Coal Emerges as the Dominant Driver of Deadly PM2.5 Pollution in Landmark Machine Learning Study

Fine particulate matter, the microscopic pollutant known as PM2.5, remains one of the most consequential threats to public health worldwide, penetrating deep into the lungs and bloodstream and contributing to millions of premature deaths each year. A new study published in the journal Air Quality, Atmosphere & Health has now disentangled, with unusual precision, which parts of a nation’s energy system push this pollutant upward and which ones pull it down. The research, led by S. Arafat Ayon of Noakhali Science and Technology University together with an international team of economists and energy analysts, examines sixteen nuclear-consuming countries across the G7 and the European Union over a thirty-one-year span, from 1990 to 2021, and reaches conclusions that carry direct implications for energy policy on both sides of the Atlantic.

The study’s central finding is stark: coal is the single most powerful driver of PM2.5 pollution in the countries analyzed. The researchers calculated that a 1 percent increase in coal consumption raises fine particulate concentrations by approximately 0.22 percent, an elasticity that dwarfs the contribution of any other fossil fuel in their models. Oil consumption also worsens air quality, though its effect is considerably weaker. This hierarchy matters because it quantifies, in econometric terms, what atmospheric scientists have long suspected from emission inventories and chemical transport models: that the combustion of coal for electricity and industry releases a dense cocktail of primary particles and precursor gases, including sulfur dioxide and nitrogen oxides, that ultimately convert into secondary fine particulates in the atmosphere.

On the other side of the ledger, the analysis delivers a consistently favorable verdict for nuclear energy. Across every econometric specification the team employed, nuclear power consumption was associated with lower PM2.5 concentrations, reinforcing its role as a low-particulate alternative to fossil generation in the energy mix. Biofuels told a more nuanced story. Their pollution-reducing benefits materialized only at higher levels of adoption, suggesting that modest biofuel penetration does little for air quality while substantial deployment can meaningfully displace dirtier fuels. The authors caution, however, that the relationship is not linear, and their machine learning analysis revealed an inverted S-shaped pattern for biofuels that conventional regression techniques would have entirely missed.

Methodologically, the study is notable for fusing two analytical traditions that rarely appear in the same paper. The researchers extended the STIRPAT framework, a model rooted in the classic IPAT identity proposed by Ehrlich and Holdren in 1971, which expresses environmental impact as a function of population, affluence, and technology. Rather than treating the energy mix as a single composite diversification index, they disaggregated it into source-specific elasticities for coal, oil, nuclear, and biofuels, and even modeled the interactions between these energy sources and income levels and innovation activity. This granular approach allowed them to ask not just whether the energy mix matters, but which components of it matter, and under what economic conditions.

To capture the linear structure of these relationships, the team built a baseline panel regression using fixed-effects estimation with Driscoll-Kraay standard errors, a technique designed to remain reliable when observations across countries and years are correlated, a common feature of environmental and economic data. They then stress-tested the results with quantile regression, which examines whether drivers behave differently at the low and high ends of the pollution distribution, and with Feasible Generalized Least Squares, which corrects for both serial correlation and heteroskedasticity in the error terms. The consistency of the core findings across all these specifications lends considerable weight to the conclusions.

The machine learning component is where the study breaks genuinely new ground. The researchers trained an XGBoost model, a gradient-boosted decision tree algorithm prized for its ability to detect nonlinearities and interactions in complex datasets, and then interpreted it using SHAP, or Shapley Additive Explanations, a game-theoretic method that assigns each variable a fair share of the model’s predictions. The SHAP analysis confirmed the econometric results, including the inverted U-shaped relationship between income and pollution that supports the Environmental Kuznets Curve hypothesis, the idea that environmental degradation first rises with economic development and then falls as societies grow wealthy enough to prioritize and afford cleaner technologies. But SHAP also uncovered dynamics invisible to traditional methods, including S-shaped patterns for oil consumption, in which pollution effects shift abruptly across certain thresholds of use.

Several secondary findings round out the picture. Technological innovation, measured across all models, significantly mitigated PM2.5 pollution, underscoring that research and development in cleaner processes and emission controls is not merely a co-benefit of prosperity but an active agent of environmental improvement. Perhaps more surprising, urbanization was associated with lower emissions in these advanced economies, a result that challenges the assumption that densifying cities inevitably dirty the air. In countries with modern infrastructure, stringent regulation, and efficient public transport, concentrating population and economic activity in urban areas may actually reduce per capita energy waste and facilitate cleaner heating, power, and mobility systems.

The policy implications the authors draw are organized around four priorities. First, accelerating the phase-out of coal stands out as the highest-leverage intervention, given its dominant elasticity. Second, biofuel adoption should be scaled above the critical thresholds identified in the analysis, since partial deployment delivers little air quality benefit. Third, nuclear capacity can be cautiously expanded as a proven low-particulate energy source, though the authors note evidence of diminishing returns at high penetration levels that warrant monitoring. Fourth, and perhaps most importantly, energy transitions should be explicitly linked to innovation investment, because the pollution benefits of switching fuels are amplified when they are accompanied by technological progress. Each of these priorities is grounded not in advocacy but in the quantified elasticities and machine-learned response curves the study produced.

The timing of this research is significant. The International Energy Agency has projected that fossil fuel use will peak before 2030 under current stated policies, and both the European Union and the G7 have committed to decarbonizing their electricity sectors, with the IEA publishing dedicated roadmaps for achieving net-zero electricity in G7 members. At the same time, European air quality has improved markedly over the past three decades, yet monitoring by the European Environment Agency shows that fine particulate levels across much of the continent still exceed the guidelines set by the World Health Organization. Understanding precisely which levers in the energy system deliver the largest air quality gains, and at what thresholds, is therefore not an academic exercise but a practical necessity for regulators weighing the pace and composition of the energy transition.

What makes this study likely to influence both the economics and the atmospheric science communities is its demonstration that interpretable machine learning and rigorous panel econometrics are complementary rather than competing tools. The econometric models provided causal structure, statistical robustness, and elasticities that policymakers can plug directly into cost-benefit calculations, while XGBoost and SHAP exposed the threshold effects and nonlinearities that linear models flatten away. The authors have made their data available upon reasonable request, and the code underlying the machine learning analyses can likewise be obtained from the corresponding authors. As countries across the G7 and the European Union confront the final, hardest stretch of their coal phase-outs and debate the future role of nuclear power and biofuels, this analysis offers a data-driven map of which choices will clear the air fastest, and which will leave the most dangerous particles hanging over their cities.

Subject of Research: Determinants of PM2.5 air pollution in relation to energy mix, economic growth, and technological innovation in G7 and EU nuclear-consuming countries

Article Title: Energy mix, technological innovation, and air quality nexus: A STIRPAT-extended analysis using machine learning and econometric techniques

Article References: Ayon, S. A., Ridwan, M., Hossain, M. E., Joy, M. I. H., Rehman, M. Z., & Esquivias, M. A. (2026). Energy mix, technological innovation, and air quality nexus: A STIRPAT-extended analysis using machine learning and econometric techniques. Air Quality, Atmosphere & Health, 19(10), Article 221. https://doi.org/10.1007/s11869-026-02103-4

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02103-4

Keywords: PM2.5, air quality, coal consumption, nuclear energy, biofuels, energy mix, technological innovation, XGBoost, SHAP, Environmental Kuznets Curve, STIRPAT, energy transition