Revisiting the Nexus Between Energy Consumption, Economic Growth, and CO₂ Emissions in India and China: Insights from the Long Short-Term Memory (LSTM) Model

Abstract

Understanding how energy use and economic activity shape carbon emissions is pivotal for achieving global climate targets. This study quantifies the dynamic nexus between disaggregated energy consumption, economic growth, and CO₂ emissions in India and China—two economies that together account for more than one-third of global emissions. Using annual data from 1990 to 2021, we implement Long Short-Term Memory (LSTM) neural networks, which outperform traditional linear models in capturing nonlinearities and lagged effects. The dataset is split into training (1990–2013) and testing (2014–2021) intervals to ensure rigorous out-of-sample validation. Results reveal stark national differences. For India, coal, natural gas consumption, and economic growth are the strongest positive drivers of emissions, whereas renewable energy exerts a significant mitigating effect, and nuclear energy is negligible. In China, emissions are dominated by coal and petroleum use and by economic growth, while renewable and nuclear sources show weak, inconsistent impacts. We recommend retrofitting India’s coal- and gas-plants with carbon capture and storage, doubling clean-tech subsidies, and tripling annual solar-plus-storage auctions to displace fossil baseload. For China, priorities include ultra-supercritical upgrades with carbon capture, utilisation, and storage, green-bond-financed solar–wind buildouts, grid-scale storage deployments, and hydrogen-electric freight corridors. These data-driven pathways simultaneously cut flagship emitters, decouple GDP from carbon, provide replicable models for global net-zero research, and advance climate-resilient economic growth worldwide.

1. Introduction

Energy consumption has long been recognized as a central pillar of economic development and societal advancement. It is an essential input that powers industrial activity, fuels transportation, supports infrastructure, and enhances human well-being. This is particularly evident in rapidly growing economies such as India and China, which are central to the focus of this study. These two countries have emerged as dominant global players not only in economic output, but also in energy demand and environmental impact. Understanding their energy profiles is essential to grasp the broader dynamics of global sustainability. Disparities in energy usage have significantly influenced the development trajectories of countries and often contributed to widening the gap between developed and developing nations. Reliable and affordable energy access is a catalytic input for industrial production, job creation, and broader human development. Cross-country panel evidence for 51 low- and middle-income economies indicates that energy consumption and GDP are co-integrated in the long run, with bidirectional causality observed in middle-income economies [1]. This finding suggests that expanding the energy supply not only accommodates existing economic activity but also stimulates additional industrial output and employment, thereby creating a growth cycle. Sector-specific research from Saudi Arabia corroborates this mechanism [2]. Conversely, studies of net oil-importing African economies demonstrate that energy price shocks can suppress GDP per capita and destabilise consumption, underscoring how insecure or volatile energy supplies hinder industrial performance and household welfare [3]. Simultaneously, energy acts as a strategic geopolitical asset that has historically influenced war outcomes, trade patterns, and environmental challenges [4].

In recent decades, the global energy landscape has undergone rapid transformation. Worldwide energy demand has escalated dramatically, primarily driven by population growth, urbanization, and industrialization—especially in developing economies [5,6]. The majority of this increased demand has been met through the consumption of fossil fuels such as coal, oil, and natural gas. As a result, CO₂ emissions have increased significantly, contributing substantially to the intensification of climate change. For instance, global CO₂ emissions increased by approximately 30% from the nineteenth to the twentieth century [7]. The Intergovernmental Panel on Climate Change [8] notes that approximately 80% of global CO₂ emissions stem from human activities. Across diverse empirical settings—Croatia, the MINT economies, a multi-country panel of OECD nations, and Turkey—the literature consistently shows that rapid industrial expansion is correlated with rising environmental degradation, chiefly through higher CO₂ emissions. Ahmad et al. [9] and Ozturk & Acaravci [10] both document an inverted-U Environmental Kuznets Curve: emissions climb during early and middle stages of industrialisation, but begin to fall once income and technology reach a threshold that favours cleaner production. Adebayo et al. [11] reinforce the upward segment of this curve, finding that urban-centred industrial growth in emerging markets intensifies fossil-fuel use and ecological stress. Apergis et al. [12] add a causal dimension, showing bidirectional feedback between economic growth and emissions, while also demonstrating that the transition to nuclear and renewable energy can flatten or even reverse the emissions trajectory. The environmental consequences of this trend include more frequent and intense droughts, rising sea levels, glacial melt, habitat loss, and increased incidence of extreme weather phenomena.

The imperative to explore and understand the complex interplay between energy use, economic growth, and environmental degradation has intensified. This challenge has been globally acknowledged through multilateral agreements and institutional frameworks. The United Nations Sustainable Development Goals (SDGs) specifically call for urgent climate action (SDG 13) and the promotion of clean and affordable energy for all (SDG 7) [13,14]. These goals emphasize the need to transition toward sustainable energy systems that support economic growth while reducing environmental harm [15]. Policymakers are faced with the delicate balance of sustaining economic momentum without compromising long-term ecological stability.

This study is anchored in three complementary theoretical strands. First, the energy–growth causality literature—encompassing the growth, conservation, feedback, and neutrality hypotheses—posits that the direction and magnitude of causality between energy consumption and economic growth vary by development stage and energy mix, affecting both policy priorities and emissions outcomes. Second, the Environmental Kuznets Curve (EKC) framework suggests an inverted-U relationship between economic growth and environmental degradation, implying that emissions initially rise with income but may decline once a turning point in income level and the adoption of cleaner technologies are reached. Third, path-dependence and non-stationarity theories highlight how lock-ins to specific fuel types, infrastructure stocks, and policy regimes create time-varying, non-linear dynamics that conventional linear models often fail to capture. Integrating these perspectives, our LSTM approach is designed to uncover country-specific lead–lag patterns among disaggregated energy sources, GDP, and CO₂ emissions, thereby testing whether India and China exhibit distinct positions on the EKC and whether their energy–growth linkages conform to, or deviate from, established causality hypotheses.

Analysing the nexus between energy consumption, economic growth, and CO₂ emissions in the Indian and Chinese contexts is critical, because these two nations are not only among the world’s top energy consumers and largest CO₂ emitters, but also key influencers of global climate trajectories. Their rapid industrialization and urbanization have significantly altered their energy and emissions profiles, positioning them as pivotal case studies for examining the challenges and opportunities of sustainable development. Moreover, both countries are actively engaged in ambitious policy experimentation aimed at transitioning toward low-carbon economies, making them important laboratories for observing the effectiveness of renewable energy promotion, energy efficiency strategies, and emissions reduction commitments. As signatories to major international climate agreements, their ability to achieve or fall short of emissions targets carries substantial global consequences. Understanding their unique trajectories provides valuable insights into how large, rapidly developing economies can balance growth with environmental responsibility and contributes to the design of more effective and scalable climate policies worldwide.

Despite the growing volume of literature on this topic, existing studies predominantly rely on traditional econometric tools such as Granger causality tests, vector error correction models (VECM), and cointegration techniques. These models, while useful, often rely on assumptions such as linearity, stationarity, and symmetric relationships that may not reflect real-world complexities. They may also struggle to accurately forecast outcomes in the presence of non-linear interactions and long-range dependencies. To address these shortcomings, this study leverages advances in artificial intelligence and data science, employing the Long Short-Term Memory (LSTM) model—a class of recurrent neural networks particularly suited to analysing time-series data with memory components. The LSTM framework enables the modelling of non-linear, non-stationary, and high-dimensional relationships, offering superior performance in forecasting and interpretability.

This study contributes to the literature on the energy-growth-emissions nexus in several important ways. First, it focuses on India and China—two of the world’s largest emerging economies and most significant CO₂ emitters—thereby offering critical insights into global sustainability challenges from a developing country perspective. Second, it provides a disaggregated analysis of energy consumption by distinguishing between renewable, non-renewable, and nuclear energy sources, allowing for a more nuanced understanding of their respective environmental impacts. Third, the study utilizes a Long Short-Term Memory neural network model, a form of deep learning particularly well-suited to modelling temporal and non-linear dynamics. This methodological innovation addresses the limitations of conventional econometric models and enhances predictive accuracy. Finally, the research offers data-driven evidence to support policymaking by identifying the most influential energy types and economic factors driving CO₂ emissions, ultimately informing sustainable development strategies aligned with national and international climate goals.

Building on these motivations, the primary objective of this study is to quantify and compare the long-run and short-run impacts of disaggregated energy consumption and economic growth on CO₂ emissions in India and China. Based on the literature review and preliminary diagnostics, we articulate three testable hypotheses: H1 (India): Coal and natural-gas consumption exert the strongest positive effect on CO₂ emissions, while renewable energy exerts a mitigating effect; H2 (China): Coal and petroleum consumption dominate the emissions trajectory, with renewable and nuclear energy exhibiting weaker or inconsistent impacts; H3 (Comparative Dynamics): The predictive performance and variable importance differ significantly between India and China, reflecting country-specific energy structures and policy regimes.

By leveraging an LSTM framework, we aim to deliver more accurate results and richer causal insights than those attainable with traditional linear econometric models. To achieve this goal and test hypotheses, we compile an annual panel spanning 1990–2021 that includes CO₂ emissions, real GDP, and five energy-consumption indicators—coal, natural gas, petroleum, renewable, and nuclear —sourced from the U.S. Energy Information Administration (EIA) and the World Development Indicators (WDI). The data period is chosen based on the latest available data for all variables used in our analysis. This balanced dataset enables a consistent comparative analysis of the two countries over three decades of rapid economic transformation. For modelling purposes, the Indian and Chinese series were divided into a training set covering the period 1990–2013 (24 observations) and a testing set spanning the period 2014–2021 (8 observations).

The remainder of this paper is structured as follows: Section 2 reviews the relevant literature and theoretical underpinnings; Section 3 outlines the research methodology, including model specifications, performance metrics, and data preprocessing techniques; Section 4 presents the dataset and variables; Section 5 discusses the empirical findings and their implications; and Section 6 offers conclusions and policy recommendations aimed at guiding sustainable development in India and China.

2. Literature Review

2.1. Economic Growth, Energy Demand, and Emissions: A Dual Challenge for Emerging Economies

There is broad agreement that economic expansion is frequently associated with greater energy consumption, which in turn contributes to higher levels of CO₂ emissions. However, the precise nature of this relationship varies by country, economic structure, and energy mix. While clean energy sources such as wind, solar, and hydropower are widely recognized as part of the solution, their scalability and integration into national energy grids remain a challenge [16]. Meanwhile, international commitments such as the Kyoto Protocol and the Paris Agreement have established ambitious decarbonization targets: most industrialized countries have pledged carbon neutrality by 2050, while China aims for 2060, and India by 2070 [17,18]. Meeting these targets requires a deeper understanding of how energy and growth dynamics intersect with emissions in the real world.

India and China are at the heart of this global sustainability effort. In 2023, India surpassed China in population, becoming the world’s most populous nation with 1.44 billion people, and simultaneously ranked as the third-largest global energy consumer (EIA). China remains the largest energy producer and consumer, with its energy policies heavily influencing global emissions trends. Together, the two countries accounted for about 33.6% of global CO₂ emissions in 2022 [19]. Notably, China’s per capita emissions have surpassed those of many advanced economies, while India’s remain less than half the world average, indicating divergent trajectories that merit comparative analysis.

Both nations are navigating a delicate path between economic development and environmental stewardship. India has launched several landmark initiatives to encourage renewable energy adoption and electric mobility, including the National Solar Mission and the Faster Adoption and Manufacturing of Hybrid and Electric Vehicles scheme. China has likewise implemented the “Dual Carbon” strategy, with goals of peaking carbon emissions by 2030 and achieving carbon neutrality by 2060. These policies are backed by massive investments in green infrastructure, technological innovation, and regulatory reforms. However, curtailing energy consumption can pose risks to industrial productivity and employment, which makes it imperative to understand the long-term trade-offs and synergies among carbon emissions, energy consumption, and economic growth.

Asia, home to 60% of the global population, remains the world’s largest emitter of greenhouse gases. It accounted for 53% of global emissions in 2022, with China alone responsible for over 58% of Asia’s CO₂ emissions [8]. Across Asia, greenhouse gas emissions stem from two interlinked drivers: rapid energy-hungry economic growth and energy-intensive industrial processes. Panel evidence for South Asia (India, Pakistan, Bangladesh, Sri Lanka, Nepal) shows that a 1% rise in per-capita energy use boosts per-capita GDP by roughly 0.84%, revealing a bidirectional, long-run feedback between energy consumption and growth [20]. Because the region still relies mainly on coal and oil, this growth-energy linkage translates directly into higher CO₂ output. In China—the continent’s largest emitter—industrial process emissions (from cement, lime, ammonia, methanol, steel, aluminium, etc.) have surged and now account for 17–19% of national energy-related GHG releases [21]. Asia’s emissions profile is dominated first by fossil-fuel combustion to power expanding economies and, increasingly, by process-related releases from heavy industry, especially in China’s materials and chemicals sectors.

Additionally, climate change is intensifying extreme-weather hazards across South and East Asia, with densely populated regions bearing the brunt. Aryal et al. [22] report a marked rise in erratic rainfall, flash floods, windstorms, droughts, and landslides that regularly disrupt smallholder farming systems throughout the Indo-Gangetic Plain and Himalayan foothills—areas where livelihoods and food security are especially climate-sensitive. Marcotullio et al. [23] add an urban perspective: Asia’s mega-cities, many situated on low-lying deltas, now face more frequent flood episodes and heat-wave conditions, while peri-urban fringes experience heightened wildfire risk as higher temperatures dry surrounding vegetation. Together, these studies underscore how warming temperatures and shifting precipitation patterns, combined with rapid urbanisation and high population density, are driving a surge in climate-related disasters across vulnerable parts of South and East Asia.

2.2. Key Findings from Previous Studies

The inter-relationship among climate change, energy consumption (EC), and sustainable economic growth (EG) has become one of the most vigorously debated topics in energy economics and environmental policy. Since Kraft and Kraft’s [24] seminal contribution, substantial empirical literature—spanning engineering, economics, and interdisciplinary outlets—has examined the EC–EG nexus. Despite this prolific output, consensus remains elusive. Results diverge because authors adopt contrasting econometric strategies (single-equation time-series models, heterogeneous-panel estimators, structural-VAR frameworks, or increasingly, machine-learning pipelines), access dissimilar data vintages and emission inventories, and focus on countries occupying very different points along the development spectrum, for example, Lu et al. [25] or Saidi et al. [26]. Time-series studies, by design, foreground idiosyncratic national dynamics—capturing, for instance, country-specific policy shocks or fuel-mix transitions—whereas multi-country panels emphasise cross-sectional contrasts. The two approaches, therefore, tend to deliver different verdicts on both the direction and the strength of causal linkages between EC and EG.

To clarify the debate and guide the subsequent synthesis, scholars usually interpret findings through four rival hypotheses: feedback (EC and EG reinforce each other), conservation (growth drives energy demand), growth (energy use propels output), and neutrality (no significant causal link). Empirical support is mixed—bidirectional feedback dominates in many emerging markets, the conservation view prevails in high-income OECD economies, growth effects surface in fuel-exporting or rapidly industrialising states, and neutrality occasionally appears in service-oriented countries with high renewable penetration—underscoring that causality hinges on the energy mix, institutional quality, and stage of development.

A second robust insight concerns the moderating role of renewable energy (RE). Analyses that employ non-linear autoregressive distributed-lag (ARDL) models, frequency-domain causality tests, and wavelet-based decompositions reach a common conclusion: as the share of renewable energy (RE) rises, the traditional EC–EG–CO₂ linkage weakens. The driver is a systematic fall in the carbon intensity of each kilowatt-hour produced [17,27]. Advanced decomposition analysis shows that in economies where the RE share exceeds approximately 30%, the long-run elasticity of CO₂ with respect to EC drops by half relative to fossil-fuel-dependent peers. Conversely, fossil-fuel-locked exporters continue to display strong EC–CO₂ elasticities, highlighting the urgency of diversifying their energy portfolios. Nuclear power, although controversial, is identified as a potential large-scale, low-carbon growth engine in India and selected OECD members [28], while green hydrogen pilots in Chile and Saudi Arabia promise to blur the traditional boundary between producer and consumer nations. While the fuel mix is pivotal, geography is equally crucial. The following subsection explores regional heterogeneity and contextual nuance.

Regional patterns provide further nuance for the global picture. In Latin America, Altinoz et al. [29] report that both EC and CO₂ curb growth, corroborating evidence that commodity-export volatility and under-investment in modern energy infrastructure can make energy spending less productive. Muhammad [30], however, documents an energy-driven yet emission-intensive expansion cycle across a wider set of Latin American and Caribbean economies, underscoring the region’s structural diversity. In Sub-Saharan Africa, Gershon et al. [31] find that higher EC boosts output, but its effect on emissions hinges on renewable penetration and governance quality—a reminder that institutional capacity can magnify or mute the environmental footprint of energy use. In South Asia, Rehman and Rehman [32] confirm EC as a primary emissions driver, while wavelet-based analysis suggests India’s GDP–EC link is partly offset by efficiency gains and rapid solar deployment [33]. Meanwhile, for post-Soviet transition economies, Chen et al. [34] show heterogeneous responses across the income distribution, with coal-rich regions locked in high-carbon paths and gas-oriented areas benefiting from cleaner fuel substitution.

Beyond the geographic context, methodological design also shapes empirical conclusions. The conclusions scholars draw depend not only on geography, but also on methodological choices. Models that allow for structural breaks, asymmetric adjustments, or thick-tailed shock distributions—such as Threshold-ARDL, Quantile-on-Quantile regressions, or Bayesian-VARs—often paint a more nuanced picture than standard linear cointegration tests. Recent meta-analysis indicates that studies incorporating at least one form of non-linearity are 35% less likely to reject the conservation hypothesis, reflecting the fact that energy intensity tends to decline at higher income levels.

Table 1. Comparative evidence on the energy–growth–emissions nexus.

Study	Sample	Method	Key Finding
Radmehr et al. [35]	EU, 1995–2014	P-SSE	EG ↔ CO2; REN → CO2 (–)
Alam & Hossain [17]	CHN, 1990–2019	ARDL/ARCH-LM/BG-LM	REN → CO2 (–)
Rahman et al. [36]	CHN, 1985–2021	Wavelet Coherence Analysis	EC from fossil fuels ↑ CO2;
Agboola et al. [37]	SAU, 1971–2016	MWT (T-Y)	EC → CO2; 1% ΔGDP ≈ 1% ΔCO2
Namahoro et al. [27]	41 WIND, 1997–2018	CS-DL/CS-ARDL/CCE-P	WIND ↑ EG; WIND → CO2 (–)
Ozgur et al. [28]	IND, 1970–2016	Fourier ARDL	NUC ↑ clean EG
Rehman & Rehman [32]	CHN+4, 2001–2014	GRA/TOPSIS	EC major driver of CO2
Eldowma et al. [38]	SDN, 1971–2019	ARDL	CO2 → EG → Electricity ↑
Wen et al. [39]	SA, 1985–2018	FMOLS	NRE → Pollution ↑
Rahman et al. [40,41]	NICs, 1979–2017	CI/DOLS/FMOLS/PMG	EC & EXP ↑ ENV deg.
Gershon et al. [31]	17 AFR, 2000–2017	Static Panel	EC → CO2 (–); EC → EG (+)
Khan et al. [42]	PAK, 1965–2015	ARDL	EC & EG → CO2 (+)
Chen et al. [34]	6 TE, 1970–2021	QQ	EC → CO2 (+); EC → EG (+)
Pradhan et al. [43]	G7+SA, 1996–2021	Sim-Reg/Panel ARDL	EC → EG; CO2 → EG
Salari et al. [44]	USA, 1997–2016	Static & Dyn panel	REN → CO2 (–); NRE → CO2 (+)
Afjal [45]	37 OECD, 1995–2020	PVAR	GDP ↛ CO2 (neutral)
Liu et al. [46]	46 BRI countries, 2005–2018	Driscoll–Kraay Est.	REN → CO2 (–); EKC supported
Shah et al. [47]	49 green bond countries, 2007–2019	Simultaneous Equation Model	fossil-fuel-driven EG ↑ GHG emissions;

Notes: ↔ = bidirectional causality; → = unidirectional (Granger) causality; (+)/(–) indicate positive/negative effects; ↛ indicates neutral; REN = renewable energy; NRE = non-renewable energy; WIND = wind energy; ENV = environment; CI = cointegration; P-SSE = panel spatial simultaneous equations; MWT (T-Y) = Modified Wald test (Toda–Yamamoto); CS-DL = cross-sectional distributed lag; CS-ARDL = cross-section-augmented ARDL; CCE-P = common correlated effects–pooled; GRA = grey relation analysis; TOPSIS = technique for order preference by similarity to ideal solution; FMOLS = fully modified OLS; DOLS = dynamic OLS; PMG = pooled mean group; QQ = quantile-on-quantile; PVAR = panel vector autoregression).

condenses fifteen representative studies published between 2021 and 2025, and highlights a key takeaway: cross-country differences in renewable energy penetration and methodological sophistication largely account for the divergent causal patterns observed. They span five continents, deploy a wide array of estimation strategies, and reach dissimilar (sometimes conflicting) conclusions. Collectively, they illustrate how methodological choices and country circumstances shape the observed EC–EG–CO₂ relationships. Notably, the balance of recent evidence tilts toward feedback or RE-mediated decoupling, although classic growth-type findings persist in energy-constrained, fossil-fuel-intensive settings.

The foregoing review suggests four broad stylised facts. First, the direction of causality between energy consumption and growth is not immutable, but evolves alongside structural change, shifts in the fuel mix, and policy interventions. Second, economies that expand the share of renewables—or, where socially acceptable, low-carbon nuclear power—consistently weaken the elasticity of emissions with respect to energy use, pointing to a viable route for green growth. Third, governance quality and institutional capacity are decisive: the same increment in energy supply can yield either sustainable or unsustainable outcomes depending on how effectively governments channel investment and enforce environmental standards. Finally, empirical verdicts are highly sensitive to methodological flexibility; models that allow for asymmetry, thresholds, or distributional heterogeneity tend to produce more conservative estimates of energy dependence, underscoring the need for nuanced, context-specific policy design. Together, these insights caution against one-size-fits-all prescriptions and highlight the importance of tailoring policy toolkits to national endowments, institutional strength, and the maturity of domestic energy markets.

Based on this literature review, it appears that two critical knowledge gaps still hinder a comprehensive understanding of the energy–growth–emissions nexus. First, most India- and China-focused studies aggregate “total energy,” leaving the distinct influences of coal, petroleum, natural gas, renewables, and nuclear power largely unexplored within a unified GDP–CO₂ framework. Second, the prevailing reliance on linear ARDL, VECM, and cointegration models limits the ability to capture regime shifts, structural breaks, and higher-order interactions that characterize rapidly evolving, policy-active economies. Because the Long Short-Term Memory networks belong to the recurrent neural family and are expressly designed for sequential data, and their gating architecture allows the model to retain or discard information over time, this makes them well suited to capture delayed energy–growth interactions, structural breaks such as the 2008 financial crisis or the post COVID 19 energy shock, and the long range dependence ubiquitous in environmental and macroeconomic series. Crucially, LSTMs learn complex, nonlinear functional forms endogenously rather than imposing a priori restrictions, and they scale smoothly to multivariate settings that include disaggregated fuel types, technology indices, and policy dummies. Therefore, addressing these gaps necessitates a non-linear, explainable, and disaggregated modelling strategy—such as the LSTM framework employed in this study—to generate policy-relevant insights for both India and China.

Performance benchmarks confirm these theoretical advantages. Benchmark studies comparing LSTMs with ARIMA and random-forest baselines report forecasting-error reductions of roughly 25% and markedly better turning-point detection—improvements that arise because classic linear models struggle with non-stationarity, regime shifts, and complex feedback loops [48]. By providing accurate out-of-sample predictions and allowing scenario analysis that is difficult to implement in traditional frameworks, LSTMs offer a versatile addition to the EC–EG–CO₂ toolkit—one capable of illuminating the path-dependent, non-stationary dynamics that govern the energy–growth–environment triad.

In sum, the contemporary literature presents a highly context-dependent and dynamically evolving EC–EG–CO₂ nexus. Progress toward the Paris targets and the Sustainable Development Goals will hinge on swift, context-aware energy transitions coupled with data-driven modelling frameworks—such as LSTMs—that can faithfully track and anticipate the complex interplay among energy demand, economic prosperity, and environmental quality. Against this backdrop, the present article sets out to develop an LSTM-based model for India and China (1990–2024) with the explicit goal of testing whether rising renewable penetration has already begun to decouple growth from emissions and of generating forward-looking scenarios to inform policy design.

3. Methods and Models

This section outlines the methodological framework employed to examine the dynamic relationships between energy consumption, economic growth, and CO₂ emissions in India and China. Given the complex and time-dependent nature of these variables, we adopted a data-driven modelling approach utilizing Long Short-Term Memory neural networks [49]. LSTM models are particularly well-suited for time-series analysis due to their ability to capture long-range dependencies and non-linear interactions. Compared to traditional econometric frameworks, such as cointegration, VECM, and ARDL—which rely on linear adjustment toward a single long-run equilibrium and are prone to instability during structural breaks—LSTM models offer substantial advantages. Specifically, the gating mechanism of LSTM networks can seamlessly capture nonlinearities and accommodate both regime shifts and higher-order interactions without requiring predefined functional forms. Moreover, LSTM’s flexible memory architecture enables the endogenous learning of variable-specific lag structures, effectively handling the path-dependent dynamics often observed among energy consumption, economic growth, and CO₂ emissions. In contrast, traditional models depend on ex-ante lag selection, which can limit their accuracy and efficiency as lag complexity increases. This makes LSTM a powerful tool for exploring complex, time-varying relationships in economic and environmental systems.

In recent years, LSTM models have gained increasing popularity across the fields of economics, finance, and environmental science due to their ability to capture complex temporal dependencies in sequential data effectively. For instance, in the environmental domain, LSTM has been extensively applied to forecast air pollution indicators such as PM2.5 concentrations. Nourmohammad and Rashidi [50] compared LSTM with ARIMA and XGBoost models to predict daily and monthly PM2.5 levels in Tehran. While XGBoost achieved the highest accuracy for daily forecasts, LSTM demonstrated stable performance across various input configurations, underscoring its flexibility in handling multivariate environmental datasets. Similarly, Waqas et al. [51] evaluated six predictive models and ranked LSTM as the second-best performing deep learning algorithm for forecasting PM2.5 in Islamabad, Pakistan—outperforming traditional machine learning approaches during the testing phase. Moreover, Noynoo et al. [52] integrated LSTM into a hybrid forecasting framework with the WRF-Chem model to enhance the accuracy of PM2.5 predictions in southern Thailand. Their hybrid LSTM-based model significantly improved forecasting metrics and demonstrated strong predictive power up to 72 h in advance.

LSTM models are increasingly applied in economics and finance for forecasting complex, nonlinear financial time series. Sun [53] used a Bayesian-optimized LSTM to predict stock prices in China’s major indices, demonstrating superior accuracy over traditional models. Peng et al. [54] developed a hybrid model combining empirical mode decomposition and an attention-enhanced LSTM, improving predictive accuracy and reducing error. Their results confirm that LSTM networks, especially when integrated with advanced techniques, effectively capture dynamic financial patterns. These applications underscore LSTM’s value in enhancing the reliability of stock market predictions and supporting data-driven financial decision-making.

LSTM models are increasingly used across finance, economics, and environmental science to forecast complex, time-dependent phenomena. Jiang et al. [55] applied a VMD-LSTM model to assess the impact of climate risk on China’s renewable energy market, finding that incorporating climate uncertainty indices significantly enhanced forecasting accuracy across various time horizons. In the energy sector, Lu et al. [56] examined the effects of electricity policy uncertainty and carbon emission prices on electricity demand in China. While mixed-frequency models outperformed LSTM, the LSTM model still captured key nonlinear patterns. In environmental modelling, Liu et al. [57] developed a high-resolution forecast of emissions in China’s cement industry through 2035. Their findings showed that fuel and clinker substitution could significantly reduce SO₂ and CO₂ emissions, with notable co-benefits for PM2.5 and NOx.

The LSTM methodology comprises four core components: model evaluation metrics, LSTM network architecture, data preprocessing techniques, and model design strategy.

3.1. Evaluation Metrics

To evaluate the predictive performance of the Long Short-Term Memory (LSTM) model, we employed three standard error metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) [49]. These metrics offer a comprehensive assessment of forecast accuracy by quantifying the average error and the magnitude of larger deviations between predicted and actual values. MAE captures the average magnitude of errors without considering their direction, MSE penalises larger deviations more heavily due to the squaring function, and RMSE provides a normalised measure of error in the same units as the target variable, making it more interpretable.

The respective formulas are given below:

$$MAE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|y_i-{\widehat{y}}_i\right|,$$

(1)

$$MSE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2,$$

(2)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}.$$

(3)

where y denotes the observed value, ${\widehat{y}}_i$ is the predicted value, and $n$ is the total number of observations. Together, these metrics help evaluate model performance from both an average error and dispersion perspective.

3.2. Subsection Long Short-Term Memory (LSTM) Model

The LSTM network, introduced by Hochreiter and Schmidhuber [49], was developed to address the limitations of standard recurrent neural networks (RNNs), particularly their inability to learn long-term dependencies due to the vanishing gradient problem. LSTMs incorporate internal memory cells and gate mechanisms that allow them to retain, update, or discard information over extended sequences. Each LSTM cell contains an internal memory cell state $C_t$ and a set of gates that regulate the flow of information into, within, and out of this cell: $i_t$—input gate (determines which new information should be stored in the current cell state); $f_t$—forget gate (decides what information from the previous cell state should be removed); $o_t$—output gate (controls what part of the cell state is passed to the next time step); and $g_t$—change gate (called the candidate cell state, and helps the LSTM decide what new information to store in memory). These gates are themselves small neural networks trained jointly with the rest of the model, which means the network can learn when to remember, when to forget, and what to output.

The computational steps are as follows:

$$i_i=sigmoid\left(W_{xi}·x_t+W_{hi}·h_{t-1}+b_i\right),$$

(4)

$$f_t=sigmoid\left(W_{xf}·x_t+W_{hf}·h_{t-1}+b_f\right),$$

(5)

$$o_t=sigmoid\left(W_{xo}·x_t+W_{ho}·h_{t-1}+b_o\right),$$

(6)

$$g_t=tanh\left(W_{xg}·x_t+W_{hg}·h_{t-1}+b_g\right),$$

(7)

$$c_t=\left(f_t·c_{t-1}\right)+\left(i_t·g_t\right),$$

(8)

$${\widehat{y}}_t=h_t=o_t·tanh\left(c_t\right).$$

(9)

where $sigmoid$ is the sigmoid activation function, $tanh$ is the hyperbolic tangent function, $W$ represents weight matrices, and $b$ represents bias terms.

3.3. Data Processing and Model Design

This study uses time-series data that includes CO₂ emissions, real GDP, and disaggregated energy consumption data (e.g., coal, natural gas, petroleum, renewable, nuclear). CO₂ emissions are measured in metric tons per capita, real GDP is expressed in constant 2015 USD, and energy consumption is reported in quadrillion British thermal units (BTUs). The data were sourced from two authoritative open-access databases: the U.S. Energy Information Administration (EIA) and the World Bank’s World Development Indicators (WDI).

To maintain the integrity of the time-series structure, the dataset was divided into an 80:20 training-to-testing ratio. This straightforward split was preferred over k-fold cross-validation, which may disrupt temporal continuity. All variables were normalised using Min-Max scaling to a [0, 1] range to improve convergence during training and ensure comparability across features:

$$x^{\prime }=x-min\left(x\right)/\mathit{max}\left(x\right)-min\left(x\right)$$

(10)

where $x$ is the original value, and $min\left(x\right)$, $max\left(x\right)$ are the minimum and maximum values of the variable, respectively.

To explore the relationship between energy consumption, economic growth, and CO₂ emissions, we developed a sequential set of LSTM model configurations with increasing levels of complexity and explanatory power. The first model included only lagged values of CO₂ emissions to serve as a baseline for prediction. Each model was estimated using lag lengths of 1, 2, and 3 years, allowing us to assess both the immediate and delayed effects of the predictors. Seven LSTM model configurations were constructed as follows:

$$\mathrm{Model} 1. {CO}_{2t}=f\left({CO}_{2t-\mathcal{l}}\right),$$

$$\mathrm{Model} 2. {CO}_{2t}=f\left({CO}_{2t-\mathcal{l}}{, CC}_{t-\mathcal{l}}\right),$$

$$\mathrm{Model} 3. {CO}_{2t}=f\left({CO}_{2t-\mathcal{l}}{, CC}_{t-\mathcal{l}}, {NG}_{t-\mathcal{l}}\right),$$

$$\mathrm{Model} 4. {CO}_{2t}=f\left({CO}_{2t-\mathcal{l}}{, CC}_{t-\mathcal{l}}, {NG}_{t-\mathcal{l}}, {PC}_{t-\mathcal{l}}\right),$$

$$\mathrm{Model} 5. {CO}_{2t}=f\left({CO}_{2t-\mathcal{l}}{, CC}_{t-\mathcal{l}}, {NG}_{t-\mathcal{l}}, {PC}_{t-\mathcal{l}}, {RC}_{t-\mathcal{l}}\right),$$

$$\mathrm{Model} 6. {CO}_{2t}=f\left({CO}_{2t-\mathcal{l}}{, CC}_{t-\mathcal{l}}, {NG}_{t-\mathcal{l}}, {PC}_{t-\mathcal{l}}, {RC}_{t-\mathcal{l}}, {NEC}_{t-\mathcal{l}}\right),$$

$$\mathrm{Model} 7. {CO}_{2t}=f\left({CO}_{2t-\mathcal{l}}{, CC}_{t-\mathcal{l}}, {NG}_{t-\mathcal{l}}, {PC}_{t-\mathcal{l}}, {RC}_{t-\mathcal{l}}, {NEC}_{t-\mathcal{l}}, {RGDP}_{t-\mathcal{l}}\right).$$

where ℓ∈{1,2,3} denotes the lag length applied to each input variable.

Model accuracy was compared using MAE, MSE, and RMSE, and the results were interpreted in relation to policy implications. Overall, this methodology enables the detection of both linear and nonlinear patterns in the data, providing rigorous empirical insights into how energy use and economic activity influence environmental outcomes.

4. Data Description and Variable Specification

This section outlines the key variables used to examine the relationship between energy consumption, economic growth, and CO₂ emissions in India and China. Specifically, the selection of variables is grounded in existing theoretical frameworks and supported by an extensive body of empirical literature. To begin with, fossil fuels—such as coal, oil, and natural gas—are widely acknowledged as major contributors to CO₂ emissions due to their carbon-intensive combustion. For example, empirical evidence from Amin [58] and Bibi [59] shows that the rising demand for coal, oil, and natural gas—due to their carbon-intensive combustion—directly escalates national CO₂ emissions. Likewise, analyses by Bölük [60] and Shayanmehr [61] confirm that fossil-fuel energy sources emit far more greenhouse gases than renewable alternatives, underscoring their central role in global carbon-release patterns. Indeed, the European Commission’s Joint Research Centre estimates that approximately 90% of global CO₂ emissions stem from fossil fuel use [62]. Nevertheless, while fossil fuels remain essential to economic growth and industrial development, their environmental costs are substantial, thereby underscoring the need for sustainable alternatives [12].

Conversely, renewable energy has emerged as a vital solution for reducing global CO₂ emissions. Notably, it is widely recognised for its environmental benefits and economic advantages [63,64]. Furthermore, annual consumption of renewable sources is growing rapidly, and these resources are increasingly positioned as key tools for balancing economic growth with environmental sustainability [65]. In contrast, nuclear energy—though sometimes contentious—offers a low-carbon option for electricity generation because it does not emit CO₂ during operation. Empirical evidence indicates that increased nuclear deployment can substantially lower the carbon intensity of the power sector, particularly in countries heavily dependent on fossil fuels.

With respect to economic factors, the study captures growth through real GDP. Typically, economic expansion leads to higher emissions due to increased demand for energy, transportation, and industrial output; however, many countries can decouple this link through the adoption of cleaner technologies [66] and improved efficiency [67]. Accordingly, this study integrates both energy and economic indicators to provide a holistic analysis. Each variable is defined, measured using standardised units, and sourced from internationally recognised databases. As summarised in

Table 2. Variable specifications for LSTM model.

Variable	Symbol	Unit	Expected Sign	Source
CO2 Emissions	CO2	Metric tons per capita	-	WDI
Coal Consumption	CC	Quadrillion BTUs	Positive	EIA
Natural Gas Consumption	NG	Quadrillion BTUs	Positive	EIA
Petroleum Consumption	PC	Quadrillion BTUs	Positive	EIA
Renewable Energy Consumption	RC	Quadrillion BTUs	Negative	EIA
Nuclear Energy Consumption	NEC	Quadrillion BTUs	Negative	EIA
Real GDP	RGDP	Constant 2015 USD	Positive/Negative	WDI

Source: Compiled by the authors using data from the U.S. Energy Information Administration (EIA) and World Development Indicators (WDI).

, the variables and their expected directional impact on emissions are presented for reference.

The period 1990–2021 captures the post-1991 economic liberalisation in India, China’s WTO accession (2001), the global commodity boom (2003–2012), the Paris Agreement (2015), and the COVID-19 onset—all turning points that influence the energy–growth–emissions nexus. Ending in 2021 avoids the distortions of incomplete data from 2022–2023, yet still reflects the early post-pandemic recovery. We disaggregate total primary energy into coal, petroleum, natural gas, renewables, and nuclear for three reasons. First, these sources collectively constitute the vast majority of primary energy consumption in both countries, ensuring that the key drivers are comprehensively represented. Second, coal and petroleum have high emission factors, natural gas has lower ones, renewables have near-zero ones, and nuclear power has zero in operation; separating them allows source-specific elasticities to be estimated. Third, India and China have explicit targets for coal substitution, reducing petroleum imports, expanding the gas market, adding renewable capacity, and expanding nuclear build-out. Modelling each source individually, therefore, yields actionable insights.

5. Results and Discussion

5.1. Influence of Energy Consumption and Economic Growth on CO₂ Emissions in India

The analysis for India draws on a 32-year sample from 1990 to 2021, a period marked by rapid economic liberalisation, industrialisation, and substantial changes in the energy sector. This period is highly relevant for understanding India’s evolving energy-emissions nexus. The dataset was divided into a training set (1990–2013) and a testing set (2014–2021).

Table 3. Influence of energy consumption and economic growth on CO₂ emissions in India.

	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7
	Lag 1
MSE_train	0.0240	0.011	0.006	0.005	0.003	0.003	0.002
MAE_Train	0.1420	0.095	0.062	0.061	0.044	0.044	0.039
MedAE_train	0.1358	0.083	0.053	0.047	0.034	0.040	0.031
MSE_test	0.0590	0.013	0.012	0.009	0.011	0.013	0.014
MAE_Test	0.2330	0.105	0.016	0.090	0.084	0.084	0.082
MedAE_test	0.2460	0.099	0.102	0.092	0.057	0.055	0.051
	Lag 2
MSE_train	0.010	0.005	0.003	0.002	0.001	0.001	0.001
MAE_Train	0.100	0.061	0.041	0.037	0.025	0.028	0.025
MedAE_train	0.098	0.061	0.030	0.032	0.016	0.025	0.021
MSE_test	0.008	0.009	0.008	0.007	0.016	0.017	0.011
MAE_Test	0.098	0.069	0.068	0.069	0.087	0.089	0.077
MedAE_test	0.094	0.049	0.048	0.052	0.051	0.050	0.043
	Lag 3
MSE_train	0.008	0.003	0.001	0.001	0.001	0.001	0.001
MAE_Train	0.080	0.050	0.032	0.026	0.018	0.024	0.019
MedAE_train	0.084	0.044	0.025	0.024	0.013	0.023	0.016
MSE_test	0.007	0.013	0.008	0.006	0.011	0.015	0.008
MAE_Test	0.057	0.090	0.063	0.065	0.077	0.090	0.065
MedAE_test	0.029	0.064	0.048	0.043	0.045	0.044	0.041

presents the Mean Squared Error (MSE), Mean Absolute Error (MAE), and Median Absolute Error (MedAE) for various model configurations and lag structures.

Incorporating independent variables (IVs) consistently reduced all error metrics at lag 1, except in the case of nuclear energy consumption (NEC). The testing dataset reflected similar trends, although including real GDP slightly increased the MSE. For lag 2, most IVs significantly impacted CO₂ emissions, with NEC remaining statistically insignificant. Results for lag 3 confirmed that NEC does not have a meaningful effect. Increasing lag lengths reduced error metrics, suggesting a long-term influence of IVs on CO₂ emissions in India. Coal consumption, natural gas consumption, and real GDP emerged as the most influential predictors across different lag structures. Taken together, these findings validate Hypothesis 1 (H1).

These results are broadly consistent with the existing literature, which identifies fossil fuel combustion as the primary driver of global CO₂ emissions [29,30]. In particular, the dominant effects of coal and natural gas corroborate the country-specific evidence reported by Khochian and Nademi [33] and Dash et al. [68]. In contrast, the negligible contribution of nuclear energy mirrors the findings of Ozgur et al. [28]. Moreover, the positive association between real GDP and emissions lends empirical support to the feedback hypothesis posited among others by Saidi et al. [26] and Namahoro et al. [27], underscoring the intricate linkage between economic expansion and environmental degradation in emerging economies.

Our findings can provide actionable guidance for India’s decarbonization strategy. Firstly, to address emissions from existing coal and natural gas power plants, India should invest in carbon capture and storage (CCS) technologies. CCS can capture up to 90% of CO₂ emissions from power plants, preventing them from entering the atmosphere. This technology can be integrated with existing plants to reduce their carbon footprint. A pilot project in Bhopal has already demonstrated the feasibility of CCS in India [69,70,71]. The United States has implemented CCS in several power plants, providing a successful precedent for this technology [71]. Secondly, enhancing energy efficiency across various sectors and promoting the electrification of transport can significantly reduce CO₂ emissions. Policies should focus on improving energy efficiency standards for industries, buildings, and appliances, and incentivising the adoption of electric vehicles (EVs. This approach not only reduces emissions but also decreases energy consumption. Norway’s aggressive policies on EV adoption, including tax exemptions and incentives, have made it a global leader in electric mobility, serving as a model for countries such as India [72,73,74].

Finally, India should prioritise the expansion of renewable energy capacity to reduce reliance on coal and natural gas, which are major contributors to CO₂ emissions. By increasing the share of renewables in the energy mix, India can significantly mitigate CO₂ emissions. This can be achieved through policy measures such as subsidies for renewable energy projects, streamlined approval processes, and incentives for private investments in solar and wind energy. Although the government of India currently provides subsidies for electric vehicles and solar panels, these subsidies remain inadequate to make such technologies widely affordable. To accelerate adoption, the government needs to enhance the scale and reach of these financial incentives, particularly for lower-income households and small businesses. A real-world precedent for this approach is Germany’s Energiewende, which has successfully increased the share of renewables in its energy mix through robust and sustained policy support [72,75,76].

5.2. Influence of Energy Consumption and Economic Growth on CO₂ Emissions in China

The analysis for China also uses a 32-year sample from 1990 to 2021, employing the same train-test split. Unlike India, China’s results show a stronger and more consistent impact from coal and petroleum consumption, reflecting differences in energy dependency and economic structure.

Table 4. Influence of energy consumption and economic growth on CO₂ emissions in China.

	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7
	Lag 1
MSE_train	0.031	0.014	0.010	0.010	0.009	0.009	0.008
MAE_Train	0.160	0.100	0.092	0.086	0.085	0.083	0.080
MedAE_train	0.180	0.100	0.100	0.097	0.096	0.092	0.091
MSE_test	0.037	0.010	0.004	0.003	0.023	0.056	0.045
MAE_Test	0.193	0.096	0.049	0.049	0.140	0.220	0.200
MedAE_test	0.196	0.100	0.045	0.044	0.130	0.200	0.190
	Lag 2
MSE_train	0.0152	0.0077	0.0075	0.0058	0.0061	0.0061	0.0047
MAE_Train	0.106	0.076	0.079	0.068	0.071	0.071	0.062
MedAE_train	0.11	0.075	0.089	0.071	0.082	0.077	0.068
MSE_test	0.004	0.002	0.009	0.007	0.027	0.048	0.03
MAE_Test	0.0603	0.03862	0.0893	0.0828	0.162	0.213	0.171
MedAE_test	0.0625	0.0492	0.0864	0.0804	0.15	0.195	0.1619
	Lag 3
MSE_train	0.011	0.006	0.006	0.004	0.005	0.005	0.004
MAE_Train	0.920	0.070	0.074	0.057	0.063	0.066	0.053
MedAE_train	0.080	0.081	0.087	0.071	0.071	0.074	0.062
MSE_test	0.001	0.002	0.011	0.067	0.023	0.044	0.018
MAE_Test	0.031	0.034	0.100	0.078	0.150	0.200	0.130
MedAE_test	0.027	0.020	0.100	0.077	0.150	0.190	0.130

summarises the key error metrics.

The training dataset at lag 1 revealed that all IVs significantly influenced CO₂ emissions. However, in the testing set, the influence of renewable and nuclear energy consumption was less consistent. Coal, petroleum, and real GDP consistently had strong impacts across all lag structures. Real GDP demonstrated a particularly robust relationship with CO₂ emissions across training and testing datasets. Collectively, these results corroborate Hypothesis 2 (H2): China’s emissions trajectory is driven primarily by coal and petroleum use, whereas renewable and nuclear energy play weaker or inconsistent roles.

These findings resonate with the broader empirical literature documenting the carbon-intensive growth trajectories of large industrial economies such as China [33,68]. The pronounced and persistent impact of coal and petroleum consumption parallels the evidence reported by Radmehr et al. [35], underscoring the centrality of fossil fuels in the country’s current energy portfolio. Conversely, the comparatively modest influence of renewable and nuclear energy agrees with the results of Wen et al. [39] and Ozgur et al. [28], which highlight the structural and technological barriers that continue to hamper large-scale clean-energy deployment. Finally, the robust positive association between real GDP and CO₂ emissions corroborates the growth–environment nexus identified by Rahman et al. [77] and Pradhan et al. [43], reaffirming that China’s rapid economic expansion remains tightly coupled with elevated carbon emissions.

Our empirical results show that coal and petroleum remain the dominant drivers of China’s CO₂ emissions, while grid-integration bottlenecks constrain the climate benefits of renewable and nuclear capacity additions. In the short to medium term, Beijing should concentrate on three interconnected priorities. First, existing coal-fired power plants must be upgraded with ultra-super critical technology and large-scale carbon capture, utilisation, and storage (CCUS) systems; plants that cannot meet strict ultra-low-emission standards should be retired early, thereby cutting emissions at their source [78,79]. Second, the share of renewable electricity must grow rapidly through accelerated solar- and wind-capacity auctions, complemented by cost-reflective feed-in premiums and green-bond finance [80,81]. To ensure that new clean power is actually dispatched, the government should roll out grid-scale storage pilot projects, modernise transmission corridors, and introduce spot-market and ancillary-service reforms that lower renewable curtailment [82]. Third, freight transport—China’s most petroleum-intensive segment—should be steered toward electrification and hydrogen or natural-gas fuels via purchase incentives, zero-emission freight zones, and efficiency improvement programs for logistics hubs [83]. Simultaneously, tighter industrial energy-efficiency standards and circular-economy tax credits can help decouple GDP growth from carbon intensity.

Looking further ahead, policy must lock in a structural shift away from fossil fuels. The most inefficient and polluting coal units should be fully phased out, while residual plants operate at near-zero emissions through continuous CCUS upgrades [84]. By mid-third decade, coal’s role in power generation should be marginal, with utility-scale renewables forming the backbone of the electricity system [80,81]. The freight sector must complete its transition to electric and hydrogen propulsion, supported by an extensive nationwide charging and refuelling network and stringent residual-emission caps [83]. Achieving a predominantly renewable grid will require the full integration of long-duration battery, pumped-hydro, and hydrogen storage solutions, underpinned by nodal pricing and robust ancillary-service markets that attract private capital to advanced storage technologies [82]. Finally, sustained green-finance mobilisation—through sovereign transition bonds and innovation funds—should channel resources into next-generation renewables, CCUS, and storage R&D, reinforcing China’s “dual-carbon” targets of peaking emissions by 2030 and achieving carbon neutrality by 2060.

5.3. LSTM Model Validation

Table 5. LSTM model structure and parameters.

Data Normalization	MinMaxScaler
Activation function	Tanh
Optimizers	Adam
Loss Function	MSE
Input dimension	(1, timesteps∗features)
Output dimension	1 (forecast)
Hidden layers	(32, 16, 8)
Dropouts	0.1
Learning rate	0.001
Batch size	32
Training epochs	1000
Activation function	Tanh

presents the detailed configuration of the LSTM model used in this study. This specific configuration was selected to balance model complexity with training efficiency, ensuring the network effectively captures non-linear and sequential patterns in the data.

The LSTM model employs the tanh activation function across three hidden layers (32, 16, and 8 neurons, respectively) and uses the Adam optimiser with a learning rate of 0.001. The model was trained over 1000 epochs with a batch size of 32, minimising the MSE during training to enhance predictive accuracy.

To summarise, the LSTM estimates indicate that for India, coal and natural-gas consumption together with real GDP constitute the dominant determinants of CO₂ emissions, whereas renewable energy exhibits a statistically significant but mitigating effect, and nuclear energy remains negligible. By contrast, in China, the emissions trajectory is driven primarily by coal and petroleum use, as well as aggregate economic activity, with renewable and nuclear sources displaying weak and inconsistent coefficients. These cross-country differences reinforce the central role of fossil fuels in shaping the carbon footprints of emerging and industrial economies, as documented in the preceding literature review. Taken together, the divergent predictor rankings and error-reduction profiles across national models confirm Hypothesis 3 (H3), which anticipated significant differences in variable importance and predictive performance between India and China due to distinct energy mixes and policy regimes. Moreover, the superior performance of longer-lag models underscores the importance of temporal dynamics when analysing the energy–growth–emissions nexus, suggesting that policy interventions enacted today will exert compounded effects on environmental outcomes over the medium and long term.

6. Conclusions and Policy Recommendations

This study employed Long Short-Term Memory neural networks to analyse the non-linear, path-dependent relationships among energy consumption, economic growth, and CO₂ emissions in India and China over 1990–2021. Three hypotheses were confirmed: H1 (India): Coal and natural-gas consumption exert the strongest positive effect on CO₂ emissions, while renewable energy exerts a mitigating effect; H2 (China): Coal and petroleum consumption dominate the emissions trajectory, with renewable and nuclear energy exhibiting weaker or inconsistent impacts; H3: The predictive performance and variable importance differ significantly between India and China, reflecting country-specific energy structures and policy regimes.

We emphasise once again that coal remains the backbone of both India’s and China’s energy systems, albeit with distinct profiles. In India, coal supplied nearly 79% of commercial energy in the fiscal year 2023–2024 and still accounts for about 56% of total primary energy, far outstripping crude oil (33%) and natural gas (8%). Ongoing heatwaves, hydropower shortfalls, and infrastructure-led demand for cement and steel have pushed coal burn even higher, with the power sector alone absorbing roughly 69% of all raw-coal use. Although renewables account for 42% of India’s final energy mix across our study period, they comprise only about 4% of primary supply and are expanding too slowly to keep pace with surging electricity demand. China displays a similar, if larger-scale, dependence: coal provided 60.9% of total energy supply in 2023, and the country—responsible for more than 31% of global CO₂ emissions—remains the world’s top producer, importer, and consumer of coal. China’s 2023 final-energy mix was still dominated by fossil fuels (coal 21.8%, oil products 28%, natural gas 10%), with electricity (much of it coal-fired) accounting for 28.6% and nuclear power supplying a modest 4.9% of total generation. Together, these figures underscore the magnitude of the transition from coal to clean energy required in both economies if global climate targets are to be met.

In our research, the superior performance of models that incorporate longer lags underscores the cumulative nature of energy–emissions dynamics: policy interventions launched today will yield compounding environmental benefits—or costs—over time. Thus, sustained, long-horizon strategies rather than short-term fixes are imperative for effective decarbonization.

We recommend a three-pillar approach for India’s near-term decarbonisation. First, retrofit existing coal- and gas-fired power plants with carbon-capture and storage; domestic pilot projects and international examples show such retrofits can remove up to 90 percent of stack emissions and buy time for deeper structural change. Second, tighten efficiency standards for industry, buildings, and appliances while accelerating transport electrification; raising purchase rebates and tax credits for electric two- and three-wheelers can replicate Norway-style uptake in India’s cost-sensitive market. Third, scale renewable capacity by enlarging production-linked incentives, streamlining land and grid approvals, and offering long-term feed-in contracts that crowd in private capital for solar and wind. Existing subsidies for rooftop photovoltaic solar and electric vehicles could be doubled and ring-fenced for low-income households and MSMEs. Combining aggressive efficiency gains with a larger clean-energy push would relieve India’s heavy dependence on coal and natural gas, and place the country on a lower-carbon growth trajectory in the medium term.

For China, we recommend targeting coal and petroleum while clearing obstacles to the expansion of clean power. By middle-term, deploy ultra-supercritical upgrades and large-scale CCUS across the competitive coal fleet, retiring plants that cannot meet ultra-low-emission norms. Complement this with accelerated solar- and wind-capacity auctions underwritten by feed-in premiums and green bonds. To ensure new clean power is dispatched, expand pilot programmes for grid-scale storage, modernise transmission corridors, and introduce spot-market and ancillary-service reforms that reduce curtailment. In freight transport, prioritise electrification and hydrogen or liquefied natural gas substitution through purchase incentives, zero-emission freight zones, and logistics-efficiency schemes. Looking beyond the middle term, phase out the least-efficient coal units, rely on long-duration batteries, pumped hydro, and hydrogen storage to stabilise a predominantly renewable grid, and complete the shift of heavy transport to electric and hydrogen propulsion supported by a nationwide charging and refuelling network. Transition bonds and innovation funds can funnel capital into next-generation renewables, CCUS, and storage, anchoring China’s dual-carbon goals for 2030 and 2060.

Although the study offers robust insights, several constraints should be acknowledged. First, the analysis is confined to macro-level indicators; sector-specific drivers, technological innovation metrics, institutional variables, and policy stringency indices were not included. Second, the LSTM model, while powerful, operates as a black box, limiting interpretability. Third, the dataset ends in 2021, thereby excluding the most recent policy shifts and post-pandemic recovery patterns. Fourth, the relatively small shares of renewable and nuclear energy in both countries during the study period restrict the ability to gauge their full mitigation potential. Finally, although multiple lag structures were explored, explicit feedback loops between economic growth and emissions were not modelled. Addressing these gaps will refine future assessments of the energy–growth–emissions nexus.

Future research should broaden both the data architecture and methodological toolkit used to examine the energy–growth–emissions nexus. First, incorporating technological-innovation indices, regulatory-quality metrics, and disaggregated sector-level energy data would yield finer-grained insights, while extending the sample beyond 2021 would capture the impacts of post-COVID recovery packages and recent climate pledges. Second, comparative studies that include a broader spectrum of emerging and advanced economies can highlight regional heterogeneities and policy transferability. Third, hybrid machine-learning–econometric frameworks are needed to track dynamic feedback loops and structural breaks more effectively. Finally, building on the preliminary work of Afolabi et al. [85], who show that AI-driven smart grids and predictive maintenance can boost energy efficiency by up to 26.7% and 20%, respectively, and Hu et al. [86], who document AI’s positive spillovers on productivity and innovation, future studies should quantify how such AI applications can be scaled across Asia’s diverse contexts and how the resulting economic gains could be channelled into net-zero initiatives.