Introduction

The growing concern about climate change’s effects on human health, lifestyles, and habitats highlights the urgency of this issue1,2,3. The rapid advancement of urbanization and industrialization exacerbates the detrimental effects of air pollution on health4,5,6,7,8. A complex interplay exists between climate change and air pollution. Climate-induced alterations in meteorological factors, such as temperature, rainfall, and wind speed, also change the concentration and distribution of air pollutants through different physical and chemical processes9,10,11,12,13.

The relationship between air pollution and climate change is inherently bidirectional. While the influence of air pollution on climate—with certain pollutants acting as short-lived climate forcers (SLCFs)—is a critical field of study extensively detailed in the IPCC Sixth Assessment Report (AR6)14, the reverse pathway presents distinct and complex challenges. This study focuses specifically on systematically reviewing the literature concerning this latter direction: the impact of climate change on air quality. The majority of studies predominantly use observational data and statistical analyses to present evidence and assess the severity of such impact. For instance, increased heatwave incidents due to climate change are shown to exacerbate air pollution events. Mechanistic studies examining the effects of climate change and various meteorological factors on air pollution are generally limited to specific countries or regions and are conducted using regional-scale models15,16. There’s still a lack of a broader consensus in this area.

Given the variability in air pollutants of concern across different regions and the diverse mechanisms across times and places, most studies focus narrowly on the effects of climate change on specific pollutants in specific areas. Review articles also tend to concentrate on particular regions and/or pollutants, often covering only a limited number of studies17,18. This implies that, although the scientific community is not lacking in interest or research quantity on the issue of climate change’s impact on air pollution, attaining a comprehensive, large-scale understanding remains elusive due to the challenges in integrating information. This study aims to bridge this gap by sifting and systematically reviewing related literature in this field. We use a large language model (LLM) based framework to provide a large-scale, quantitative, and systematic map of the literature. Our objective is not only to validate consensus with robust data but also to uncover deeper patterns in the evolution and distribution of scientific effort.

Specifically, we intend to address the following questions:

  1. 1.

    How have the volume and focus of research on climate change’s impact on air pollution evolved? Which climatic impact-drivers (CIDs) and air pollutants have been the focus of attention?

  2. 2.

    In which regions is there evidence of climate change impacting air pollution?

  3. 3.

    Which CIDs may impact air pollution, and what are the mechanisms of their influence?

  4. 4.

    What factors are associated with the research attention on different countries/regions? Are there any research “blind spots”?

To investigate these questions, we initially employ a LLM (GPT 3.5 in particular) to filter literature relevant to the topic of “Climate Change’s Impact on Air Pollution” from the search results. We then integrate the LLM with expert-curated keyword lists and geographical entity extraction techniques to extract crucial information in the literature, such as CIDs, air pollutants, and research regions. This allows us to conduct analysis and synthesis (Fig.1). Our approach not only enables the identification and extraction of information from literature that exceeds the capacity of manual review but also represents a novel endeavor aimed at reducing the technical barriers and challenges faced by researchers in the field of natural sciences when applying advanced natural language processing (NLP) technology to their research.

Fig. 1: Workflow schematic of the systematic review process.
figure 1

The workflow is segmented into four major stages: retrieval, screening, information extraction, and systematic review. During retrieval, a combination of climate and air pollution keywords are used to perform 1200 searches, yielding over 430,000 search results. Screening involves using large language models (LLMs) for scrutiny and manual annotation for validation, narrowing down to over 10,000 relevant documents. Information extraction combines the outputs from LLMs to identify article types and methods, whereas drivers and pollutants are determined using expert-defined lexicons, and study regions are pinpointed via geoparsing techniques. The systematic review is divided into two analyses: a large-scale research status statistical analysis examines temporal dynamics, spatial differences, and research attention, and a quantified synthesis of mechanisms employs tailored retrieval-augmented generation (RAG) technology within a deliberately constructed mechanistic framework to focus on and quantify the mechanisms. The color-coded blocks represent various components: red for documents, green for LLMs or other natural language processing (NLP) techniques, dashed outlines for processes where human efforts and guidance are employed, orange for statistical analysis, and blue for mechanism synthesis.

Full size image

Results

Evolves in research volume and focus

Our search on Web of Science with more than 1200 keywords resulted in over one million publications, which included over 430,000 unique papers. Leveraging a LLM, specifically GPT 3.5, we sifted through this vast corpus and identified 11,674 relevant documents from 1990 to 2021, focusing on “Climate Change’s Impact on Air Pollution” (see “Methods” for details). According to the CID framework introduced by the AR6, we classified climate impact drivers into eight major categories. Our study included ten air pollutants, comprising the six standard air pollutants specified by the EPA (O3, PM, CO, Pb, SO2, NO2), as well as four additional pollutants: dust, black carbon (BC), sulfate, and nitrate. Overall, Heat and Cold, Wind, and Wet and Dry emerged as the most frequently studied CIDs, while PM and O3 were the most commonly examined pollutants. In terms of regional distribution, research was most concentrated in Asia, Europe, and North America (Fig. 2).

Fig. 2: Schematic representation of the interconnections between climate impact-drivers, air pollutants, and geographic regions.
figure 2

This diagram illustrates the multidimensional relationships among climate impact-drivers (CIDs), air pollutants, and regions, based on literature frequency. The left panel presents a word cloud of meteorological factors, grouped into four CID categories (heat and cold, wind, wet and dry, and others), with text size and color intensity indicating relative prevalence in the literature. The central flow diagram traces the directional influence from CIDs to associated air pollutants and then to regions where these pollutants are most studied; the width of each connection reflects the proportion of studies reporting both ends. The right panel displays a word cloud of countries, where larger and darker labels correspond to regions with a higher volume of research on climate change impacts on air pollution.

Full size image

Historically, research on the impact of climate change on air quality began around the 1990s. Since then, the volume of literature in this field has grown rapidly, with an increasingly accelerated pace—particularly in recent years. Between 2017 and 2021 alone, 3745 relevant articles were published, accounting for nearly half of the total literature to date (Fig. 3a). Most studies have focused on two major pollutants: O3 and PM. Notably, since 2017, research attention has progressively shifted from O3 to PM, making particulate matter the dominant focus in recent years.

Fig. 3: Shifts of Volume and Focus in Subject Literature.
figure 3

a The stacked bar chart displays the proportion of literature focusing on each of ten air pollutants (PM, O3, NO2, SO2, CO, Dust, BC, Sulfate, Nitrate, Pb) during four consecutive periods (1990–1997, 1998–2005, 2006–2013, and 2014–2021) from 1990 to 2010, illustrating shifts in research attention over time. The accompanying line graph shows the total volume of related publications, marking a notable rise after 2010. b Pie charts detail the composition of research types for each of the four specified periods, with segments representing “Evidence” studies that directly report the impacts of climate change on air pollution, “Mechanism” studies including real-world experiments and model optimizations, and “Discussion” articles providing analyses or comments of existing literature. c The bar and corresponding pie chart visualize the categorization of literature by climate impact-drivers mentioned, emphasizing dominant themes such as heat and cold, wind, and wet and dry conditions.

Full size image

In this study, we classified the literature into three categories: “Evidence” (empirical studies), “Mechanism” (mechanistic investigations), and “Discussion” (reviews and commentary). The majority of the research falls under the “Evidence” category, accounting for 78.2% of all studies, followed by “Mechanism” at 16.6%, and “Discussion” at 5.2%. Over time, the proportion of “Evidence” studies has shown a slight upward trend, while “Mechanism” studies have experienced a modest decline, indicating that an increasing number of empirical cases have documented the impacts of climate change on air pollution (Fig. 3b).

In terms of air pollutants, researchers are most focused on PM (28.8%) and O3 (21.5%). The focus has shifted over the decades, with a gradual decrease in the proportion of studies related to O3, a rapid increase in PM-related studies, and a significant decrease in SO2-related studies (Fig. 3a). Among CID categories, the most attention is given to heat and cold (29.1%), wind (26.3%), and wet and dry (23.4%), while other CID categories receive relatively less attention. Little change was shown in the proportion of literature related to each CID category (Fig. 3c). A more detailed analysis of CID keyword frequencies is presented in Supplementary Fig. S1, where “temperature” and “wind” are the most frequently occurring terms, followed by a variety of other CID-related keywords. Additionally, we extracted the main methods used in the articles. At least 150 methods were applied in related research. The most frequently used methods include observation, model simulation, scenario analysis, experiments. Technologies like remote sensing, positive matrix factorization (PMF), and meta-analysis were also employed (Supplementary Fig. S2).

Characteristics of different study areas

The Mordecai geographical entity extractor was employed to determine the research area of all relevant literature, a methodology proven effective in a similar extensive literature review. The research areas were categorized by continents, with Asia (44.8%) being the most studied region, followed by Europe (23.1%) and North America (20.5%). Research pertaining to Africa (5.0%), South America (3.1%), and Oceania (2.9%) was comparatively less. Notably, studies on China (26.5%), the United States (14.7%), and India (6.5%) outnumbered those on other countries. Europe and North America have seen a steady increase in related research since the 1990s, indicating an earlier commencement in these studies, while a marked rise in other continents began around 2010 (Fig. 4, Supplementary Fig. S3).

Fig. 4: Number of studies and changes in focus across different research regions.
figure 4

This composite visualization presents the global distribution of research volume by country and the temporal shift in pollutant focus across continents, excluding Antarctica. The central map employs circles whose size and color intensity categorize the volume of literature studying each country, identified through geoparsing technology. Surrounding the map are six subfigures corresponding to North America, Europe, Asia, South America, Africa, and Oceania, akin to Fig. 3a. These graphs track the changes in the number of studies over time, alongside the evolution of research attention to various pollutants within each continent across four distinct periods (1990–1997, 1998–2005, 2006–2013, and 2014–2021).

Full size image

The focus of studies across different regions also differed with respect to pollutants. For example, studies of Africa emphasized dust more than other continents. Studies of Europe, North America, and Oceania commonly concentrated on PM and O3. Research concerning Asia and South America predominantly focused on PM. Most regions, apart from Africa and South America, demonstrated a trend of shifting attention from O3 to PM (Fig. 4). Various factors influence the air pollutants of interest in different regions, including geographical elements (like Africa’s desert landscape prone to sandstorms19), source of pollution (like Asia’s rapid economic boom and intensive industrialization leading to significant particulate emissions20,21,22) and climatic conditions (like Europe’s conducive summer weather for ozone generation23,24). These regional differences in pollutant focus not only reflect local geographical and socioeconomic characteristics, but also hint at varying interactions between specific CIDs and air pollutants.

The mechanisms of CID’s influence on air pollutants

Studies that mention both a CID and an air pollutant often indicate that the discussed CID might be influencing that pollutant. It also implies that when investigating specific air pollutants, researchers tend to focus on the potential roles of specific CIDs. We have examined such studies, grouping them by pollutant type and analyzing the proportion of focused CIDs within each group (Supplementary Fig. S4). For PM, the emphasis is on the effects of heat and dry, wind, and air pollution weather, suggesting a concentration on PM’s dispersion process. With O3, the focus shifts to heat and dry, radiation, while less emphasis is placed on wet and dry, indicating an interest in O3’s formation process. In the case of sulfate and nitrate, attention is paid to wet and dry, pointing to an interest in their dry and wet deposition processes. For dust, the focus is on wind, with lesser attention to heat and cold, reflecting a concern for dust’s migration process and impact. In researching BC, the emphasis is on radiation, open ocean, and snow and ice, underscoring concerns about the radiative effects of BC.

Meteorological factors influence the concentration and distribution of air pollutants through a range of physical and chemical mechanisms, affecting emissions, reactions, formations, dispersions, transports, and degradations of air pollutants. We’ve synthesized from previous research the impact mechanisms and principal effects of key meteorological factors like temperature, wind, precipitation, radiation and air pressure on common air pollutants including PM, O3, NO2, SO2, CO, classifying these impacts by their processes (Supplementary Table S1). Our synthesis reiterates the complexity of meteorological factors’ impacts on air pollutants. A single meteorological factor can produce contrasting effects on the same pollutant, depending on the influence process and action mode. This adds layers of uncertainty and complexity to research on climate change’s overall impact on air pollution. For instance, considering the increased evaporation loss of PM due to high temperatures suggests a negative impact of temperature on PM pollution. Conversely, if the increased rate of photochemical reactions due to high temperatures and strong solar radiation is considered, which speeds up the generation of PM precursors and secondary aerosols, then the temperature effect is positive. Thus, research into the effects of meteorological factor on air pollution must consider the studied area, focused season, and dominant influencing process, and also the potential interactions between different meteorological factors.

Assessment of attention degree and influencing factors across diverse regions

Furthermore, we sought to examine how regional socioeconomic development levels relate to both research focus and the sufficiency of research attention. We found that high-income regions such as North America, Western Europe, and Australia predominantly focus on O3 as the main air pollutant of interest. In contrast, middle-income countries such as China, India, and Brazil primarily focus on PM, while several other regions center their research on dust. Notably, the most severe data gaps are concentrated in low-income regions, particularly across Africa (Fig. 5a).

Fig. 5: Research attention and influencing factors in different regions.
figure 5

a Country-level distribution of the most frequently studied air pollutant, categorized by World Bank income group classifications. Color shading indicates the predominant pollutant of research focus in each country, while black denotes countries with insufficient data (fewer than three publications on the top pollutant), highlighting critical research gaps. High-income countries tend to focus on O3, whereas upper- and lower-middle-income countries primarily focus on PM. Many low-income countries lack sufficient research to identify a dominant pollutant. b The Research Attention Degree Index (RADI), standardized on a scale from –1 to 1, quantifies the relative sufficiency or deficiency of research activity across countries, adjusted for population size. Higher RADI values indicate relatively more research attention. c Correlation analysis between RADI and selected factors, including socioeconomic indicators (e.g., GDP per capita), climatic variables (e.g., temperature and precipitation), and pollution metrics (e.g., average PM2.5 exposure). A strong positive correlation is observed with GDP per capita, while pollution exposure and health risk show inverse correlations, revealing a potential imbalance where more polluted regions receive less scientific attention.

Full size image

To systematically assess global disparities in scientific focus, we developed a Research Attention Degree Index (RADI), scaled from –1 to 1, which accounts for both the number of relevant studies and population size at the national level. Our analysis reveals that China, along with most developed countries in Europe and North America, receives relatively sufficient research attention regarding the impacts of climate change on air pollution. In contrast, significant research deficiencies are evident in India, Southeast Asia, Africa, and South America (Fig. 5b). This spatial pattern closely mirrors the geographic distribution of empirical evidence on climate–pollution interactions, suggesting that the lack of evidence in many low- and middle-income countries may be attributable more to insufficient research investment than to a true absence of climate-driven air quality effects. These areas thus represent critical blind spots in the global understanding of compound climate–pollution risks.

Further exploration was conducted on the correlation between the RADI and local climate, economy, population, the degree of impact from climate change, air pollution severity, and other factors, to investigate the causes of the “research gap”. Results revealed that the RADI positively correlates with GDP per capita. It negatively correlates with air pollution exposure (measured by annual average PM2.5 exposure per person) and health risks (indicated by the number of deaths per 100,000 people due to air pollution and the proportion of such deaths among all risk factors). There was minimal correlation observed with trends in pollution levels (Fig. 5c). These insights underscore the research deficiencies in areas where air pollution is intense but economic growth is limited, and highlight the overlooked urgency and need for increased scrutiny in regions where climate change might considerably exacerbate the threats posed by air pollution to human health. Indeed, our understanding of which regions are at risk and the degree of these risks is still inadequately comprehensive and accurate.

Discussion

In this study, we employed a LLM-based framework to systematically analyze the vast literature on the impacts of climate change on air pollution. We provided not only the synthesis of findings, but also the large-scale, quantitative validation of macro-level trends that were previously understood only qualitatively or through smaller-scale reviews. By integrating insights from numerous, often siloed studies and converting qualitative consensus and anecdotal knowledge into a quantitative, data-driven evidence map, our work offers a new, holistic understanding of the field’s state and trajectory. Utilizing modern NLP technologies enabled us to undertake this extensive task, encompassing the screening of over 430,000 documents and the in-depth analysis of more than 10,000 papers.

Traditionally, employing NLP technology, particularly models requiring fine-tuning, presented substantial hurdles for researchers outside of computer science due to complexities like data preparation, manual labeling, model training, and securing computational resources. It also demands substantial preparatory work and unavoidable human efforts. In our study, we explored an approach demonstrating that LLMs based on prompts such as GPT, even without extensive sample training and customization, can effectively perform literature analysis tasks using straightforward prompt engineering and minimal examples. Our manual annotation of 1328 papers to evaluate the reliability of the LLM revealed a good 86.1% accuracy in assessing topic relevance.

The emergent abilities like advanced reasoning and knowledge inference of LLMs significantly broaden their applicability in scientific fields. For instance, our study required filtering literature where climate change or meteorological factors are independent variables and air pollution is the dependent variable. This task necessitates logical reasoning and judgment based on contextual knowledge, challenging for conventional statistical or topic-relevance methods. In an era characterized by abundant and rapidly expanding scientific literature, prompt-based LLMs present a novel approach, liberating researchers from laborious data annotation and extensive dataset fine-tuning. This method allows for efficient extraction of information from massive unstructured texts, enabling researchers to focus on more innovative aspects of their work.

Our research shows that the study of “Climate Change’s Impact on Air Pollution” emerged around the 1990s and experienced a surge in growth after 201025,26. The primary pollutants of interest, PM and O3, align with conventional understandings. However, our findings reveal a noticeable shift over the past three decades: while the focus on O3 has gradually diminished, attention to PM has increased significantly. This shift highlights the evolving priorities of researchers regarding different air pollutants.

We have systematically reviewed the complex mechanisms by which meteorological factors affect air pollutants. It appears that the same meteorological factor can yield divergent, sometimes contradictory, effects on a pollutant, varying by region and period due to different operational mechanisms and pollutant characteristics27,28. This complexity might explain why many studies have focused on the impact of certain meteorological factors on specific pollutants in certain areas and time periods, or have concentrated on investigating specific historical pollution events. Despite advancements in high-precision and high-resolution regional air quality models that capture various physical and chemical processes of air pollutants, accurately projecting and quantifying the overall effects of climate change on diverse air pollutants globally remains an uncertain and challenging task. This uncertainty is not merely a matter of computational capacity and model development; it is also intricately linked to the complexity of how climate change influences air pollution.

Our study also uncovers how evidence and studies on the effects of climate change on air pollution are globally distributed. Regions such as Europe, North America, and Oceania show significant interest in both PM and O3, while Asia and South America display a marked preference for PM studies21,23,29,30. When considered from a “research per capita” perspective, research in China and most developed nations in Europe and America is more extensive, but significantly less so in India, Southeast Asia, Africa, and most South American countries. This disparity appears to stem from economic factors, like GDP per capita, rather than the severity of climate change impacts or air pollution levels in these areas.

Although we employed extensive keyword searches and leveraged large language models (LLMs) to process a vast literature corpus, some relevant studies may have been missed. LLMs can efficiently handle volumes beyond human capacity, but their accuracy is not perfect, and expert-led systematic reviews remain essential. Our analysis relied only on titles and abstracts, which suits large-scale synthesis; however, deeper examination of full texts could reveal more nuanced mechanisms linking climate change and air pollution. As NLP and multimodal technologies advance, LLMs are best viewed as powerful tools to enhance—not replace—scientific judgment and discovery.

We specifically focus on the unidirectional impact of climate change on air quality. It should be noted that the critical bidirectional effect, where air pollutants influence climate, represents an equally important dimension of the nexus. Future research could powerfully leverage the large-scale analytical methods demonstrated in this study to systematically map these mutual interactions. Such work could illuminate complex feedback loops and provide a more holistic evidence base for integrated mitigation policies.

Despite growing interest and an expanding body of literature, our findings highlight that the current understanding of how climate change influences air quality remains fragmented and uneven. Mechanistic clarity is still limited, particularly regarding the complex, and sometimes opposing, pathways through which meteorological factors affect different pollutants. Empirical evidence is concentrated in a few high-income regions, while large areas—especially in low- and middle-income countries—remain understudied, raising concerns about global research equity. This imbalance is particularly troubling given that many of these underrepresented regions are highly vulnerable to the combined health burdens of climate change and air pollution. It signals a need for a more globally equitable distribution of research efforts. Enhancing international collaboration in these areas is essential.

The methodological approach of this study offers a powerful paradigm for addressing the broader poly-crises faced by humanity. As intergovernmental bodies like the IPCC and the Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services (IPBES) increasingly undertake collaborative assessments on human health and biodiversity effects of both climate change and environmental pollution31, the need for new synthesis tools becomes paramount. LLM-based analyses, such as the one demonstrated here, could be helpful in assessing and finding links and connections in a vast and rapidly growing body of literature, including assessment reports. This framework represents a scalable approach for such endeavors, capable of identifying crucial research trends and knowledge gaps across complex scientific domains, thereby supporting more timely and comprehensive global assessments. Our data-driven identification of “research blind spots” provides an example that can be referenced to support international efforts. We recommend that such evidence synthesis be utilized to prioritize collaborative research and capacity-building in vulnerable regions where the dual risks of climate change and environmental pollution are high, yet scientific attention is critically low. This proactive identification of attention gaps is helpful for ensuring that global policy responses in an era of poly-crises are guided by a more equitable and risk-informed evidence base.

Looking ahead, there is an urgent need for more geographically inclusive, mechanistically explicit, and policy-relevant research that can inform mitigation and adaptation strategies at multiple scales. Strengthening observational networks, enhancing regional modeling capacity, and promoting interdisciplinary collaboration will be essential. In particular, future studies should aim to disentangle the compound interactions between climatic drivers and pollutant behaviors, quantify their health implications, and identify high-risk areas where intervention is most needed. Only through such coordinated efforts can we advance a more complete and actionable understanding of the air quality–climate nexus. Fulfilling this research agenda is a critical prerequisite for crafting the effective, evidence-based, and equitable environmental policies needed to navigate the dual challenges of climate change and air pollution.

Method

Overview

This study addresses the interactions between climate change and air pollution by systematically reviewing an extensive body of literature. Given the sheer volume of studies on this topic, manual review would be inefficient and impractical. To overcome this challenge, we leverage advanced large language models (LLMs) for automated literature screening, and information extraction32,33. This approach enables efficient and accurate handling of the substantial number of publications, providing a comprehensive analysis across a wide range of pollutants, climate impact drivers, and geographical regions. We followed the checklist for scoping review suggested in Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) (Supplementary Table S2).

Large language models

In this study, we employed three advanced large language models (LLMs) for literature screening and information extraction: GPT-3.5 by OpenAI, Gemini Pro by Google, and Qwen1.5-72B by Alibaba. Utilizing multiple models aims to mitigate potential biases inherent in individual models, thereby enhancing the reliability and comprehensiveness of our analysis. By integrating the strengths of these diverse models, we strive to achieve more objective and accurate results.

Literature retrieval and screening Protocol

We constructed a set of 48 climate change keywords and 25 air pollution keywords (see Supplementary Table S3). By combining these terms, we performed 1200 searches on Web of Science, resulting in over 430,000 unique records after deduplication. All searches were done between October 2022 and September 2023.

To identify literature relevant to our focus on the impacts of climate change on air quality, we employed the three LLMs. Each model independently assessed the titles and abstracts to determine whether a paper was relevant. A paper was considered relevant only if all three models agreed on its relevance. This screening process identified more than 10,000 relevant papers. Additionally, we manually annotated 1300 papers to validate the accuracy of the LLM-based screening, achieving an accuracy rate of 86%. Human annotators followed the same screening protocol as the LLMs (excluding output format requirements). See Supplementary Information for details on the screening protocol (LLM prompts).

Information extraction

For the relevant literature, we utilized three large language models (LLMs) to extract three categories of information: (1) article type and methodology, (2) climate change drivers and affected air pollutants, and (3) the geographical region of the study.

All relevant articles were classified into three categories: evidence, mechanism, and discussion. “Evidence” refers to observational or model-based case studies; “Mechanism” pertains to studies exploring how climate change impacts air pollution; and “Discussion” includes review or commentary articles. Additionally, the methodologies employed in the studies were identified. Over 500 methods were applied in the related research, with the most frequently used being observation, model simulation, scenario analysis, and experiments. Technologies such as remote sensing, PMF, and meta-analysis were also employed34,35.

Climate change drivers and affected air pollutants mentioned in the articles were extracted and categorized into eight major climate change drivers and ten pollutants. This process was achieved through keyword extraction and matching, with 150 climate change keywords and 24 pollutant keywords predefined. The categories for climate change drivers are: Heat and Cold, Wet and Dry, Wind, Snow and Ice, Coastal, Open Ocean, Air Pollution Weather, and Radiation. The ten pollutants are: CO, Pb, NO2, O3, Black Carbon, Dust, Sulfate, Nitrate, Particulate Matter, SO2 (see Supplementary Table S4). For instance, the “PM” category was identified by terms including “PM2.5”, “PM10” and “particulate matter”, whereas the “Dust” category was identified by “dust”. Our counting methodology was designed to reflect the focus of each publication: if an article’s title or abstract contained keywords from both categories, it was included in the count for each. This ensures our analysis captures all relevant contributions to each topic. The correspondence between keywords and categories is detailed in Supplementary Table S5.

We utilized the Python library Mordecai to extract and geolocate place names mentioned in the articles. Mordecai is an open-source geoparsing tool that identifies place names in English-language text, resolves them to their correct locations, and returns their coordinates along with structured geographic information. This tool has been effectively employed in other language model-based studies, such as the one by Callaghan 36, to systematically identify geographical regions addressed.

Regional research sufficiency analysis

To evaluate the sufficiency of research across various regions, we developed an index \({S}_{r}\) based on the number of publications (\({{Pub}}_{r}\)) and population (\({{Pop}}_{r}\)) for each region \(r\). Using the global average of publications per capita as a benchmark, we calculated each region’s research sufficiency by measuring the difference between its actual publication count and the expected value based on population. This difference serves as an indicator of whether a region has sufficient research relative to its population.

$${S}_{r}={{Pub}}_{r}^{{need}}-{{Pub}}_{r}={{Pop}}_{r}\frac{{\sum }_{i}{{Pub}}_{i}}{{\sum }_{i}{{Pop}}_{i}}-{{Pub}}_{r}$$
(1)

If \({S}_{r} > 0\), this indicates that the research sufficiency in the region is above the global average. Conversely, if \({S}_{r} < 0\), the research sufficiency in the region is below the global average. We categorize them into two sets as follows:

$${S}^{+}=\left\{{S}_{i}\in S,|,{S}_{i}\ge 0\right\},\,{S}^{-}=\left\{{S}_{i}\in S,|,{S}_{i} < 0\right\}$$
(2)

For each element in \({S}^{+}\) and \({S}^{-}\), we calculate its percentile within its respective set, denoted as \({P}_{i}\). We then define a new indicator, the Research Attention Index (RAI), as the formular below.

$${{RAI}}_{i}=\left\{\begin{array}{l}{P}_{i}\,if\,{S}_{i}\in {S}^{+}\\ {P}_{i}-1\,if\,{S}_{i}\in {S}^{-}\end{array}\right.$$
(3)

The RAI possesses the following properties: its values range between −1 and 1, where values closer to 1 indicate higher research sufficiency in the region, values closer to −1 indicate lower research sufficiency, and values near 0 indicate a research sufficiency close to the global average. By performing statistical analyses of the RAI against various factors, we explore which elements may influence Research Attention.

Quantified synthesis framework for mechanism research

We consider the impact of temperature changes on concentrations of particulate matter (PM) and ozone (O3), specifically through its influence on emission, reaction, and dispersion processes, to be a critical area of focus24,37. This study centers on these domains, extending the analysis from title and abstract screening to a full-text review, leveraging large language models to extract mechanisms influencing PM and O3 concentrations. To ensure the authority and reliability of the sources, we selected highly-cited literature, with 40% of these documents comprising 80% of all citations.

Our analysis is thus concentrated on studies with substantial influence in the field. Full-text documents were parsed and divided into segments to facilitate academic analysis. Each segment was embedded using a vectorization method, creating vector representations stored in a database for efficient retrieval based on semantic relevance. This structured approach allowed us to isolate specific content relevant to our focus areas, while supporting detailed comprehension at a segment level.

We then utilized GPT-4-Turbo to conduct chunk-by-chunk analysis, automating judgment and reading processes while enhancing accuracy through manual review where necessary. Finally, we synthesized quantitative insights from the selected document segments, providing a comprehensive quantification of mechanisms associated with temperature, PM, O3, emissions, reactions, and dispersion, thus enabling a nuanced understanding of their interactions.