significant in non-resource-based cities but insignificant in resource-based cities. Finally, spatial spillover analysis shows that intelligent transformation generates a significant “competitive spillover effect” under geographical proximity, reducing the industrial carbon emission efficiency of adjacent cities, while it generates a significant “demonstration spillover effect” under spatial economic proximity, improving the industrial carbon emission efficiency of cities with similar economic development levels or close economic linkages.

The above findings carry important policy implications. First, local governments and relevant departments should integrate industrial intelligent transformation into their overall green development plans, which is crucial for advancing industrial sustainability. Financial support and tax incentives may be employed to encourage the development and industrial application of intelligent technologies, deepen the integration of intelligent technologies with all links of industrial production, improve energy efficiency and advance green technological innovation, thereby establishing a complete system from technology research to its industrial application and fostering an efficient and low-carbon industrial production model. Second, while intelligent technologies contribute to carbon emission reduction, their deployment demands coordinated regional planning. Increasing investment in low-carbon transition and promoting cross-city cooperation are essential measures. Without such coordination, the unregulated flow and competition of production factors, especially high-quality managerial and technical personnel, will lead to costly internal frictions. Furthermore, emission reductions in one region may be offset by increased emissions in other regions. Therefore, governments should guide geographically adjacent cities to mitigate the negative impacts of competition through rational industrial division of labor and resource sharing, support economically linked cities in establishing cooperation mechanisms to share advanced experience in low-carbon and intelligent development, and enhance cross-regional carbon emission accounting and benefit distribution systems to ensure the balanced allocation of transition gains. Third, supporting policies should be improved to strengthen the carbon mitigation effects of intelligent technologies. For peripheral cities lagging initially in low-carbon development, targeted policy preferences, resource inputs, and technical assistance can facilitate their catch-up growth and promote coordinated cross-regional low-carbon transition. Fourth, it is necessary to optimize the carbon emissions trading system (ETS) and strengthen the regulatory framework. The performance of industrial intelligent transformation can be linked to carbon quota allocation. Meanwhile, the regulatory model should be upgraded by combining smart supervision with targeted law enforcement, and appropriate policy incentives should be provided to enterprises with remarkable emission reductions, which will further enhance the driving role of market-based instruments and regulatory policies in industrial decarbonization. Fifth, given the heterogeneous impacts of intelligent transformation across regions, it is necessary to formulate region-specific strategies that align with local sustainable development goals. Economically developed regions should prioritize the R&D and comprehensive application of high-end intelligent technologies to play a leading and demonstration role. Regions with moderate development can strengthen technological cooperation with advanced areas and prioritize the intelligent upgrading of traditional industries based on their industrial foundations. Relatively underdeveloped regions should increase investment in digital infrastructure construction and talent training, and launch pilot projects of intelligent transformation tailored to local advantageous industries. Through tiered implementation and regional coordination, regional development disparities can be gradually narrowed, and the green and low-carbon transformation of industry can be effectively expanded to a wider scope.