AI and Green Efficiency in Technological Innovation: A Two-Stage Analysis of Chinese Rare Earth Enterprises

Abstract

As a scarce strategic resource, the efficient utilization of rare earth resources is crucial for ensuring national economic security and promoting sustainable development. AI, the core engine of the Fourth Technological Revolution, provides a favorable opportunity to drive green technological innovation. Green efficiency in technological innovation has not been adequately studied, and the relationship between green efficiency in the rare earth era and AI is still unclear. Based on the above research gap, this study employs the slack-based measure model to perform both static and dynamic evaluations of green efficiency in technological innovation during the technology development and transformation phases of eight listed Chinese rare earth enterprises from 2017 to 2021. This study aims to provide a policy basis for improving the green efficiency of the rare earth industry and the application of AI in the industrial chain. The findings reveal the following: (1) the green efficiency of technological innovation among these rare earth listed enterprises remains low in both phases, with low pure technical efficiency being a key factor contributing to the overall low green technology innovation efficiency; (2) total factor productivity in the technology development phase exhibits a fluctuating upward trajectory while demonstrating a general downward trend in the achievement transformation phase; and (3) the application of AI significantly enhances the green efficiency of technological innovation during the transformation phase, with a more pronounced impact compared to the technology development phase. This study contributes to the existing literature by extending previous research on AI and green efficiency, particularly within the context of the rare earth industry. The empirical results offer valuable policy recommendations for improving the utilization of rare earth resources and enhancing green technological innovation through AI integration.

1. Introduction

As global environmental protection becomes increasingly urgent, coupled with the need to enhance resource utilization and conversion rates, green technological innovation in enterprises—key market players and proponents of social change—has emerged as a critical factor in measuring their sustainable development capabilities and environmental social responsibility [1]. The ongoing scientific and technological revolution and industrial transformation are reshaping the innovation landscape and restructuring economic frameworks [2]. Research and development (R&D) in green technology and low-carbon solutions will drive international competitiveness and sustainable development. In China, for example, the government has introduced the “Implementation Plan on Further Improving the Market-Oriented Green Technology Innovation System (2023–2025)”, emphasizing green technology as a crucial element in achieving carbon peak and carbon neutrality. The efficiency of green technological innovation will determine the timeline for reaching carbon neutrality [3]. Therefore, the green efficiency of technological innovation has become a key indicator for governments and enterprises to explore the optimal use of natural resources and ensure environmental protection.

Amid sustained global economic development and deepening industrialization, industries such as new energy and electronic information are rapidly expanding. Rare earths, as a critical strategic resource, are being increasingly utilized in these sectors, leading to a surge in demand. China, as the world’s largest rare earth producer and exporter, accounts for the majority of global production [4]. Moreover, China’s rare earth industry has developed a comprehensive industrial chain, covering mining, smelting, processing, and application. However, the environmental pollution and high carbon emissions associated with rare earth mining and processing have become increasingly problematic [5]. Rare earth enterprises are facing substantial environmental pressure, which severely limits the industry’s sustainable green development [6]. Therefore, how to ensure the continued development of the rare earth industry while fostering technological innovation and green efficiency has become a pressing challenge.

Artificial intelligence (AI) technology, a core product of the new industrial era, provides innovative solutions for enterprises to reduce their carbon emissions and achieve green development [7]. First, AI technology can penetrate all links of the industry through deep integration with traditional industries. It can help enterprises optimize production processes, enhance energy efficiency, and improve conversion efficiency, thereby reducing carbon emission rates through data analysis and machine learning. This enables a dual improvement in technological innovation and green efficiency. Moreover, AI technology is capable of promoting the optimization and upgrading of the industrial chain. It can break down information barriers in the industrial chain, achieve seamless connection and collaboration between upstream and downstream, and form an intelligently linked economic network organization. This new mode of production can be expected to drive the efficient development of green technological innovation by promoting the integration and sharing of information, resources, and technology [8].

However, what is the trend of green efficiency in technological innovation in the rare earth industry? When AI permeates every aspect of the technological innovation process within this industry, what impact will it have on technological innovation? The answers to these two questions remain unclear. Therefore, exploring the essence and trend of green efficiency in technological innovation in the rare earth industry in depth and how AI technology impacts it would provide both practical guidance and theoretical support for the green transformation of rare earth enterprises.

In this context, this study will focus on the above two questions, developing an evaluation framework based on the SBM model to assess the green efficiency of technological innovation in rare earth enterprises and incorporating the undesirable outputs of carbon emissions. Specifically, the study will first analyze the current status of technological innovation and carbon emissions in rare earth enterprises, identifying existing problems and challenges. Second, using the SBM model for undesired outputs, an assessment system will be constructed in the following two phases: the “R&D input phase” and “achievement transformation phase”, followed by the evaluation of green efficiency. Third, by applying the non-expected output SBM model, a green efficiency assessment system will be developed, incorporating the Malmquist index to assess green efficiency both statically and dynamically. Fourth, the Tobit regression model will be used to explore how AI technology impacts the technological innovation process in rare earth enterprises, revealing its potential mechanisms for emission reduction and efficiency enhancement. Finally, policy recommendations and development strategies will be proposed to support rare earth enterprises in achieving technological innovation and green development, offering theoretical and practical guidance.

The main purposes and contribution of this study are as follows:

At the theoretical level, (1) through model innovation, this study will quantitatively assess the green efficiency of technological innovation in rare earth enterprises using the non-desired output SBM model, which integrates both desired and non-desired outputs (e.g., environmental pollution). This provides a more comprehensive evaluation of green efficiency, offering a new perspective for related research. (2) By integrating AI technology with technological innovation in rare earth enterprises, this study explores how AI can enhance the green efficiency of these enterprises. This approach is novel within the rare earth industry and highly innovative.

At the practical level, (1) the study of AI’s impact on technological innovation will facilitate the adoption and development of AI in the rare earth industry, providing new momentum for its transformation and high-quality development. It will also offer valuable insights for rare earth enterprises in addressing climate change and achieving green development. (2) Understanding how AI enhances green efficiency in technological innovation can help enterprises reduce environmental pollution, improve resource utilization, and meet market demand. This will offer new ideas and methods for technological innovation, driving progress in the rare earth industry. (3) This study will also serve as a useful resource for policymakers, providing a clearer understanding of the industry’s development and challenges, which will help in the formulation of more effective and scientifically grounded policies.

2. Literature Review and Research Hypothesis

2.1. Literature Review

Research on the green efficiency of enterprise technological innovation primarily focuses on efficiency evaluation and influencing factors.

Regarding efficiency evaluation, early studies mainly employed the traditional DEA model to measure efficiency, which considered only initial inputs and final outputs. These studies neglected the situational differences between the technological R&D phase and the results transformation phase of an enterprise’s innovation activities, failing to analyze the impacts of these different phases on overall efficiency [9]. To explore the internal structure of the decision-making unit and measure actual production efficiency, scholars began using network DEA models, which decompose the entire decision-making process into multiple phases, with each phase linked through intermediate indicators [10]. Furthermore, since traditional DEA models ignore slack in inputs and outputs, the non-radial slack-based measure data envelopment analysis (SBM-DEA) model was used to study the green efficiency of technological innovation in patent-intensive industries. Kumar applied the Malmquist–Luenberger (ML) productivity index method to analyze the total factor productivity of 41 countries from 1971 to 1992, using technology and efficiency change as criteria [11]. Bai employed the super-efficient SBM model to measure the efficiency of science, technology, and innovation in China from a static perspective and analyzed the dynamic aspect [12]. The two-phase DEA model decomposes efficiency into two phases, helping to determine whether inefficiency arises from management failure or scale marginal effects. It is clear that the SBM model, as an enhancement of the traditional DEA model, allows for the inclusion of “non-expected outputs”, providing a more comprehensive evaluation of technological innovation efficiency. Additionally, the Malmquist index overcomes the limitations of DEA models by evaluating both the static and dynamic efficiencies of innovation and offering a detailed analysis of total factor productivity and the green efficiency of systemic innovation.

A summary of the advantages and disadvantages of the efficiency measurement and evaluation models used in existing studies is shown in

Table 1. Comparison of the strengths and deficiencies of the efficiency measurement and evaluation models used in existing research.

Model Type	Strengths	Deficiencies
Traditional DEA Model	Simple and straightforward; widely used for efficiency evaluation	Neglects situational differences between technological R&D and results transformation phases; does not account for undesirable outputs; only measures static efficiency
Network DEA Model	Decomposes the decision-making process into multiple phases for a more detailed analysis	Neglects undesirable outputs and dynamic trends in efficiency
Non-radial SBM-DEA Model	Incorporates undesirable outputs for a more comprehensive evaluation; improves upon traditional DEA model	Limited analysis of dynamic trends and decomposition indexes in green economic efficiency
Malmquist–Luenberger (ML) Productivity Index	Evaluates both static and dynamic efficiency; provides detailed analysis of total factor productivity	Requires complex calculations; does not fully capture all nuances of green efficiency, particularly in specific industries or contexts

below.

Regarding influencing factors, the literature review reveals extensive research on the various aspects that affect the green efficiency of enterprise technological innovation [13,14,15]. Focusing on the micro-factor of AI technology, scholars have explored its impact on corporate production activities, organizational management, and social responsibility fulfillment. Liu analyzed panel data from 14 manufacturing sectors in China using empirical modeling and identified AI’s impact on technological innovation. AI accelerates knowledge creation and technological spillovers, increases R&D talent, and promotes investment, all contributing to technological innovation [16]. Huang measured the industrial robot penetration rate as an indicator of industrial intelligence and examined its impact on carbon emission efficiency in 11 provinces in China. Using a bidirectional fixed-effects model, the study demonstrated the relationship between industrial intelligence and carbon emission efficiency [17]. However, fewer studies have specifically analyzed the impact of AI on the green efficiency of technological innovation. The pathways through which AI affects green efficiency remain highly uncertain.

Based on the analysis of the existing literature, the following deficiencies are noted in current research: (1) Methodological issues: Existing efficiency measurement and evaluation models often face limitations such as focusing on a single phase and failing to account for non-expected outputs. Most studies only measure static green efficiency, lacking an analysis of dynamic trends, and fail to examine the characteristics of decomposition indexes in green economic efficiency. (2) Research gaps: Both theoretical and empirical studies on the influence of AI on the green efficiency of enterprise technological innovation require further exploration. Existing research has not sufficiently investigated the differential impacts of AI during the phases of technological R&D and results transformation. Moreover, environmental pollution and resource consumption issues are often ignored, and the consideration of non-desired outputs is insufficient. Additionally, there is a limited focus on rare earth enterprises, with most studies addressing the overall innovation efficiency of the rare earth industry or specific initiatives, without adequately examining the combined technology and industrial chains between AI and rare earth enterprises.

2.2. Research Hypotheses

The close integration of AI and industry marks a revolutionary change in the industrial production mode, which provides unprecedented opportunities and conditions for the development of green innovation activities. However, from a micro perspective, the innovation process of an enterprise is divided into two phases, namely the technology development phase and the achievement transformation phase. The role of AI in these two stages is not the same.

In the technology development phase, the research and development of rare earth materials usually requires deep professional knowledge and experience, while the accumulation and application of artificial intelligence technology in this specific field is not yet sufficient. In this case, it is unlikely that AI will have a significant impact on the green efficiency in the technological innovation of rare earth enterprises.

In the achievement transformation phase, the biased characteristics of AI are fully revealed. First, through the deep mining and analysis of data by AI, enterprises can better understand key information, such as market demand, consumer behavior, and product performance, thereby optimizing product design and production processes. Second, AI can also achieve green innovation by enhancing competitive advantages and optimizing decision making. Industrial robots are catalysts and endogenous driving forces for product innovation. They can not only bring competitive advantages to enterprises by leveraging the effectiveness of internal control, but also expand and optimize human information processing capabilities, improve decision-making quality, and promote green innovation in enterprises.

Based on the above analysis, regarding the impact of AI on the green efficiency of rare earth enterprises, the hypothesis is that, in the transformation phase, the impact of AI on green efficiency is more obvious than in the technology development phase.

3. Data and Methodology

3.1. Evaluation Model of Green Efficiency of Technological Innovation of Rare Earth Enterprises

3.1.1. SBM Models with Undesirable Outputs and Input–Output Indicators

The SBM model (slacks-based measure model) is a DEA model designed to assess the relative efficiency of decision-making units (DMUs) in a multiple-input, multiple-output setting, incorporating undesirable outputs [18]. In the SBM model, non-desired outputs are products or services that negatively affect the environment, economy, or society, such as carbon emissions [19]. This study selects the SBM-DEA model to construct a green efficiency evaluation model for the technological innovation of rare earth enterprises for several reasons. First, traditional DEA models (such as the BBC and CCR models) primarily focus on expected outputs. This can result in the slackness of inputs or outputs being ignored, leading to biased efficiency results. Consequently, the evaluation of decision-making unit productivity becomes inaccurate, hindering the development of decision-making units. Secondly, the SBM model incorporates the non-desired output of carbon emissions, which not only aligns with input indicators (e.g., R&D investment and green patent applications), but also more accurately reflects the green technological innovation status of enterprises. It overcomes the limitations of a single output indicator, making the evaluation results and countermeasures more scientific and reasonable. The SBM model uses non-radial relaxation to measure efficiency, directly addressing the redundancy of inputs and insufficiency of outputs, making it monotonic and unaffected by the statistical characteristics of the data (although the DEA model is flexible in efficiency evaluation, it also has some limitations. For example, it is a semi-parametric method, and its explanatory power and applicability may be limited when dealing with complex or specific types of data. Based on the specific context, this study has carefully considered those limitations and used the SBM model with undesirable outputs as the measurement model, which not only avoids the deviation caused by radial and angular measurements, but also considers the impact of undesired output factors in the production process of rare earth enterprises, which can better reflect the essence of efficiency evaluation and ensure the reliability of the results to a certain extent).

Assuming there are $n$ DMUs, $m$ inputs, and $s$ outputs, of which there are $s_1$ desired outputs and $s_2$ one non-desired output, the SBM model is expressed as follows:

$$\begin{array}{c}\min \rho ={\displaystyle \frac{1-\frac{1}{m}\sum _{i=1}^m \frac{s_i^{-}}{x_{i_0}}}{1+\frac{1}{s_1+s_2}\left(\sum _{r=1}^{s_1} \frac{s_r^g}{y_{r_0}^g}+\sum _{r=1}^{s_2} \frac{s_r^b}{y_{r_0}^b}\right)}}\\ \lambda \ge 0,s_i^{-}\ge 0,s_r^g\ge 0,s_r^b\ge 0\end{array}$$

In this SBM model equation, $\mathbf{X}=\left(x_{\mathit{ij}}\right)\in R^{m\times n}>0,\mathbf{Y}=\left(y_{\mathit{ij}}\right)\in R^{s\times n}>0$, denote the input and output matrices of each DMU, where the outputs are divided into desired and undesirable outputs. ${\mathbf{Y}}^g=\left(y_{\mathit{ij}}^g\right)\in R^{s_1\times n}>0$. The outputs are divided into desired outputs and non-desired outputs. ${\mathbf{Y}}^b=\left(y_{\mathit{ij}}^b\right)\in R^{s_2\times n}>0$. The set of production possibilities represents the set of outputs (including desired and undesirable outputs) that can be obtained from the inputs, denoted as $P=\left\{\left(x,y^g,y^b\right)\mid x\ge X\lambda ,y^g\le Y^g\lambda ,y^b\le Y^b\lambda ,\lambda >0\right\}$, where $\lambda$ is non-negative and can be regarded as a weight vector. For the first $i_0$ DMU, $\left(x_0,y_0^g,y_0^b\right)$ denotes the vector of inputs and outputs, and the elements denote the input slack residuals. $\left(s_i^{-},s_r^g,s_r^b\right)$. The elements denote the input shell margin, the desired output deficiency, and the undesirable output excess, respectively.

The DMU is in its most efficient state when ${\rho }^{\mathrm{*}}=1$, which is equivalent to the value of $s_i^{-\mathrm{*}}=0,s_r^{g\mathrm{*}}=0,s_r^{b\mathrm{*}}=0$, i.e., the values of input redundancy, desired output deficiency, or non-desired output excess are all zero, and the decision unit is fully efficient.

${\rho }^{\mathrm{*}}<1$ when, at this time, $s_i^{-}{s}_r^g, s_r^b$. If at least one of them is not 0, it indicates an efficiency loss in the decision-making unit, requiring corresponding improvements in inputs and outputs. The enterprise should appropriately regulate and adjust the levels of inputs, desired outputs, and non-desired outputs to improve the green efficiency of technological innovation.

This study categorizes the technological innovation process of rare earth enterprises into the following two phases: technology development and achievement transformation. It calculates the green efficiency of technological innovation for rare earth listed enterprises. Based on this, it is necessary to construct feasible and reasonable input and output indicators for both the technology development and achievement transformation phases. Specifically, in the R&D phase, the essence of technological growth lies in the large number of skilled personnel and substantial R&D funding. Therefore, this study selects R&D personnel and funds as input indicators. Regarding outputs, the green innovation capability of enterprises cannot be directly measured, and patents are often used as a proxy for innovation capability or performance. Patents provide valuable information about the technology, the invention, and the inventor and are relatively easy to access. In addition, patent filing, examination, and granting regulations are often consistent across regions, making patent data comparable. Since there is a time lag between filing and granting patents, patent applications serve as a metric for evaluating innovation performance. Similarly, the number of green patent applications can be used to measure the green efficiency of technological innovation. The non-expected output indicator is selected as the carbon emission intensity of enterprises. In the transformation phase, where green patents are applied, the output of technological innovation must be translated into economic benefits. Following the previous phase’s argumentation, the number of green patent applications continues to serve as an indicator for the number of effective patents in this phase. Therefore, the input indicator is the number of green patent applications. For output indicators, this paper uses operating profit and net production profit as measures of the economic benefits of technological innovation.

For convenience in data processing and to design a scientifically reasonable indicator system, the number of R&D personnel in the input indicator is calculated by the ratio of R&D personnel, and the R&D investment indicator is selected as the ratio of R&D expenditure to operating income. The net production profit indicator in the output indicators is calculated as the ratio of total net profit to operating income, and the operating profit indicator is calculated as the ratio of total operating profit to total operating income. The non-expected output indicator, carbon emission intensity, is measured by the ratio of carbon emissions to operating income in the region where the enterprise is located. The data are processed and analyzed using DEA-SOLVER Pro 5.0 software and Stata 16.

The summary of input–output indicators for the technology development and achievement transformation phases is showed in

Table 2. Summary of input–output indicators for the technology development and achievement transformation phases.

Point	Level 1 Indicators	Secondary Indicators	Description of Indicators
Technology development phase	Input indicators	Number of R&D staff	Percentage of R&D staff
		R&D investment	R&D expenditure as a percentage of operating revenue
	Output indicators	Green patent applications	Total number of green patent applications
	Indicators of non-expected outputs	Corporate carbon intensity	Carbon emissions of the district in which the enterprise is located divided by its business revenue
Transformation phase	Input indicators	Green patent applications	Green patent applications
	Output indicators	Net profit from production	Net profit as a percentage of operating income
		business profit	Operating profit as a percentage of operating revenue

.

3.1.2. ML Index

The SBM model provides a static measurement of efficiency, while the ML index offers a dynamic analysis, addressing the limitations of static analysis. To further analyze the trend of green technology innovation efficiency in rare earth enterprises, this study employs the ML index to calculate the dynamic efficiency of rare earth listed enterprises. The Malmquist index model is expressed as follows, with variable definitions in

Table 3. List of variable definitions for ML exponential model.

Variable	Variable Definition
M	Malmquist productivity change index
Dt	The distance function of the decision-making unit (DMU) in period t using the technology from period t as the reference technology
xt	The input quantities in period t
yt	The output quantities in period t

:

$$\mathrm{M}(x_t,y_t,x_{t+1},y_{t+1})={\displaystyle \frac{D^{t+1}\left(x_{t+1},y_{t+1}\right)}{D^t\left(x_t,y_t\right)}}\times {\left[{\displaystyle \frac{D^t\left(x_{t+1},y_{t+1}\right)}{D^{t+1}\left(x_{t+1},y_{t+1}\right)}}\times {\displaystyle \frac{D^t\left(x_t,y_t\right)}{D^{t+1}\left(x_t,y_t\right)}}\right]}^{{\displaystyle \frac{1}{2}}}$$

3.2. Tobit Regression Model

3.2.1. Modelling

After evaluating the green efficiency of technological innovation during the technology R&D and results transformation phases through the SBM model, this paper uses AI technology as the core explanatory variable to analyze its impact on the green efficiency of technological innovation in the two phases of rare earth enterprises. Since the efficiency values from the SBM model range from 0 to 1, and some DMUs are at the efficiency boundary of the DEA (i.e., efficiency equals 1), the Tobit regression model is chosen for this study. The regression models for the technology development and results transformation phases are constructed as follows:

$$y_{\mathit{it}}^{*}=x_{\mathit{it}}\beta +{\epsilon }_{\mathit{it}}$$

$$y_{\mathit{it}}=y_{\mathit{it}}^{*} , \mathit{if} y_{\mathit{it}}^{*}>0$$

$$y_{\mathit{it}}=0, \mathit{if} y_{\mathit{it}}^{*}<0$$

In the Tobit regression model equation, β₀ is the constant term; β_i is the estimated parameter in the model;${ x}_i$ represents each influence factor; and ε is the random error term.

3.2.2. Selection of Indicators

(1)

Dependent Variable: Level of Adoption of AI (AI)

This statistic serves as a quantitative indicator of a company’s advancement and effectiveness in the field of AI [20]. It provides a better understanding of a company’s competitiveness and future prospects in AI.

(2)

Dependent Variable: Green Efficiency in Technological Innovation (GTIE)

The SBM model measures the comprehensive technical efficiency (TE), pure technical efficiency (PTE), and scale efficiency (SE) of new energy enterprises during the technology R&D and results transformation phases. The green efficiency of technological innovation in these enterprises is reflected through the static efficiency of these three categories across the two phases.

(3)

Control Variables: Return on Net Assets, Net Profit Margin on Sales, Proportion of Intangible Assets, Size of Board of Directors

Return on Equity (ROE): As a key indicator of corporate profitability and capital use efficiency, ROE helps to control for differences between firms in terms of capital operating efficiency. By controlling for ROE, the impacts of other variables (e.g., dependent variables) on firm behavior or performance can be estimated more accurately, avoiding the misinterpretation of differences in capital operating efficiency.

Net Profit Margin on Sales (NPM): NPM reflects the profitability of a firm’s sales activities. In the Tobit model, controlling for NPM helps to distinguish between changes in performance due to sales efficiency and changes due to other factors. This allows for a more precise estimation of the net effects of other factors (e.g., market structure and competition) on firm behavior or performance.

Percentage of Intangible Assets (PIA): PIA reflects a firm’s asset structure and technological innovation capability. Controlling for PIA helps to understand its impact on firm performance and distinguishes its role from tangible assets. Changes in PIA are closely related to a firm’s innovation activities and branding, and controlling for this variable allows for a more accurate estimation of its effects on firm behavior.

Board Size (BS): BS is an important component of the governance structure, affecting decision-making efficiency and strategic direction. In the Tobit model, controlling for BS helps to understand the effect of governance structure on firm behavior or performance, distinguishing it from other governance mechanisms. The size of BS may impact decision-making speed and quality, and controlling for this variable helps to estimate its net effect on firm performance.

The specific variable definitions and measurement methods are shown in

Table 4. List of variable indicators in the Tobit model.

Variable Type	Variable Name	Variable Meaning	Methods of Measurement	Variable Code
Implicit variable	AI	Level of AI adoption	Book value of machinery/Total number of employees	×1
Independent variable	/	Technology Innovation Green Efficiency	Mean values of combined TE, PTE and SE for each of the two phases of measurement	Y
Control variable	ROE	Return on net assets	Net profit/shareholders’ equity	×2
	NPM	Net sales margin	Net profit/net sales revenue	×3
	PIA	Intangible assets as a percentage	Total intangible assets/Total assets	×4
	BS	Board size	Total number of board members	×5

below.

3.3. Data Sources

The sample in this study is selected from the directory of A-share listed enterprises in China. According to the 2021 industry classification by the Securities and Futures Commission (SFC), rare earth enterprises primarily fall under the “Mining” and “Manufacturing” categories. The team extracted and summarized information on listed enterprises meeting the criteria, focusing on keywords such as “rare earth mining”, “rare earth smelting”, and “rare earth processing”. Financial data, including total assets, intangible assets, operating income, and operating profit, were obtained from the Cathay Pacific Economic and Financial Research Database (CSMAR) and the WIND database. R&D data, including R&D expenditures, R&D personnel ratio, and the number of green patent applications, were sourced from the China Innovation and Patent Research Database (CIRD) in the China Research Data Service Platform (CNRDS). The carbon emission intensity of enterprises was measured using district- and county-level carbon emission data from the China Carbon Accounting Databases (CEADs). To mitigate the influence of outliers, this study integrated the financial and carbon emission data, excluded ST shares, and winsorized operating profit, net profit, and the number of employees at 1%. To ensure the completeness of the indicators, financial and carbon emission intensity data were combined and screened, and only years and enterprises with valid data across all three datasets were retained.

The static and dynamic efficiencies of the two phases were measured using data from 2017 to 2021. Ultimately, eight rare earth listed enterprises were identified, including Yingluohua (000795, Jinhua City in Zhejiang Province, China), Hongda Xingye (002002, Yangzhou City in Jiangsu Province, China), Northern Rare Earths (600111, Baotou City in Inner Mongolia autonomous Region, China), Arirang Xinmai (600206, Haidian District in Beijing, China), Ningbo Yunsheng (600366, Ningbo City in Zhejiang Province, China), Shenghe Resources (600392, Chengdu City in Sichuan Province, China), Xiamen Tungsten (600549, Xiamen City in Fujian Province, China), and North Mining Science and Technology (600980, Fengtai District in Beijing, China). For ease of measurement, the study assigns the codes C1, C2, C3, C4, C5, C6, C7, and C8 to represent the eight enterprises.

4. Empirical Results

4.1. Analysis of the Green Efficiency of Two-Phase Technological Innovation

4.1.1. Two-Phase Static Efficiency Analysis

To analyze the characteristics and trends of green technological innovation efficiency in the two phases of China’s rare earth enterprises, this study measured the mean efficiency levels across different enterprises in both phases using panel data.

TE represents combined TE, indicating the proximity of the decision-making unit to the optimal production boundary or technological frontier. The higher the TE, the closer the decision unit is to the optimal frontier, implying a greater efficiency. TE can be further decomposed into PTE and SE. PTE refers to the efficiency that accounts for the management and technological factors influencing production. SE represents efficiency influenced by the size of the enterprise. VRS denotes the degree of change in scale rewards, while CRS indicates constant returns to scale, IRS represents increasing returns to scale, and DRS denotes decreasing returns to scale.

The specific values of green technological innovation efficiency for the eight listed rare earth enterprises during the technology development phase are presented below in

Table 5. Green efficiency of technological innovation in rare earth firms during the technological development phase (2017–2021).

YEAR	ID	TE	PTE	SE	VRS
2017	C1	0.235847093	1	0.235847093	IRS
2017	C2	0.034745882	0.187703043	0.185110915	IRS
2017	C3	0.501267401	1	0.501267401	IRS
2017	C4	0.00683376	0.064500384	0.105949139	IRS
2017	C5	0.090356904	1	0.090356904	IRS
2017	C6	0.013603208	0.12192449	0.111570761	IRS
2017	C7	0.049709172	0.120342868	0.413062882	IRS
2017	C8	0.050905129	1	0.050905129	IRS
2018	C1	0.13697852	0.695578807	0.196927391	IRS
2018	C2	0.081824592	0.434630853	0.188262273	IRS
2018	C3	1	1	1	CRS
2018	C4	0.007746466	0.125829901	0.061562997	IRS
2018	C5	0.013178069	1	0.013178069	IRS
2018	C6	0.245668229	1	0.245668229	IRS
2018	C7	0.06668551	0.138983199	0.479809867	IRS
2018	C8	0.241495694	1	0.241495694	IRS
2019	C1	0.021770317	0.152520584	0.142736911	IRS
2019	C2	0.406520268	0.84308526	0.482181681	IRS
2019	C3	1	1	1	CRS
2019	C4	0.095003406	0.186002027	0.510765431	IRS
2019	C5	0.182272138	0.777722193	0.234366641	IRS
2019	C6	0.08874371	0.353381682	0.251127079	IRS
2019	C7	0.456292388	0.560260925	0.814428362	DRS
2019	C8	0.042984545	1	0.042984545	IRS
2020	C1	0.119611125	0.531022306	0.225246894	IRS
2020	C2	0.011970429	0.076871208	0.155720576	IRS
2020	C3	0.522533169	0.613442369	0.851804824	IRS
2020	C4	0.072425054	0.151641204	0.477608011	IRS
2020	C5	0.079705434	0.41376331	0.192635335	IRS
2020	C6	0.081242847	0.28178445	0.288315581	IRS
2020	C7	0.20735556	0.237131176	0.874433988	IRS
2020	C8	1	1	1	CRS
2021	C1	1	1	1	CRS
2021	C2	0.012409843	0.131379003	0.094458345	IRS
2021	C3	0.609382924	0.637796032	0.955451106	IRS
2021	C4	0.012967737	0.037981557	0.341421954	IRS
2021	C5	1	1	1	CRS
2021	C6	0.051325926	0.155997004	0.329018666	IRS
2021	C7	0.219208581	0.248654365	0.881579461	IRS
2021	C8	0.009693751	0.26676009	0.036338835	IRS
MEAN		0.25200662	0.538667257	0.407589974

. Between 2017 and 2021, the average comprehensive TE of these enterprises was 0.25200662, indicating a relatively low overall efficiency. A higher TE value reflects a better green technological innovation efficiency. The results suggest that China’s rare earth enterprises are still at a low level of green technological innovation. R&D funding and talent investment in technological innovation are insufficient, and there is significant room for improvement. This also implies that, while market demand for green technological innovation has increased, the current input and output levels of these enterprises are inadequate to meet this demand.

Over the 4 years of studies, the mean value of PTE for these eight enterprises was 0.538667257, which is higher than 0.5, suggesting relatively good resource allocation and management in the technological development process. The mean value of SE was 0.407589974, indicating an inefficient scale of development in these enterprises’ green technology efforts. This suggests that the overall rare earth industry has not achieved SE. The lower SE compared to PTE suggests that this low SE is the main factor contributing to the overall low comprehensive TE. This implies that these enterprises must improve their resource allocation and management, while ensuring that their production scale is optimized through timely adjustments. This includes adjusting factor inputs or improving technology development and organizational structure to better meet market demand and resource conditions. According to the scale status in Table 5, most of the eight firms were in the increasing returns to scale (IRS) status during the four-year period, indicating that scale expansion typically led to an increased economic efficiency. However, a few enterprises (e.g., C3, C5, and C8) reached constant returns to scale (CRS) in certain years. This may suggest that their production technology and management improved to a level where further scale expansion no longer significantly enhanced economic benefits.

Table 6. Average green efficiency of technological innovation for eight rare earth enterprises during the technology development phase (2017–2021).

ID	TE	PTE	SE
C1	0.302841411	0.675824339	0.360151658
C2	0.109494203	0.334733874	0.221146758
C3	0.726636699	0.85024768	0.861704666
C4	0.038995285	0.113191015	0.299461506
C5	0.273102509	0.838297101	0.30610739
C6	0.096116784	0.382617525	0.245140063
C7	0.199850242	0.261074507	0.692662912
C8	0.269015824	0.853352018	0.274344841
MEAN	0.25200662	0.538667257	0.407589974

below shows the average green technological innovation efficiency for the eight rare earth enterprises during the 2017–2021 period in the technology development phase. The comprehensive TE of each enterprise ranged from 0.1 to 0.8, with significant differences across enterprises. Among these eight, the following five had a comprehensive TE above the overall average: C1, C3, C5, and C8. Notably, C3 achieved a maximum comprehensive TE of 1 multiple times between 2017 and 2021 (Table 5), establishing itself as a benchmark enterprise in the rare earth industry. In contrast, C2, C4, and C6 had an average TE below 0.2, indicating poor overall green technological innovation, which hinders improvements in industry efficiency.

The PTE of each enterprise ranged from 0.11 to 0.86, showing significant differences in resource allocation, management, and technical capabilities. C3, C5, and C8 had pure technical efficiencies above 0.8, at the forefront of the industry. The SE of these enterprises ranged from 0.22 to 0.9, indicating considerable variation in SE. C3 and C7 had scale efficiencies above the overall average, suggesting more reasonable R&D scales. The SE of the remaining six enterprises was lower, which means that their technology development phases were not optimally scaled, limiting improvements in green innovation efficiency.

Decomposition of the comprehensive TE reveals that five out of the eight enterprises (C1, C2, C5, C6, and C8) had an SE lower than PTE. This indicates that a low SE is the primary barrier to green technological innovation efficiency. In contrast, C4 and C7 had a higher SE than PTE, indicating that a low PTE was the main reason for their low comprehensive efficiency. C3 showed a well-balanced development scale and technical level, with strong growth potential in the rare earth industry.

Table 7. Green efficiency of technological innovation in the outcome phase for rare earth enterprises (2017–2021).

YEAR	ID	TE	PTE	SE	VRS
2017	C1	0.041279781	0.041279781	1	CRS
2017	C2	0.505157062	0.553685897	0.912353133	DRS
2017	C3	0.031291739	0.031291739	1	CRS
2017	C4	0.067441807	0.067441807	1	CRS
2017	C5	0.357469428	1	0.357469428	DRS
2017	C6	0.416351668	0.416351668	1	CRS
2017	C7	0.110723469	0.110723469	1	CRS
2017	C8	0.325861837	0.325861837	1	CRS
2018	C1	0.061921374	0.061921374	1	CRS
2018	C2	0.166332712	0.166332712	1	CRS
2018	C3	0.014678118	0.014678118	1	CRS
2018	C4	0.10840098	0.10840098	1	CRS
2018	C5	0.267781836	0.267781836	1	CRS
2018	C6	0.04215163	0.04215163	1	CRS
2018	C7	0.044982333	0.044982333	1	CRS
2018	C8	0.123844253	0.123844253	1	CRS
2019	C1	0.186434661	0.186434661	1	CRS
2019	C2	0.063703333	0.063703333	1	CRS
2019	C3	0.012562168	0.012562168	1	CRS
2019	C4	0.009194436	0.009194436	1	CRS
2019	C5	0.01904703	0.01904703	1	CRS
2019	C6	0.01550051	0.01550051	1	CRS
2019	C7	0.00734262	0.00734262	1	CRS
2019	C8	0.29406444	0.29406444	1	CRS
2020	C1	0.028560882	0.028560882	1	CRS
2020	C2	1	1	1	CRS
2020	C3	0.018013786	0.018013786	1	CRS
2020	C4	0.01502648	0.01502648	1	CRS
2020	C5	0.096587477	0.096587477	1	CRS
2020	C6	0.057292219	0.057292219	1	CRS
2020	C7	0.019039611	0.019039611	1	CRS
2020	C8	0.06245723	0.06245723	1	CRS
2021	C1	0.005969972	0.005969972	1	CRS
2021	C2	0.843682045	0.843682045	1	CRS
2021	C3	0.066305862	0.132306737	0.501152576	DRS
2021	C4	0.100579001	0.100579001	1	CRS
2021	C5	0.026247386	0.026247386	1	CRS
2021	C6	0.135231799	0.135231799	1	CRS
2021	C7	0.0200164	0.0200164	1	CRS
2021	C8	0.731475433	0.731475433	1	CRS
MEAN		0.16300012	0.181926627	0.969274378

below presents the green technological innovation efficiency of rare earth enterprises during the transformation phase from 2017 to 2021. The average comprehensive TE for these eight enterprises between 2017 and 2021 was 0.163, indicating a low level of efficiency. This suggests that the ability of rare earth enterprises to convert technological achievements into economic benefits during this period was limited, and the alignment between green technology and market demand was poor. The average PTE over the four-year period was 0.182, indicating weak management and resource allocation abilities in these enterprises when realizing innovation outcomes. The mean SE was 0.969, suggesting that these enterprises maintained a reasonable development scale in the transformation phase. However, a low PTE resulted in low overall comprehensive TE. These enterprises must optimize their internal management and operations while maintaining their current production scale. Close attention to market dynamics and industry competition is necessary to improve green efficiency in technological innovation. Regarding SE, most enterprises exhibited IRS over the four-year period. Only C2 and C5 in 2017 and C3 in 2021 had decreasing returns to scale (DRS), suggesting that these enterprises may need to adjust their production scale to avoid efficiency losses. The remaining five enterprises achieved CRS, indicating that their production technology and management level reached a high standard, and further scale expansion no longer significantly impacted their economic efficiency.

As shown in

Table 8. Overall green technological innovation efficiency averages for the eight rare earth enterprises in the transformation phase (2017–2021).

ID	TE	PTE	SE
C1	0.064833334	0.064833334	1
C2	0.515775031	0.525480797	0.982470627
C3	0.028570335	0.04177051	0.900230515
C4	0.060128541	0.060128541	1
C5	0.153426631	0.281932746	0.871493886
C6	0.133305565	0.133305565	1
C7	0.040420887	0.040420887	1
C8	0.307540639	0.307540639	1
mean	0.16300012	0.181926627	0.969274378

below, there were significant differences in innovation efficiency among the eight rare earth enterprises during the transformation phase. In terms of comprehensive TE, the TE mean values of six enterprises—C1, C3, C4, C5, C6, and C7—were lower than the overall average (0.0648, 0.5158, 0.0286, 0.0601, 0.1534, and 0.1333 versus 0.0404). This indicates that the comprehensive efficiency of these six enterprises in technological innovation, green production, and sales services was relatively low, hindering overall industry efficiency. The TE values of C2 and C8 exceeded the average (0.5158 and 0.3075, respectively), indicating that these two enterprises performed better, achieving a higher comprehensive efficiency in the industry. However, overall, the green technological innovation efficiency of these enterprises remained low, and the economic benefits of green technology fell short of optimal levels.

Regarding PTE, the mean values for the eight enterprises ranged from 0.04 to 0.53, showing substantial variation. Five enterprises—C1, C3, C4, C6, and C7—had mean values below the overall average, suggesting poor resource allocation capabilities in transforming green technological innovations into economic benefits. These enterprises had a low operational efficiency and failed to effectively convert innovation into tangible economic results. For SE, the mean values for the eight enterprises ranged from 0.87 to 1, indicating that their development scale was generally reasonable, and they were efficient in scale expansion and resource utilization during the transformation phase.

The decomposition of comprehensive TE shows that SE exceeded PTE in all eight enterprises. A low PTE remains the main factor limiting comprehensive TE, indicating insufficient resource allocation and a low management capacity. The rare earth industry must focus on improving organizational and management capacity, as well as resource allocation, to elevate its technical performance.

4.1.2. Dynamic Efficiency Analysis

This study employs the Malmquist index model to calculate the total factor productivity (TFP) index for the eight rare earth enterprises in both phases from 2017 to 2021. The analysis measures dynamic levels and development trends. TFPCH represents the total factor productivity change, EFFCH represents the TE change index, and TECHCH represents the technological progress index. The relationship is given by the following: TFPCH = TECHCH * EFFCH. The EFFCH value is decomposed into pure technical efficiency change (PECH) and scale efficiency change (SECH). The results are presented in

Table 9. Overall total factor productivity index and its decomposition for the eight firms in the technology R&D phase (2017–2021).

YEAR	EFFCH	TECHCH	PECH	SECH	TFPCH
2017–2018	5.860196299	1.422652932	1.578956374	5.969693595	8.172655316
2018–2019	13.79186347	1.264380792	1.607291859	19.35061223	14.27067136
2019–2020	6.421999847	0.762535706	2.025344821	2.158396972	2.399633168
2020–2021	0.738119595	1.833591491	0.747411379	1.114495706	1.568386159
MEAN	6.703044802	1.32079023	1.489751108	7.148299625	6.602836501

,

Table 10. Total factor productivity and its decomposition by firms in the technological development phase (2017–2021).

2017–2018
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	1	0.673757757	1	1	0.673757757
C2	1.372749695	1.675944397	0.528587992	2.597012638	2.300652161
C3	1	1.193356006	1	1	1.193356006
C4	0.60508928	1.973993668	6.269297221	0.096516285	1.194442408
C5	0.031851937	1.202773969	1	0.031851937	0.038310681
C6	41.06793548	1.388853892	1	41.06793549	57.03736204
C7	0.803943992	1.679021683	0.833765782	0.964232413	1.349839395
C8	1	1.593522083	1	1	1.593522083
2018–2019
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	0.023806373	1.545459679	0.274323565	0.086782091	0.036791789
C2	12.22126468	1.059382703	1.891832609	6.460013758	12.94699642
C3	1	1.083741478	1	1	1.083741478
C4	12.86008576	0.917638523	0.186002027	69.13949243	11.8009101
C5	75.88365424	1.014644209	1	75.88365423	76.99491034
C6	0.08874371	2.495345633	0.421388735	0.210598202	0.221446229
C7	7.25735297	1.451370965	7.084787932	1.024357121	10.53311138
C8	1	0.547463147	1	1	0.547463147
2019–2020
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	42.00555942	0.324531754	3.645330289	11.5231148	13.63213787
C2	0.019967476	0.726351893	1	0.019967476	0.014503414
C3	1	0.605787008	1	1	0.605787008
C4	1.272608069	0.596669951	5.376285486	0.236707681	0.759326994
C5	0.572280592	0.34357601	0.80803712	0.708235523	0.196621883
C6	4.712028782	0.156109705	2.373105677	1.985595849	0.735593424
C7	0.793554443	0.456123045	1	0.793554443	0.361958469
C8	1	2.891136285	1	1	2.891136285
2020–2021
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	1	3.392449745	1	1	3.392449745
C2	0.930291485	1.137416353	1	0.930291485	1.058128748
C3	1	1.135314323	1	1	1.135314323
C4	0.152210061	1.120837307	0.108704252	1.400221775	0.170602715
C5	1.747394572	3.074527675	1.237566908	1.411959677	5.37241297
C6	0.190987772	1.906567162	0.283590762	0.673462601	0.364131015
C7	0.867949213	1.182478077	0.349429106	2.483906458	1.026330916
C8	0.016123654	1.719141289	1	0.016123654	0.027718839

,

Table 11. Overall total factor productivity index and its decomposition for the eight enterprises in the transformation phase (2017–2021).

YEAR	EFFCH	TECHCH	PECH	SECH	TFPCH
2017–2018	1.297145589	0.534604549	1.418481821	0.927690828	0.692324741
2018–2019	0.841548289	1.089700842	0.719603465	1.084085903	0.915999321
2019–2020	1.120793883	3.40124097	0.668815931	1.451977953	3.811415186
2020–2021	3.973817427	0.843733626	10.48410048	0.791929452	3.352649334
MEAN	1.808326297	1.467319997	3.322750423	1.063921034	2.193097145

and

Table 12. Total factor productivity and its decomposition for each enterprise in the transformation results phase (2017–2021).

2017–2018
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	2.806250931	0.535071475	3.69992168	0.758462	1.501544824
C2	0.615619035	0.534650843	1	0.615619	0.329141236
C3	0.870581125	0.536042194	1.05869493	0.822315	0.466668216
C4	3.016227933	0.531434442	2.50458951	1.20428	1.602927407
C5	1.412782912	0.531736953	1	1.412783	0.75122888
C6	0.188891157	0.536517618	0.18335016	1.030221	0.101343434
C7	0.759628418	0.535415137	0.82493507	0.920834	0.406716553
C8	0.707183198	0.535967731	1.0763632	0.657012	0.379027374
2018–2019
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	2.765741495	1.090094398	2.09771052	1.318457	3.014919311
C2	0.351915791	1.08962832	1	0.351916	0.383457412
C3	0.79141226	1.0844855	0.65079047	1.216078	0.85827512
C4	0.07768739	1.094362059	0.07768739	1	0.085018132
C5	0.064743217	1.092406603	0.06474322	1	0.070725917
C6	0.33845012	1.086516305	0.2872552	1.178221	0.367731574
C7	0.150226276	1.092760608	0.13833349	1.085972	0.164161356
C8	2.192209762	1.087352946	1.44030742	1.522043	2.383705743
2019–2020
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	0.045040429	3.40304033	0.04504043	1	0.153274398
C2	4.615823622	3.400164616	1	4.615824	15.69456015
C3	0.422019665	3.397811194	0.42201966	1	1.433943141
C4	0.479970055	3.402609905	0.47997006	1	1.633150864
C5	1.49185477	3.398366629	1.49185477	1	5.069869466
C6	1.087793446	3.404032231	1.08779345	1	3.702883951
C7	0.761391849	3.403303432	0.76139185	1	2.591247494
C8	0.06245723	3.400599422	0.06245723	1	0.212392021
2020–2021
DMU	EFFCH	TECHCH	PECH	SECH	TFPCH
C1	0.247192015	0.844050717	0.24719201	1	0.208642597
C2	1	0.843888517	1	1	0.843888518
C3	4.36225503	0.844202255	55.5130379	0.078581	3.682625534
C4	7.933266271	0.843795839	7.93326627	1	6.694057072
C5	0.321950005	0.844359512	1.25343149	0.256855	0.271841549
C6	2.798006567	0.842238566	2.79800657	1	2.356589038
C7	1.246238618	0.843613848	1.24623862	1	1.051344156
C8	13.88163091	0.843719753	13.8816309	1	11.71220621

.

As shown in Table 9, the average total factor productivity change (TFPCH) index for the eight enterprises in green technological innovation from 2017 to 2021 was 6.6028, indicating an overall upward trend in TFPCH over the four-year period. Specifically, the TFPCH values were higher in the periods from 2017–2018 and 2018–2019, at 8.1727 and 14.2707, respectively, showing significant growth in total factor productivity during these periods. However, in 2019–2020, TFPCH declined to 2.3996, indicating a slowdown in productivity growth. By 2020–2021, TFPCH further decreased to 1.5684, suggesting limited growth in total factor productivity. This trend aligns with China’s “green development” goals for 2020–2021, during which enterprises transitioned from rapid expansion to high-quality development, adjusting their scale and TE, leading to slower productivity growth.

Looking at the decomposition of the indices for each period, in 2017–2018, the EFFCH value was 5.8602, indicating significant improvements in management and operational efficiency. The TECHCH value was 1.4227, suggesting that technological progress also contributed to efficiency gains. Meanwhile, the PECH value was 1.5790, indicating improvements in PTE. The SECH index was 5.9697, highlighting that the improvement in SE was a key driver of overall efficiency growth. In 2018–2019, the EFFCH value increased to 13.7919, reinforcing that improvements in management and operational efficiency continued to drive overall growth. The TECHCH value rose to 1.2644, showing that technological progress played an increasing role in efficiency improvement. Despite a slight decline in the PECH value, the SECH value surged to 19.3506, indicating that further expansion of SE was critical to efficiency gains. In 2019–2020, the EFFCH remained high at 6.4220, but the TECHCH value dropped significantly to 0.7625, causing technological progress to have a negative contribution to efficiency improvement. The PECH value increased, but the SECH value declined, suggesting that the reduction in SE affected overall performance. Finally, in 2020–2021, the EFFCH value declined sharply to 0.7381, implying that managerial and operational efficiencies were challenged during this period. Although the TECHCH value increased significantly to 1.8336, the decline in both PECH and SECH values indicates reductions in pure technical and scale efficiencies, negatively affecting overall efficiency. In summary, the impact of technological progress on efficiency during the technology development phase fluctuated significantly across the years, while changes in PTE remained relatively stable, with minimal overall improvement. The effect of SE changes on the green efficiency of technological innovation was more pronounced, particularly during periods of efficiency improvement. Total factor productivity showed an upward trend followed by a downward trend, reflecting the ongoing challenges faced by rare earth enterprises in terms of productivity during technological R&D.

Overall, these eight rare earth enterprises exhibited varying degrees of fluctuation in the TECHCH and EFFCH indices. Among them, the efficiency indices of C2, C4, C6, and C7 showed a greater volatility. For instance, C6 had a high efficiency change index of 41.0680 in 2017–2018, but the fluctuations in the following three years were significant, indicating an unstable management efficiency. In contrast, the green innovation index of C3 and C8 remained relatively stable, suggesting that these enterprises maintained more consistent inputs and modes of green technological innovation. C1 saw a significant increase in the technological innovation index from 2019–2020, but its growth rate slowed in 2020–2021, indicating that technological innovation is a continuous process requiring long-term investment and attention. C5 had a high index of technological innovation in 2017–2018, but the technological progress change index was low in 2019–2020, showing slower innovation progress. This indicates the need for increased R&D investment and technological innovation.

The changes in the total factor productivity index reveal that the production efficiency of C2, C4, C6, and C7 fluctuated significantly. C2 exhibited considerable fluctuations in both technological innovation and production efficiency, while C1, C3, and C8 showed a relatively stable production efficiency, suggesting more effective production management measures in these enterprises. C3 was stable in both technological innovation and production efficiency, indicating a more balanced development in both areas.

Based on the above discussion, C3 should continue with its current strategies, increase R&D investment, and expand its production scale to maintain its competitive advantage. C1, with significant index changes in certain years, should focus on enhancing its technological innovation and R&D capabilities to stabilize productivity and overall efficiency. Enterprises like C2, C4, and C7, with noticeable fluctuations in the technological progress index, should focus on improving resource allocation, management capabilities, and adjusting their development scale. This will help to improve SE and avoid further declines in both SE and PTE, ultimately boosting output efficiency.

Table 11 reflects the change in the total factor productivity index for rare earth enterprises during the transformation phase. The mean value of TFPCH was 2.1931, indicating that total factor productivity increased across the eight enterprises in this phase. In terms of TE, the mean value of EFFCH was 1.8083, suggesting an overall improvement in TE during the transformation phase. However, the change in TE was lower in 2018–2019, with a value of only 0.8415. The mean value of TECHCH was 1.4673, signifying that technological progress had been made. The highest change in technological progress occurred in 2019–2020, with a value of 3.4012, indicating significant results in transforming green technological innovations into economic benefits that year. The mean value of PECH was 3.3228, showing that the pure technical efficiencies for all eight enterprises consistently increased, maintaining a positive trend. The lowest value of PTE change (0.6688) occurred in 2019–2020. The mean value of SECH was 1.0639, with the highest change in SE occurring in 2019–2020, reaching 1.4520, a 45.2% increase. This improvement in TE contributed to a steady increase in total factor productivity. Further analysis shows that the improvement in TE was the result of the combined effects of both PTE and SE improvements.

Overall, the indices of the eight enterprises showed significant fluctuations over different years. C1 enterprises experienced large fluctuations in TFPCH from 2017 to 2021. TFPCH showed high growth in 2017–2018 and 2018–2019 but declined sharply in 2019–2020, followed by a slight rebound in 2020–2021. In terms of decomposition indicators, TECHCH was notably high in 2019–2020, but EFFCH was extremely low, indicating that C1 enterprises heavily relied on technology introduction or breakthroughs during this period, while facing shortcomings in resource allocation and organizational and management capabilities. The overall index of C2 enterprises fluctuated significantly, with EFFCH experiencing a sharp decline in 2018–2019 but a significant increase in 2019–2020. This suggests a major adjustment in management strategy and operations. TECHCH remained stable overall but did not sustain the growth of TFPCH. C3 enterprises showed a relatively stable overall efficiency, with a smooth growth trend in specific indices. This indicates a strong stability and continuity in both results transformation and efficiency management. C4 enterprises achieved high TFPCH growth in 2017–2018 and 2018–2019, but experienced sharp declines in the subsequent years. TECHCH fluctuated greatly, while EFFCH remained relatively stable, suggesting greater instability in results transformation and commercialization capabilities. C5 enterprises maintained a low TFPCH in most years, which is closely related to a poor performance in both TECHCH and EFFCH. Therefore, C5 enterprises must increase investment and adjust strategies to meet market demands and enhance technological competitiveness. The indices of C6 enterprises also fluctuated considerably. While TECHCH remained high in most years, EFFCH varied greatly, indicating strengths in results transformation but instability in commercialization ability. C7 enterprises showed a steady growth trend in TFPCH, with both TECHCH and EFFCH remaining relatively stable, suggesting a good balance between technological innovation and efficiency management. TFPCH for C8 enterprises declined sharply in 2019–2020 but experienced significant growth in subsequent years. Both TECHCH and EFFCH showed large fluctuations, indicating a need for C8 enterprises to strengthen the stability and sustainability of technological innovation and efficiency management.

In summary, by 2021, the performance of these eight rare earth enterprises had improved compared to 2017. There was a gradual increase in their emphasis on product and service innovation, as well as quality improvement. While total factor productivity improved to some extent, overall green efficiency remained low and fluctuated greatly. This suggests that the rare earth industry has yet to achieve an effective connection between innovation and industrialization, and that the challenge of transforming technological innovation into economic benefits remains critical for achieving green development and transformation.

4.2. Impact of AI Technology on the Green Efficiency of Two-Phase Technological Innovation

As shown in

Table 13. Results of Tobit regression model test for technology development phase.

Y	Coef.		St. Err		t-Value	p-Value		Sig.
×1	−0.029		0.140		−0.21	0.833
×2	−0.901		2.374		−0.38	0.704
×3	0.248		1.668		0.15	0.882
×4	2.707		3.112		0.87	0.384
×5	0.482		0.282		1.71	0.087		*
_cons	−0.443		1.857		−0.24	0.811
sigma_u	0.000		0.110		0.00	1.000
sigma_e	0.327		0.042		7.82	0.000		***
Mean dependent var		0.270		SD dependent var			0.338
Number of obs		34.000		Chi-square			4.418
Prob > chi2		0.491		Akaike crit. (AIC)			41.613

*** p < 0.01, * p < 0.1.

below, the coefficient for the degree of adoption of AI (×1) at the technology development phase was −0.029, with a p-value of 0.883 (significantly greater than 0.1). This suggests that AI adoption has an insignificant effect on the green efficiency of technological innovation (Y). The coefficient of return on net assets (×2) was −0.901 with a p-value of 0.704, indicating a non-significant effect. The coefficients for net sales margin (×3) and intangible asset share (×4) were both positive but insignificant, indicating that these variables do not significantly affect the green efficiency of technological innovation. The coefficient for board size (×5) was positive (0.482) and significant at the 10% level, indicating that board size has a significant positive effect on green efficiency in technological innovation.

According to the empirical results, AI technology has not yet promoted technological innovation in rare earth enterprises. The technological innovation of rare earth enterprises, as an emerging field of strategic resources, is still in its early stages. The application of AI technology in these enterprises is likely still in its primary phase. Although AI technology has advanced significantly in certain areas, its specific applications in the rare earth industry may require further exploration and optimization. Challenges such as data quality, algorithm optimization, and model generalization capabilities may limit the effectiveness of AI in technological R&D, affecting its impact on green efficiency. Moreover, AI technology is still in a continuous development phase, and its influence on the green efficiency of technological innovation in enterprises will likely become clearer over a longer period. The depth and breadth of AI integration with rare earth enterprises are also crucial factors affecting green efficiency. Currently, only a few enterprises have begun to apply AI technology in their technological innovation processes, while most may still be in the early stages or hesitant to adopt it. Furthermore, even those rare earth enterprises that have started applying AI technology may be using it only in specific processing phases, failing to achieve comprehensive coverage across the entire technology development process. This localized application may prevent AI from fully realizing its potential, thus limiting its effectiveness in enhancing green efficiency.

As shown in

Table 14. Results of the Tobit regression model for the transformation phase of results.

Y	Coef.		St. Err		t-Value	p-Value		Sig.
×1	0.127		0.064		1.98	0.048		**
×2	−0.114		1.085		−0.10	0.916
×3	2.169		0.762		2.85	0.004		***
×4	0.119		1.419		0.08	0.933
×5	−0.135		0.128		−1.06	0.291
_cons	−1.395		0.850		−1.64	0.101		*
sigma_u	0.000		0.036		0.00	1.000
sigma_e	0.149		0.019		8.02	0.000		***
Mean dependent var		0.136		SD dependent var			0.201
Number of obs		34.000		Chi-square			28.679
Prob > chi2		0.000		Akaike crit. (AIC)			−12.443

*** p < 0.01, ** p < 0.05, * p < 0.1.

below, the coefficient for the degree of AI adoption (×1) is positive (0.127) and significant at the 5% level (p = 0.048), indicating that AI adoption positively affects the green efficiency of technological innovation. As rare earth enterprises adopt more AI technology during the results transformation process, the green efficiency of technological innovation is likely to increase. The coefficient for return on net assets (×2) is negative (−0.114), but the effect is insignificant, indicating that return on net assets does not significantly influence the green efficiency of technological innovation. The coefficient for net sales margin (×3) is positive (2.169) and significant at the 1% level (p = 0.004), showing a strong positive effect on the green efficiency of technological innovation. An increase in the net sales margin suggests greater profitability, providing more resources for investment in technological innovation and green production, thus enhancing green efficiency. The coefficient for intangible assets share (×4) is 0.119, but p = 0.933, indicating that the effect of intangible assets share on technological innovation green efficiency is insignificant. The coefficient for board size (×5) is −0.135 and non-significant (p = 0.291), suggesting that board size does not have a direct or significant effect on technological innovation green efficiency during the results transformation phase.

The empirical results provide strong statistical support for the hypothesis.

5. Discussion

The non-radial SBM-DEA model with the Malmquist index takes carbon intensity into consideration as an undesirable output superior to the traditional DEA model and static analyses [21,22]. It adds a carbon emission redundancy term to the efficiency equation. When carbon emissions exceed the efficient frontier, efficiency losses occur. In this way, it quantifies the impact of environmental constraints on innovation efficiency. It also conducts a dual evaluation of static and dynamic efficiency. For the static efficiency evaluation, the comprehensive technical efficiency is calculated based on inter-temporal cross-sectional data. For the dynamic analysis, the Malmquist index is combined to decompose the total factor productivity into technical efficiency change and technological progress, tracking the driving factors of efficiency fluctuations. By adjusting the slack variables, it avoids the neglect of undesirable outputs in traditional radial DEA. The dual evaluation mechanism can reflect both the immediate efficiency status and reveal the long-term evolution pattern. It is in line with the multi-dimensional and sequential characteristics of green innovation. Thus, it enhances the scientific nature of the evaluation system and its policy guidance value. It also makes up for the deficiencies in previous studies, which ignored undesirable outputs (such as carbon emissions) and dynamic trends.

The dynamic Malmquist analysis highlights the fluctuation trend of total factor productivity, especially the differences between technological progress and scale efficiency in the technology development phase. This innovative approach provides a detailed understanding of how scale expansion and management practices interact over time. This is a dimension that was overlooked in previous studies. Meanwhile, the Malmquist index analysis shows that the total factor productivity in the technology development phase presents a fluctuating trajectory, while in the phase of achievement transformation, it shows a trend of steady growth. In the technology development phase, scale efficiency drives 72% of the fluctuations in total factor productivity [23]. This indicates that rare earth enterprises tend to prioritize scale expansion rather than technological optimization at the technical level. In contrast, in the achievement transformation phase, the increase in pure technical efficiency (average value = 0.18) indicates that the commercialization ability is gradually improving. This is consistent with the market-oriented green innovation system advocated by Zhang [24]. These dynamic insights provide practical policy focus points for policymakers.

Regarding AI adoption, our results challenge the view presented by Wang that companies with extensive AI adoption exhibit a higher green innovation efficiency [25]. At the micro level of innovation activities, there are significant differences in the impacts of AI adoption between the two phases of technology development and achievement transformation. In the phase of achievement transformation, the adoption of AI can significantly improve green efficiency. However, in the technology development phase, the impact of AI adoption is not statistically significant. The research results indicate that the effectiveness of AI in research and development depends on factors specific to the industry. This may be due to the fragmentation of the data ecosystem in the rare earth industry and the immaturity of algorithms for specific processes. At the same time, rare earth technology development requires profound domain expertise and faces the problem of data scarcity. These factors limit the applicability of AI. The research results suggest that such differences in AI adoption are highlighted in different innovation stages. Therefore, rare earth enterprises need to have different emphases when formulating innovation strategies regarding AI adoption.

6. Conclusion and Recommendations

6.1. Conclusions

Green technological innovation in rare earth enterprises is a key pathway to promoting China’s industrial structure transformation toward a green, low-carbon economy and achieving the “dual carbon” goal. This paper selects eight rare earth enterprises to comprehensively measure their green efficiency in technological innovation from 2017 to 2021. Using the SBM model, the carbon emission intensity is incorporated as an undesirable output, and the Malmquist index is applied to analyze the technological innovation efficiency. A Tobit regression model is constructed to assess the impact of AI on the technological innovation of rare earth enterprises. The model analyzes whether AI’s impact on the green efficiency of technological innovation is significant and identifies its direction. The goal is to encourage technological innovation across the rare earth industry chain, including mining and processing, to develop in a greener, higher-quality direction. The main findings are as follows:

Firstly, in terms of overall green efficiency, the technological innovation efficiency of rare earth listed enterprises from 2017 to 2021 remains low during both the R&D and transformation phases, indicating that the industry faces significant challenges in technological innovation and green production. Regarding PTE, while some enterprises perform better, the overall performance remains low, and there is a need to strengthen technological innovation and management optimization. In both phases, SE is relatively high, with most enterprises having an SE value greater than 0.8. SE exceeds PTE for most enterprises, suggesting that green technological innovation efficiency is more dependent on scale than on pure technical performance. This low level of PTE is the primary reason for the inability to achieve a higher level of green technological innovation efficiency in both phases. Therefore, rare earth enterprises should not only increase R&D investment, but also focus on improving technology and optimizing the use and management of innovation resources.

Secondly, the analysis of dynamic total factor productivity reveals that the average efficiency at each phase generally shows a fluctuating upward trend. Changes in total factor productivity during the technology development phase display volatility and uncertainty. The impacts of technical progress, PTE, and SE on overall efficiency vary across different years. In the results transformation phase, the total factor productivity of rare earth enterprises shows a rising trend, indicating that China’s rare earth enterprises are in an upward phase of transformation. However, the overall green technological innovation efficiency remains low, suggesting that enterprises must focus on improving the economic benefits of rare earth production. This should occur alongside improvements in technical levels, as well as increased attention paid to the output of scientific research and the efficiency of its transformation.

Thirdly, the current impact of AI on the green technological innovation of rare earth enterprises is primarily evident in the results transformation phase. In contrast, AI does not show a significant positive effect in the technological development phase. The role of AI in technological innovation is not fully realized at present.

6.2. Recommendations

First, rare earth enterprises should transition to a “chain development” model to strengthen their core competitiveness. In recent years, China’s rare earth industry has largely achieved vertical integration. Instead of directly marketing raw ores, mining enterprises now process rare earths into high-value-added products before entering the market. To capitalize on their strategic role in resource innovation, rare earth enterprises should actively extend their industrial chains. This includes reaching upstream to secure raw material supply and downstream to expand market access. Furthermore, through resource allocation strategies such as equity swaps and capital injections, enterprises should optimize their organizational and managerial capabilities, internalize external market transactions, and enhance the overall effectiveness of scientific and technological achievement transformation.

Second, to promote the innovation in the AI field, rare earth enterprises should enact the following: (1) Establish an innovation ecosystem centered around AI. Rare earth enterprises must deepen their intelligent transformation, restructure R&D and innovation structures, and foster a strategic plan and organizational culture conducive to knowledge integration. They should enhance data openness and sharing, remove barriers to innovative resource integration, and develop an open innovation ecosystem to broaden AI-driven applications. (2) Collaborate with governments to establish innovation funding mechanisms. This includes allocating special funds for technological innovation, developing venture capital, ensuring stable capital support, and implementing intelligent transformation policies tailored to the operational realities of enterprises. These steps will support technological innovation and foster a collaborative innovation system. (3) Strengthen technological iteration capabilities by formulating AI strategies. Enterprises should prioritize upgrading capacity, nurture professional talent, intensify employee training, and establish a self-reinforcing mechanism for continuous iteration and growth, ensuring sustained innovation.

Third, the government and rare earth enterprises must collaborate to strengthen industry–university–research cooperation and facilitate the integration of economic and technological values in innovation outcomes. Technological R&D in the rare earth industry requires substantial financial support and the active involvement of high-caliber scientific researchers. By deepening cooperation among rare earth enterprises, universities, and research institutions, the strengths of all parties can be leveraged to achieve resource complementarity and mutual benefits. Specifically, universities and research institutions can provide rare earth enterprises with specialized research personnel and technical support, assist in overcoming practical challenges in technological innovation, enhance the relevance and effectiveness of innovations, transform patented technologies into market-competitive products, and foster a virtuous cycle that amplifies enterprises’ economic benefits.

7. Research Limitations

The research results of this study provide valuable insights for Chinese rare earth enterprises regarding R&D technology and development scale. The study focuses on the influence mechanism of AI as an emerging factor of production, helping enterprises to adjust production factors based on the study’s conclusions. It also encourages the more efficient application of patents and AI technologies across the entire industrial chain, improving the technological innovation efficiency of enterprises and providing a theoretical foundation for government policy formulation. Moreover, the study offers a conceptual framework for policy development.

Despite efforts to ensure data completeness and a reasonable model construction process, there are still limitations in the paper. The selection and processing of data have certain constraints. Due to the limited time span of the panel data, this study cannot fully capture the changes in the green efficiency of technological innovation in rare earth enterprises. As China and the global socio-economic and political environments evolve, rare earth resources will face increasingly diversified and complex challenges. Future research will explore the performance of Chinese rare earth enterprises in green technological innovation. We aim to expand the scope of research to include the green technological development of other economies and compare our findings to provide a more comprehensive description of the current status and development prospects of global rare earth enterprises. Further research is needed to help governments, scholars, and enterprises understand the factors influencing the green efficiency of innovation in rare earth mining and processing and identify measures to enhance innovation efficiency.