AI has transformed aerosol science, and a new analysis maps exactly how

Their bibliometric and text-mining study, published in the Asian Journal of Atmospheric Environment, reveals a field that has expanded explosively since 2019 and is undergoing a fundamental shift in purpose: away from simple concentration prediction and toward multisource data integration, satellite retrieval, spatiotemporal modeling, and exposure-oriented applications.

The scale of the growth is striking. In 2019, the Web of Science indexed 207 articles applying AI to aerosol and PM research. By 2020 the count had risen to 349, reaching 568 in 2021, 781 in both 2022 and 2023, 980 in 2024, and 1,192 in 2025. The authors note that this sustained increase indicates AI-based approaches became increasingly utilized in aerosol and particulate-matter research during the early 2020s. They also caution that the 2025 figures should be interpreted with reference to the retrieval date, since recent annual records can be skewed by indexing and database updates. As a supplementary context, the team separately examined 99 publications from the Korea Citation Index, the national Korean database, which showed domestic annual output rising from one article in 2017 to 22 in 2025, although those records were deliberately excluded from all core analyses.

Why did AI take hold so quickly in this particular corner of atmospheric science? The answer, according to the study, lies in the nature of the data itself. Aerosol and PM variability is governed by nonlinear interactions between emissions, meteorology, topography, land use, and seasonality, which can limit the ability of conventional linear approaches to represent complex dependencies. At the same time, ground-based monitoring networks, satellite remote sensing, numerical models, and reanalysis products have produced increasingly dense and diverse observations of aerosol properties. These heterogeneous data sources differ in dimensionality, resolution, uncertainty, and sampling characteristics, creating conditions in which nonlinear and multisource modeling can be genuinely useful. Machine learning and deep learning methods can model these nonlinear relationships, integrate disparate observations, and generate high-resolution spatiotemporal information that no single observation system could provide alone.

Geographically, the field is heavily concentrated in Asia. China dominated the analyzed dataset with 2,310 articles, followed by the United States with 550, India with 270, and South Korea with 252. China also recorded the highest total citation count at 56,088, ahead of the United States at 14,113, India at 4,890, and South Korea at 4,280. But when the researchers ranked countries by mean citations per publication, the picture changed dramatically: the United Kingdom led with 37.6, followed by Australia at 30.6, France at 27.7, Italy at 26.4, and the United States at 25.7, while China averaged 24.3. The authors emphasize that publication productivity and citation intensity represent distinct characteristics of the analyzed literature and should not be interpreted interchangeably, particularly since citation accumulation depends on publication age, topic, journal, and field-specific citation practices.

The institutional landscape reveals how differently nations organize this research. In China, the Chinese Academy of Sciences led with 868 affiliation records, followed by Wuhan University, Nanjing University of Information Science and Technology, Peking University, and Fudan University. In the United States, the University of California ranked first with 195 records, followed by the University of Maryland, Harvard University, and NASA, with NOAA also prominent, highlighting the importance of large-scale satellite and atmospheric observation infrastructure. In India, the Indian Institutes of Technology produced 120 affiliation records, while in South Korea Seoul National University led with 72, and in Italy the National Research Council of Italy topped the list with 31. Notably, Italy’s profile included a health sector organization, ASL Roma 1, suggesting stronger integration of aerosol research with exposure and health applications. The authors conclude that AI-based aerosol research has not yet been conducted using a single institutional model, reflecting national differences in research infrastructure, observational data availability, and application needs.

The journals publishing this work span atmospheric science, remote sensing, environmental pollution, and environmental health, underscoring the field’s multidisciplinary character. Atmosphere published the most articles at 225, followed by Remote Sensing with 212, Atmospheric Environment with 200, Science of the Total Environment with 194, and Atmospheric Pollution Research with 145. Yet citation impact told a different story: Science of the Total Environment accumulated the largest total citation count at 8,068, and Environmental Science & Technology showed the highest mean citations per publication at 50.0, while Atmosphere and Remote Sensing had lower mean citation counts despite their high publication volumes. The journal distribution, the authors suggest, reflects three overlapping publication domains: atmospheric environment journals focused directly on aerosols and PM, remote sensing journals capturing satellite retrieval work, and broad environmental health journals reflecting the extension of AI-based aerosol research toward exposure and health-related applications.

Delving into the most influential papers, the analysis identified the ten most highly cited articles in the dataset. The most cited, by Yuan and colleagues in 2020, was a review of deep learning in environmental remote sensing with 1,403 citations. Other landmark studies included Di and colleagues’ 2019 ensemble-based model for high-resolution PM2.5 across the contiguous United States, Wei and colleagues’ space-time random forest estimation of PM2.5 across China, and Geng and colleagues’ 2021 near-real-time PM2.5 retrievals using multisource data fusion. Spatiotemporal forecasting also featured prominently, with Qi and colleagues combining graph convolutional neural networks with long short-term memory networks to represent the spatial dependence and temporal evolution of PM2.5. The clustering of highly cited papers between 2019 and 2021 suggests that period marked a key phase in the rapid expansion of AI-based PM estimation and forecasting, pointing to four major methodological developments: deep learning integrated with satellite remote sensing, high-resolution PM2.5 estimation, spatiotemporal deep-learning prediction, and multisource data fusion.

Text mining of the literature’s own language revealed the field’s conceptual architecture. Keyword co-occurrence analysis identified machine learning as the central theme linking PM prediction, deep learning, aerosol optical depth, exposure assessment, and satellite remote sensing. Four principal components emerged: machine learning and deep learning prediction of PM, PM exposure and health-related assessment, satellite-based aerosol and PM estimation, and emission, transport, and regional pollution analyses. A second, independent analysis using abstract-based clustering with TF-IDF features and non-negative matrix factorization converged on five thematic domains: prediction modeling, PM concentration and exposure, satellite-based aerosol retrieval, emission-related analysis, and aerosol monitoring and environmental applications. The convergence of these two independent text-based analyses supports a relatively stable knowledge structure, and the authors interpret it as evidence that AI has developed beyond an auxiliary statistical tool, increasingly supporting the generation of spatially continuous, high-resolution aerosol information from heterogeneous observations.

The study’s authors argue that this structural transformation matters beyond bibliometrics. High-resolution AI-derived PM fields are increasingly used as intermediate products to estimate population exposure and support environmental assessments, and the appearance of air quality and air pollution terms in the thematic clusters should be read as an application context for aerosol research rather than an expansion into all gaseous pollutants. Looking forward, they contend that future AI-based aerosol research should move beyond predictive accuracy toward interpretability, transferability, uncertainty characterization, and physical consistency. Explainable-AI approaches such as SHAP-based attribution can help quantify the contribution of input variables, but such explanations should be evaluated against atmospheric knowledge rather than treated as causal evidence. Geographical transferability remains a major challenge, since models trained in densely monitored regions may not generalize to locations with different emissions, meteorology, aerosol compositions, or observation infrastructures, making transfer learning, domain adaptation, and rigorous out-of-domain validation important priorities.

The analysis also carries clear caveats. The primary dataset was restricted to the Web of Science and therefore does not represent literature indexed exclusively in Scopus or discipline-specific databases, and the literature set was sensitive to search terms and screening decisions. Citation indicators were influenced by publication age and disciplinary citation practices, and keyword and abstract analyses do not comprehensively represent methodological details, data quality, or model performance. Author keywords were not passed through a synonym thesaurus before network construction, so near-duplicate terms may appear as separate nodes. The authors also note that the bibliometric pattern alone does not establish the causes of publication growth. Even so, they conclude that the collective evidence indicates a clear evolution of AI-based aerosol research from individual prediction models toward an integrated data science framework for producing, interpreting, and applying high-resolution aerosol information, with future progress depending on interpretable and physically informed learning, uncertainty-aware multisource integration, and robust cross-region generalization.

Subject of Research: A bibliometric and text-mining analysis of artificial intelligence applications in atmospheric aerosol and particulate matter research

Article Title: Mapping the evolution of artificial intelligence in atmospheric aerosol research: a bibliometric and text-mining analysis

Article References: Lee, S.-C., & Lee, K.-H. (2026). Mapping the evolution of artificial intelligence in atmospheric aerosol research: a bibliometric and text-mining analysis. Asian Journal of Atmospheric Environment, 20(1), Article 21. https://doi.org/10.1007/s44273-026-00101-7

Image Credits: AI Generated

DOI: 10.1007/s44273-026-00101-7

Keywords: artificial intelligence, machine learning, atmospheric aerosols, particulate matter, PM2.5, bibliometrics, text mining, satellite remote sensing, aerosol optical depth, exposure assessment, Web of Science, deep learning