Multiple studies confirm that the development of the digital economy contributes to improved urban environmental quality and reduced pollution and carbon emission intensity. Quasi-experimental evidence from studies leveraging the “Broadband China” demonstration city policy as an exogenous shock shows that digital economy advancement significantly lowers PM2.5 concentrations and the Air Quality Index (Tian et al., 2024), while also reducing both the total volume and intensity of CO₂ emissions (Feng et al., 2023; Liu et al., 2023). This negative correlation between the digital economy and environmental pollution remains robust across broader empirical settings (Yuan et al., 2024). For instance, Zhang (2023), constructing a digital economy index via principal component analysis, demonstrates that it significantly reduces urban environmental pollution—proxied by industrial sulfur dioxide emissions—through mechanisms of human capital enhancement and industrial agglomeration.

Digitalization spurs green technological innovation, which in turn suppresses pollution and carbon emissions. This mechanism has been consistently validated within spatial econometric, threshold, and Difference-in-Differences mediation frameworks. The digital economy fosters green research and development through “innovation spillovers,” an enabling effect particularly salient in specific digital contexts such as smart cities (Yang et al., 2022).

Furthermore, digitalization drives the structural transformation from secondary to tertiary industries, promotes servitization of production, and facilitates industrial upgrading, thereby significantly enhancing environmental performance. Research by Liu et al. (2025) indicates that the digital economy substantially boosts energy efficiency through intermediary pathways like industrial structure optimization.

Digital technologies and data governance also enhance energy efficiency and promote the substitution of clean energy. Studies by Zhang et al. (2022), Shahbaz et al. (2022), and Xu et al. (2022a) show that digitalization considerably improves corporate energy savings and accelerates the energy transition. At the industrial level, technologies like the Industrial Internet and big data analytics significantly reduce energy consumption per unit of output (Huang et al., 2023).

Finally, digitalization improves environmental quality by enabling cross-regional resource allocation and strengthening government environmental governance—encompassing monitoring, enforcement, and institutional capacity. Huang and Lei (2021) posit that environmental regulation can stimulate advanced industrial structures and low-carbon energy consumption via innovation compensation and investment screening effects. In environmental governance, technologies like blockchain aid in tracking transaction-based carbon emissions, providing a technical foundation for precise regulation.

In the early stages of digitalization or under inadequate supporting conditions, the expansion of the digital economy may temporarily increase pollution or emissions. Threshold regression analyses identify a “rise-then-fall” pattern, where digital expansion below the threshold level is accompanied by rising emissions (Lan and Ma, 2025). This phenomenon can be attributed to the substantial energy and resource consumption inherent in digital economic activities. For example, Jiang et al. (2021) estimate that Bitcoin blockchain operations in China would generate massive carbon emissions. From a derived-effects perspective, consumption upgrading fostered by the digital economy - such as cross-border e-commerce - increases logistics and transportation volume, thereby driving up energy consumption.

Economic development levels and institutional/market environments serve as “amplifiers” for the pollution-reduction effects of the digital economy. At lower developmental or marketization levels, the environmental benefits of digital economy development are significantly weakened, with risks of negligible or even temporarily adverse impacts. As development levels improve, these pollution-reduction effects become substantially stronger (Yang et al., 2022).

Panel threshold models focusing on CO₂ emissions show that when the digital economy index remains below the threshold, digitalization progress correlates with increased emissions. This upward pressure can only be partially counteracted through technological advancement channels. Statistical evidence does not support the existence of a dual-threshold model (Sun et al., 2024).

During early development stages, multiple factors contribute to short-term environmental pressures: rising energy consumption from data centers and communication networks (Zhao et al., 2023), increased logistics intensity driven by e-commerce and instant delivery, and rebound effects in consumption and travel resulting from digitalization. As technological efficiency improves, energy structures transform, and governance capacity strengthens, these adverse effects are gradually offset and eventually converted into net environmental improvements. This explanation aligns with the segmented evidence presented by Lan and Ma (2025). Overall, the “inhibitory effect” predominantly emerges during low digital maturity stages or below established thresholds, but can be reversed through green technological progress and optimization of industrial and energy structures.

Early research on the relationship between the digital economy and environmental pollution primarily focused on whether its role was “promoting” or “inhibiting,” leading to seemingly contradictory conclusions. To reconcile these differences, the “nonlinear theory” has gradually emerged as the prevailing consensus. Regarding environmental pollution or carbon emissions, the panel threshold analysis consistently identifies a single threshold, indicating that the marginal environmental effects of the digital economy undergo structural changes across different developmental stages. When marketization crosses the threshold (e.g., the single marketization threshold reported by Yuan et al. (2024)), the pollution-reducing effects of the digital economy become more pronounced. Threshold tests for CO₂ emissions reveal a significant single threshold but no significant dual threshold (at the 10% level), suggesting a more likely “single inflection point” structure within the sample period (Sun et al., 2024). Empirical evidence also supports an inverted U-shaped relationship: digital expansion initially increases pollution but significantly suppresses it upon reaching maturity, accompanied by spatial spillovers and the activation of intermediary channels (Lan and Ma, 2025). Chen and Fu (2023) further revealed the complexity of this relationship in their study of the Yangtze River Economic Belt, identifying innovation levels as a key threshold variable influencing the digital economy’s environmental impact. Similarly, He et al. (2019) also identified a significant nonlinear threshold effect in their analysis of green credit, providing theoretical support for this framework.

The “segmented” effect observed in regional scale grouping aligns with threshold conclusions: higher-development groups are more likely to exhibit emission reductions and environmental improvements (Li and Yang, 2024). The digitalization process of neighboring cities alters the nonlinear pathway of local environmental performance, carrying implications for regional coordination policies (Huang et al., 2023). Recent studies have found that the digital economy exerts negative spillover effects on pollution intensity in neighboring regions (Yang et al., 2024), further enriching the spatial dimension of the relationship between the digital economy and the environment.

Chen et al. (2023) revealed the mediating mechanisms through which the digital economy influences green development via three pathways: “technological innovation,” “industrial structure optimization,” and “environmental regulation.” The effectiveness of these mechanisms may also be modulated by developmental stage thresholds. At low maturity stages, inhibitory mechanisms (e.g., infrastructure energy consumption) dominate; while after threshold crossing, facilitating mechanisms (e.g., industrial green transformation, targeted regulation) take the lead, thereby achieving an overall shift in environmental effects.

The aforementioned literature lays the foundation for this study, yet certain limitations remain. First, most studies employ single or composite indicators to measure the digital economy, failing to effectively separate and distinguish the heterogeneous impacts of “green” versus “non-green” digital technologies. This makes it difficult to answer the question of “what kind of digital economy should be developed” to more effectively drive green transformation. Second, extensive research relies on linear models for inference, struggling to capture the complex nonlinear realities of the digital economy’s environmental impacts. While a few studies employ threshold models, they often focus solely on macro-level threshold variables, failing to fully reveal the micro-level mechanisms driving causal inflection points. Third, existing mechanism tests primarily validate single transmission pathways, lacking systematic analysis of the dynamic interactions among multiple pathways—particularly conflicting ones—during developmental stages.

To address these gaps, this paper contributes as follows: By innovatively categorizing city-level digital economic development into green and non-green digital technologies, it establishes an empirical foundation for examining the impact of technological heterogeneity on the environment. Second, it integrates cutting-edge econometric and machine learning methods to capture complex nonlinear relationships. By combining panel threshold regression models with the XGBoost-SHAP machine learning framework, it systematically identifies threshold effects, nonlinear patterns, and key feature contributions in the environmental impacts of the digital economy, overcoming the limitations of linear assumptions. Third, it proposes and tests the “dual-path dynamic game” theoretical mechanism. This study integrates the Environmental Kuznets Curve (EKC) with the technological innovation life cycle theory, proposing that the digital economy influences environmental pollution through two dynamically interacting pathways: the “infrastructure energy consumption effect” (dominant in the early stage) and the “green innovation effect” (dominant in the later stage). This theory provides micro-mechanism support for phased policy design. The innovation of this integrated framework lies in its dual capability: not only does it predict the existence of nonlinear relationships, but it also identifies the core drivers of transition points—namely, the dynamic interaction between the “infrastructure energy consumption effect” and the “green innovation effect” at different developmental stages. This lays the theoretical foundation for subsequently distinguishing the heterogeneous impacts of technological attributes (H3) and regional contexts (H4).

This study employs a three-tier methodological framework integrating bidirectional fixed-effects models, panel threshold models, and machine learning approaches: ‘fundamental testing—structural identification—multifactorial analysis.’ The two-way fixed effects model is employed to preliminarily validate the nonlinear relationship between the digital economy and environmental pollution. The panel threshold model further identifies the structural thresholds where the impact mechanism shifts. Machine learning methods, from a data-driven perspective, reveal the importance of key variables, their nonlinear contributions, and sample heterogeneity. These three approaches mutually corroborate and complement each other, collectively enhancing the robustness and explanatory power of the research conclusions.

Regression analysis was conducted using Formula 1 to preliminarily explore the relationship between the digital economy and environmental pollution. Table 3 presents the regression results for the pooled OLS model, time fixed-effects model, individual fixed-effects model, and dual fixed-effects model (time and individual). The findings reveal that the inverted U-shaped relationship between the digital economy and environmental pollution is gradually becoming more pronounced and stable.

In the pooled OLS model, the coefficient for the squared term of the digital economy is significantly negative, while the linear term is insignificant, indicating an initial emergence of the inverted U-shaped relationship. After incorporating time fixed effects, the absolute value of the squared term coefficient increases to −7.8193, further strengthening the relationship. In the individual fixed effects model, the coefficient for the linear term of the digital economy was significantly positive, while the coefficient for the squared term was significantly negative, formally confirming the inverted U-shaped relationship. Finally, in the dual fixed effects model, both the linear and squared term coefficients were highly significant, making the inverted U-shaped relationship the most robust. The calculated inflection point value was approximately 0.243, indicating that the inhibitory effect of the digital economy on environmental pollution begins to manifest when the digital economy index exceeds 0.24.

Regarding control variables, significant differences emerged across models: the share of secondary industry was significantly positive in the mixed OLS and time-fixed effects models but became insignificant after controlling for individual effects, indicating its pollution effect primarily stems from inherent regional differences. R&D investment was significantly negative in the individual fixed effects model, suggesting technological innovation improves the environment in the long term. Environmental regulations were significantly positive under the individual fixed effects model, potentially due to the endogeneity of regulatory intensity. Foreign direct investment was significantly positive in the two-way fixed effects model, providing evidence for the “pollution haven” hypothesis.

Table 4 reports the results of the threshold effect tests. The single-threshold F-statistic was 121.22 with a p-value of 0.008, rejecting the “no threshold” hypothesis. The dual-threshold F-statistic was −82.23 with a p-value of 0.385, accepting the “no second threshold” hypothesis. This indicates the presence of a single threshold with an approximate value of 0.25.

Table 5 shows the regression results for the single threshold model. When DE ≤ 0.25, the coefficient for the digital economy is significantly positive (3.098) and significant at the 1% level. This indicates that in the early stages of digital economic development, its growth is primarily manifested through the rapid expansion of high-energy-consuming digital infrastructure (e.g., 5G base stations, data centers). The negative environmental effects generated by this segment dominate, thereby exacerbating environmental pollution. When DE > 0.25, the coefficient is significantly negative with a high absolute value and statistical significance. This indicates that beyond the threshold, the synergistic effects of digital technology integration with industries emerge—industrial internet optimizes production processes, AI monitoring enables pollution source tracing, and the green benefits outweigh the energy consumption costs, exerting a suppressing effect on environmental pollution.

Regression results on both sides of the threshold are shown in Table 5. Regarding control variables: At the high digital economy stage, while the secondary industry share coefficient remains positive, it has decreased. The foreign direct investment coefficient is significantly negative, and the environmental regulation coefficient is significantly positive. This indicates that combining a digital economy with stringent environmental regulations can effectively curb the “pollution haven” effect of foreign investment.

Economically, a one-unit increase in the digital economy index leads to a 3.098-unit rise in environmental pollution before the threshold and a 1.963-unit decrease after the threshold, resulting in a net emission reduction effect of 5.061 units.

To identify key factors influencing environmental pollution and validate the robustness of conclusions, this study employs the XGBoost machine learning model. To mitigate the risk of model overfitting, the dataset is randomly split into a 70% training set and a 30% testing set. The model optimization process combined grid search with cross-validation, ensuring model accuracy while strictly controlling overfitting. The optimal hyperparameter combination was determined as follows: n_estimators = 200, max_depth = 6, learning_rate = 0.1, subsample = 0.8, colsample_bytree = 0.8, random_state = 42. Comprehensive testing demonstrates that the XGBoost and SHAP model frameworks exhibit robust performance in feature selection, parameter tuning, and data variability, ensuring reliable interpretability. The final model achieves performance metrics of R ² = 0.655, RMSE = 0.322, and MAE = 0.246 on the test set, indicating strong simulation and explanatory capabilities for environmental pollution.

XGBoost was employed to extract feature contributions for environmental pollution interpretation, generating a feature importance map (Figure 1) to reveal each feature’s contribution to model predictions. According to the XGBoost feature importance map, factors influencing environmental pollution are ranked in descending order: DE, R&D, pgdp, Indus, FDI, and reg. DE is the most significant factor in urban ecological environments, exerting a major impact on environmental pollution levels.

To further explore the specific direction and magnitude of each feature’s impact on environmental pollution, SHAP values were calculated for each sample. Figure 2 illustrates the marginal contributions of influencing factors, with each point representing a distinct level. Blue and red sample points denote lower and higher true values of the influencing factors, respectively. The x-axis corresponds to the SHAP value of each sample, while the y-axis is sorted in descending order based on the feature’s average absolute SHAP value. Higher values indicate a greater influence of the feature on environmental pollution. Consequently, indicators with a positive impact on pollution exhibit a “blue left, red right” pattern, while those with a negative impact show a “red left, blue right” pattern. The contribution of individual features reveals that DE negatively affects pollution—higher DE values correlate with smaller SHAP values and corresponding pollution indices. Across the city, R&D and pgdp generally exert positive effects on environmental pollution, while Indus exhibits a potentially negative impact, reg and FDI demonstrate more complex patterns in their influence on environmental pollution.

SHAP value analysis reveals that DE ranks first among all influencing factors and exerts a significant negative impact on environmental pollution. This indicates that promoting digital economic development is a key strategy for improving environmental pollution.

To further explore the contribution patterns of DE and other control variables to environmental pollution, a SHAP dependence plot was generated (Figure 3). The x-axis represents feature values, while the y-axis displays the corresponding SHAP values for each feature. Each point represents the SHAP value of a specific feature in the data sample. The SHAP dependence plot reflects the specific contribution of individual variables to environmental pollution at different values.

The subplots in the figure illustrate the univariate contributions of DE, Indus, reg, R&D, FDI, and pgdp to environmental pollution. Taking the DE subplot as an example, the SHAP values exhibit complex trends as DE itself varies. Overall, when DE is at lower levels, SHAP values are predominantly negative, indicating an adverse impact on environmental pollution. As DE levels increase, SHAP values gradually shift from negative to positive, indicating a transition from negative to positive effects on environmental pollution. Beyond a certain threshold, SHAP values exhibit significant growth, suggesting that DE’s positive contribution to environmental pollution intensifies with its own development, revealing nonlinear and threshold characteristics.

In the Indus subplot, Indus’s SHAP values show an overall upward trend. As Indus increases, SHAP values gradually shift from negative to positive and continue to grow. This indicates that the impact of industrial development on environmental pollution transitions from suppression to promotion, with the promotional effect strengthening as industrial levels rise.

In the reg subplot, overall, when reg is at a low level, the SHAP value is negative, indicating a negative impact on environmental pollution. As the reg level increases, the SHAP value gradually rises, revealing a positive impact on environmental pollution.

In the R&D subplot, SHAP values exhibit significant fluctuations and nonlinear characteristics. At different R&D levels, SHAP values can be either positive or negative, indicating that R&D’s impact on environmental pollution is complex. Its contribution direction and magnitude vary across different intervals, with multiple inflection points reflecting the intricate nonlinear relationship between R&D and environmental pollution.

In the FDI subplot, the SHAP value of FDI evolves from negative to positive as its level increases. At lower FDI levels, it inhibits environmental pollution. As FDI rises, this inhibitory effect weakens and gradually shifts to a promotional role. Beyond a certain threshold, the positive contribution becomes more pronounced.

In the pgdp subplot, the SHAP value exhibits an overall nonlinear relationship. At different pgdp levels, SHAP values are both positive and negative, indicating that the impact of per capita GDP on environmental pollution varies across different levels. The direction of contribution changes with pgdp levels, featuring multiple inflection points that reveal a complex interaction pattern between per capita GDP and environmental pollution.

Overall, these variables exhibit gradual changes or multi-inflection point patterns in their nonlinear effects on environmental pollution. Substantial positive reinforcement occurs only when variables reach critical intensity levels, further highlighting the complexity and nonlinearity of these impacts.

The machine learning analysis in this section complements and cross-validates the econometric findings on several key points. First, the top ranking of the digital economy (DE) in the XGBoost feature importance analysis corroborates its statistically significant and substantial impact in the benchmark regression. Second, the SHAP dependency plots illustrate that DE’s marginal effect on environmental pollution shifts from negative to positive as its value increases, which aligns with the single threshold and the inverted U-shaped relationship identified by the panel threshold model. Furthermore, the heterogeneity in SHAP values across samples aligns with the econometric results on regional and technological grouping, jointly highlighting the context-dependent and heterogeneous environmental effects of the digital economy. This consistency across methodologies enhances the robustness of our conclusions and provides a more solid empirical foundation for policy deliberation. Building on this, the following section delves into the mediating mechanisms through which the digital economy influences environmental pollution.

The development of the digital economy initially promotes environmental pollution before subsequently inhibiting it. To further investigate this mechanism, we now conduct an in-depth analysis to identify the specific channels through which the digital economy influences environmental pollution. This study selects electricity consumption and green digital technology innovation as key mediating variables. The mechanism testing model is constructed in Formulas 5, 6: M e d i t = υ 0 + υ 1   · D E i t + υ 2   · D E i t 2 + μ i + λ t + ε i t (5) l n E P i t = δ 0 + δ 1   · D E i t + δ 2   · D E i t 2 + δ 3   · m e d i t + μ i + λ t + ε i t (6)

In Formulas 5, Med represents the mediating variables, including power consumption (pow) and green technological innovation (GP), while other variables remain consistent with Formula 1. If υ ₁ , υ ₂ , δ ₁ , δ ₂ , δ ₃ are statistically significant, it indicates that the mediating variables exert a mediating effect. This study employs a stepwise regression approach to address limitations in mediating effect statistical testing. The Bootstrap method is used to repeatedly sample 5,000 samples, enhancing statistical reliability.

The impact of the digital economy on environmental pollution transmission mechanisms primarily relies on dual restructuring of energy and technology. At the electricity consumption level, the explosive growth of the digital economy manifests directly as an exponential expansion of computing infrastructure. Such high-energy-consumption facilities inevitably drive up electricity usage, thereby exacerbating environmental pollution. However, from a long-term perspective, leveraging digital tools like smart grid dynamic scheduling and distributed energy management systems, the digital economy can transform this situation into a “digital emission reduction dividend,” mitigating environmental pollution by reducing electricity consumption. The intermediary role of green technological innovation manifests in two ways: first, the digital economy creates an iterative optimization environment for clean technology R&D; second, digital technologies themselves integrate directly into green production processes, ultimately achieving green manufacturing and reducing environmental pollution.

Based on the regression results in Table 12, the transmission mechanism of the digital economy’s impact on environmental pollution is analyzed as follows:

Column (1) shows that the coefficient of the quadratic term for the digital economy is significantly negative at the 1% level, indicating an inverted U-shaped nonlinear effect where the digital economy first promotes and then inhibits environmental pollution. Column (2) reveals the mediating effect of electricity consumption: the coefficients for DE and DE2 are positive and negative, respectively, at the 1% significance level, suggesting the digital economy also exerts a nonlinear impact on electricity consumption, initially increasing it before reducing it. Column (3) shows that after incorporating electricity consumption as a mediating variable, the coefficient for DE2 remains significantly negative, confirming the inverted U-shaped relationship between the digital economy and environmental pollution. Notably, compared to Column (1), the coefficient for DE has decreased, indicating that electricity consumption plays a nonlinear mediating role in the digital economy’s impact on environmental pollution—specifically, amplifying the effect initially before mitigating it later. Column (4) reveals that the primary term coefficient of the digital economy exhibits a significant negative impact on green technological innovation at the 1% level, while the secondary term coefficient is significantly positive, indicating a U-shaped pattern in the digital economy’s influence on green technological innovation. Column (5) results show that after incorporating green technological innovation as a mediating variable, the coefficient of DE2 remains significantly negative, further confirming the inverted U-shaped relationship between the digital economy and environmental pollution. More importantly, compared to the benchmark model (1), the absolute value of the digital economy coefficient significantly increases, verifying that green technological innovation plays a crucial mediating role in emission reduction within the digital economy’s impact on environmental pollution.

Test results based on the Bootstrap method (5,000 repeated samples) reinforce the statistical validity of the aforementioned mediating effect. Collectively, these findings demonstrate that both electricity consumption and green technological innovation play significant transmission roles in the digital economy’s impact on environmental pollution.

Mechanism testing reveals that the inverted U-shaped trajectory of “power consumption”—rising initially before declining—illustrates the digital economy’s transition from an “infrastructure energy consumption effect” (dominant in the early phase), centered on capital deepening and energy expenditure, to a “green innovation effect” (dominant in the later phase), focused on intelligent scheduling and energy efficiency optimization. Meanwhile, the U-shaped growth trajectory of “green technological innovation” reflects how the digital economy, after overcoming initial cost constraints, leverages its general-purpose technology attributes and digital platform efficiency to systematically reshape the supply-side conditions for green innovation activities. This shift moves beyond passively inducing demand to actively empowering transformation, thereby becoming the core force driving the emergence of pollution inflection points. The transition of the “green innovation effect” from latent to dominant hinges critically on whether the scale of digital infrastructure and the cost of green technologies can surpass specific thresholds. Once this critical point is crossed, the general-purpose technology attributes of the digital economy will systematically reduce the marginal cost of green innovation, driving a dynamic shift in the dominant mechanism from the “infrastructure energy consumption effect” to the “green innovation effect.” These two complementary pathways collectively reveal the deep-seated logic underpinning the digital economy’s environmental impact. Yang et al. (2024) and Li (2023) highlight industrial upgrading and green technological progress as key mechanisms through which the digital economy influences environmental pollution, aligning with our study’s “green digital innovation” mediation pathway. However, while the “green digital innovation” pathway explains most of the moderating effect, other unobserved mechanisms may coexist, for instance, while supporting environmental regulations promote green industrial upgrading, they may also trigger spatial and sectoral restructuring in the labor market (Li et al., 2025). The impact of environmental legislation can amplify the attractiveness of foreign direct investment in renewable energy sectors under production reduction policies, thereby generating more positive spillover effects for renewable energy technologies (Lai and Wang, 2024). This study identifies a dominant core mechanism, revealing the intrinsic dynamic game process through which the digital economy impacts the environment.