1 Background

Artificial Intelligence (AI) could revolutionize our ability to understand and address climate change. AI tasks and methods can increase the speed of problem solving with applications for better understanding the causes of climate change, responding to its impacts, and formulating solutions [1, 6, 11].

Today, scholars have begun to analyze the potential role that AI could play in addressing global climate change, both through improving our scientific understanding of the causes and impacts of climate change and by helping to develop solutions [22, 57]. We are increasingly seeing examples of how AI and machine learning can be used to improve the accuracy of climate system modeling [5], fill time series data gaps [26], estimate emissions inventories [20], refine climate scenario projections [44] and climate impact assessments [12], as well as develop applications for low carbon technology deployment through power, transportation and building system optimization [7, 8].

Multiple studies have shown that that AI simulations and machine learning are being integrated into weather and climate modeling, including emulating and forecasting weather patterns and climate processes with greater consistency, data efficiency, and improved generalization [28, 32, 36, 54, 55]. AI is used in in flood risk modeling frameworks to increase the performance and accuracy of prediction methods [43, 46, 61]. Using neural networks for weather and climate modeling has improved agriculture and crop yield predictions under a range of climate scenarios, and machine learning algorithms have been applied in areas such as monitoring soil quality, managing crops, and modeling evapotranspiration, rainfall, drought, and pest outbreaks [15, 50, 60].

AI algorithms are increasingly being used for improving the efficient management of natural resources. For example, combining deep learning with statistical techniques could create more useful assessments of the impact of deforestation on rising carbon emissions in metropolitan areas [34]. In addition, machine learning approaches are being applied in developing low carbon materials [19]; for example the application of machine learning in optimizing concrete and steel production have demonstrated how AI can be integrated into supply chain modeling for heavy industries [24, 39, 51]. AI frameworks have been applied to minimize water consumption and emissions from oil and gas reservoirs, while other research has demonstrated methods using machine learning in assessing the carbon footprint of buildings [13, 16, 29, 47].

Many studies have used AI methods in renewable energy research and demonstrated the broadening number of use cases for integrating AI into renewable energy systems. AI techniques are becoming a key tool in deploying data-integrated renewable energy networks [2, 4, 23, 37]; estimating and forecasting solar radiation resources [17, 30, 31, 38] and wind energy resources; [18, 25, 63] as well as in micro-grid management [27, 42, 58].

Additionally, AI has been shown to be a powerful tool to assess and develop carbon markets and generate more accurate carbon price models, including dynamic carbon pricing mechanisms [3], and more robust comparison models for carbon price forecasting [56]. Such methods have been applied to studies of emissions trading schemes including in China [35] and the UK [45].

While we have a sense of the general scope of climate change research being undertaken [21, 49, 52, 62], and studies have previously laid out the potential for AI to improve climate research and enable the achievement of global sustainable development goals [48, 53], no studies to date have taken a systematic and comprehensive approach to characterizing the way in which AI is intersecting with climate change research at a large scale, despite the vast demonstrated potential for AI to change the way in which climate research is conducted.

In this paper we map the community of knowledge at the intersection of climate change and AI to review how AI methods are being applied to climate related research, and which countries are leading in the application of AI to climate research. In mapping the specific AI tasks or methods being applied to specific climate research fields, we highlight gaps and identify opportunities to expand the use of AI in climate-related research.

Our analysis is based on a novel merged corpus of scholarly literature which contains millions of unique scholarly documents in multiple languages, and associated research clusters which are organized into a Map of Science. This is the first such study of the application of AI tasks and methods to climate change research using such a comprehensive data set. This paper can therefore greatly improve our understanding of both the current use and the potential use of AI for climate research.

2 Methods

In order to map the community of knowledge at the intersection of climate change and AI, we use a novel merged corpus of global scholarly literature, including Digital Science’s Dimensions, Clarivate’s Web of Science, Microsoft Academic Graph, China National Knowledge Infrastructure, arXiv, and Papers with Code, with CSET’s Map of Science.Footnote 1 This dataset allows for a far more comprehensive review than most traditional bibliometric analyses. In addition, it includes more than 120,000 research clusters derived from citation relationships within the merged corpus. Research clusters are groupings of scholarly documents based on citation links, not on topic similarity or author networks; thus, research clusters are groupings of scientific publications that address similar research questions. Each research cluster includes a set of research publications and aggregated metadata generated from the member publications, such as, key areas of research (fields and topics), key researchers in the field, and key funders.Footnote 2

We perform our analysis by identifying climate change related research papers via a keyword search, linking the publications to their research clusters, and then analyzing research clusters of interest. Figure 1 illustrates our data collection pipeline, starting with a set of keyword publications and ending with a set of research clusters and their member publications. Each dot in the map of science represents a research cluster and is colored by its broad area of research.

Fig. 1
figure 1

Data Collection Pipeline using CSET’s Merged Corpus and Map of Science

Full size image

This scientific research data pipeline enables us to find research clusters of interest by locating research publications in the Map of Science. We can then look at a subset of research clusters and analyze aggregate statistics from their member papers. This approach to identifying scientific research of interest requires a seed set of publications. We generated a scientific research corpus of climate change literature (\({R}_{climate}\)) using a regular expression search. We generated a scientific research corpus of climate change literature using a regular expression search in English and Chinese, including terms for climate change, global warming, carbon emissions and low carbon (Table 1).Footnote 3 If a publication contains one of the terms in its title or abstract it is included in our climate change publication set.

Table 1 Regular expression search terms used to generate \({R}_{climate}\)
Full size table

We ran a search through the CSET merged corpus using the terms generated above; publications were selected as being related to climate change research if their title or abstract contained at least one keyword. We based these keywords on other studies that have conducted bibliometric analysis [21]. This search resulted in 947,616 climate change-related publications, which we refer to as \({R}_{climate}\). We select RCs that contain at least one of these climate change publications, which results in 46,703 research clusters.

For each research cluster selected in this initial cluster search, we computed the percentage of papers that are contained in \({R}_{climate}\) out of the total number of papers in the RC. This allows us to sort and filter these resulting RCs based on the concentration of climate change-related papers. Our research cluster analysis for climate research includes 413,303 publications pulled from the 95th percentile of climate focused literature in our dataset which linked to 2,351 research clusters that have five percent or more \({R}_{climate}\) publications [33].

Our final filtering was through an identification of clusters with high percentages of AI-related publications. We use the AI percentage from the Map of Science, which identifies the concentration of AI-related publications in a given cluster. AI relatedness in English language publications were classified using a model trained on arXiv publications [14], and Chinese-language publications were classified using a regular expression query [10]. Thus, similarly to how we filter for climate change-related RCs, we can filter for AI-related RCs.

This allowed us to sort our dataset both by climate and AI relevance. We did this by looking at the clusters in both the 95th percentile of climate research and the 95th percentile of AI research. By selecting research clusters that have both 95% or more concentrations of climate change related publications and AI-related publications we identify 111 research clusters to analyze from the starting set of 2,351 climate change clusters.

Figure 2 displays the full Map of Science and the two subsets (climate change and climate change and AI) of research clusters we identify highlighted within it.

Fig. 2
figure 2

Climate Change and Climate Change AI Research Clusters Highlighted in the Map of Science

Full size image

In the synthesis section we discuss further methods that were used to analyze and synthesize the dataset described above. This includes extracting 67 clusters that have either China, or the U.S. listed as the top country and have on average more than 2 citations per paper to filter for clusters with community engagement, and an examination of the leading AI and climate change tasks and methods by cluster at the individual publication level as described in Sects. 3.2 and 3.3.

3 Synthesis

3.1 Characterizing the climate change and AI research landscape

In order to contextualize the landscape of climate change and AI research, we compare the general research fields and countries of publication for each research cluster set. Each research cluster is assigned a broad discipline from the following list: Biology, Chemistry, Computer Science, Earth Science, Engineering, Humanities, Materials Science, Mathematics, Medicine, Physics, and Social Science. This discipline assignment represents the majority of member papers in a given research cluster and does not indicate that every member paper falls unambiguously under this area. Figure 3 displays the percentages of climate change related research clusters by their general discipline (displaying discipline areas that have at least a 1% share of publications).

Fig. 3
figure 3

Comparison of Climate and Climate + AI Research Clusters by Discipline

Full size image

The climate research cluster set is comprised of 50% earth science publications and 43% social science publications, and includes materials science, engineering and biology publications. In contrast, the climate and AI dataset is comprised of 54% earth science and 41% social science publications, along with some engineering, computer science and materials science publications. While there is not a huge difference in fields between climate research and climate and AI research, biology drops off and is replaced by computer science in the second category as a leading field.

Articles at the intersection of climate and AI research include multiple disciplines from both the natural and social sciences. While the earth sciences dominate the research clusters identified, this is very closely followed by the social sciences. It is somewhat surprising that engineering and computer science do not show up in greater percentages in this area, likely because most climate related research is in fact not being done in these fields, but rather the models and techniques are being applied by climate researchers in their respective fields. A potential limitation of these categorizations however is that much of this work is interdisciplinary and may in fact span the natural and social sciences.

3.2 Leading countries, institutions and funders

Each research publication is assigned country data using the location of the organization that an author is affiliated with. This means that if there are multiple authors from different countries, a given publication will have multiple countries assigned. For all member publications in a given research cluster, a “top country” categorization is assigned based on the country being listed on the most publications in that research cluster. We treat all EU-27 countries as one entity due to their high rates of collaboration and research funding allocation. Figure 4 displays the top five leading countries by research cluster count.

Fig. 4
figure 4

Comparison of Climate and Climate and AI Research Clusters by Top Country

Full size image

We find that China produced more research in our climate research clusters and climate and AI research clusters, with U.S. authors producing the second highest number of research in both sets. It is perhaps not surprising given China’s role in climate change research shown here, and its strong role in AI research [41]. Yet China has a more sizable publication output lead in climate research generally than in climate and AI research. The other countries that produce significant climate and AI research outputs differ from those that produce more climate research generally. The EU-27, UK, and India follow China and the United States in climate research generally, while India, the EU-27, and South Korea follow China and the United States in research on climate and AI. It is worth noting that if results were adjusted by factors such as population size or other measures of capacity, the analysis would yield different results.

Due to the publication output lead that China and the U.S. hold, we further refine our set of 111 climate change and AI research cluster to the 67 clusters that have either China or the U.S. listed as the top country and have on average more than 2 citations per paper to filter for clusters with community engagement [33]. This allows us to examine a variety of relevant variables including: (1) leading countries of author affiliation; (2) leading research fields; (3) leading author affiliations; (4) leading funding organizations; (5) leading industry affiliations; and (6) AI-related tasks and methods; thereby facilitating a more granular analysis of the research landscape at the intersection of climate and AI.

In order to identify research institutes with the highest global publication output at the intersection of climate and AI, we examine the research institutes that the study authors are associated with.Footnote 4 The top 10 institutes are listed in Table 2.

Table 2 Top 10 publishers of research on climate and AI
Full size table

As China is the leading country by author affiliation as presented above, we see that many research institutes publishing at the intersection of climate and AI research are based in China. The Chinese Academy of Sciences, the largest research institute in China, is by far the dominant research institute where research at the intersection of climate and AI is being conducted. Within the Chinese Academy of Sciences (CAS), the leading research institute associated with climate change and AI publications in our database is University of the Chinese Academy of Sciences (438 publications), followed by the Institute of Geographic Sciences and Natural Resources (277 publications), and the Institute of Remote Sensing and Digital Earth (246 publications). Other leading Chinese research institutes include Beijing Normal University, Wuhan University, and Tsinghua University.

Within the United States, the University of Maryland, College Park has the largest number of publications in our climate and AI dataset, followed by the United States Geological Survey, University of Wisconsin-Madison, and the United States Forest Service. The two other countries with research institutes that show up in the top ten are the Netherlands and Australia.

We examine the observable leading funding organizations associated with climate and AI publications and find that China-based funding organizations have supported research that contributed to the largest number of publications, including the National Natural Science Foundation of China (4,391 publications) and China’s Ministry of Science and Technology (1,938 publications) in the first and second position. In third place is the United States National Science Foundation (1,527 publications), followed by the European Commission (998 publications) and the Chinese Academy of Sciences (710) which not only conducts but also funds research. The top ten funders are listed in Table 3.

Table 3 Top 10 funders associated with climate and AI publications
Full size table

While no private companies appear as leading research institutes or funders, we took a closer look to determine which companies are the most associated with climate and AI publications in our database. The top five companies that appear in our database in either a funding capacity or research affiliation are Google based in the United States (62 publications), Science Systems and Applications based in the United States (30 publications), State Grid Corporation based in China (30 publications),Footnote 5 IBM based in the United States (22 publications), and Volkswagen Group based in Germany (15 publications).

The Chinese Academy of Sciences (CAS) is listed in Table 2 as being associated with the largest number of publications at the intersection of climate and AI by far. However, CAS is a large organization comprised of multiple research institutes distributed throughout the country. As a result, we took a closer look at the specific research institutes within CAS to better understand their contributions to research in this area. We found that the University of Chinese Academy of Sciences is the source of the highest number of publications, followed by the Institute of Geographic Sciences and Natural Resources Research, and the Institute of Remote Sensing and Digital Earth as listed in Table 4.

Table 4 Top producers within the Chinese Academy of Sciences of Climate/AI publications
Full size table

The names of the CAS institutes give some indication of the type of research where AI is being applied to climate research, including in the areas of geographic sciences and remote sensing. More detail is available at the websites provided in Table 4.

3.3 AI tasks and methods used in climate research fields

To better understand exactly how AI is being utilized within climate research, we examined the AI-related tasks and methods that are automatically assigned to individual research publications in our database using a named entity recognition model trained on tasks and methods as developed in [59]. Each task and method label falls under several broad areas, such as “natural language processing” or “causal inference.” For our analysis, we aggregated the tasks and methods that appeared in member publications of our 67 research clusters of interest. For each RC, we looked at the top five most frequent tasks and methods from the research clusters’ member publications and represented them in nine distinct categorizations from the “Papers with Code” taxonomy: causal inference, computer vision, graphs, methodology, natural language processing, neural networks, reinforcement learning, robots, and time series [40].

Next, we manually verified nine climate-related categorization labels based on the occurrence of keywords in the research cluster metadata: climate impacts, climate modeling, emission trends, energy efficiency, energy technology, energy trends, land use change, public perception, and transportation, based in part on the categories used in [48]. We then identified all distinct pairings between the nine AI-related tasks and methods and the nine climate-related categories. For example, if a research cluster had both climate modeling and neural networks labels, that would be represented in Table 5 by a checkmark.Footnote 6

Table 5 Mapping AI Tasks and Methods within Climate Change Research Subfields
Full size table

In Table 5 we see a wide range of AI tasks and methods being applied to the 9 climate research areas that we extract from our climate and AI RC dataset. For example, we identify six AI tasks and methods being used in studies of climate impacts, including causal interference, computer vision, natural language processing, neural networks, robots and time series. Studies involving climate modeling are using at least five AI tasks and methods including computer vision, graphs, neural networks, robots and time series.

This analysis also reveals some areas of climate research that are using fewer AI tasks and methods. While energy technologies research is using multiple methods (examples include computer vision, AI methodology, natural language processing, and reinforcement learning), we see other areas of energy research such as energy trends studies and public perception studies using fewer methods. As a result, there appear to be gaps in certain climate research areas where AI tasks and methods are not being used as widely and where there may be useful applications. Exploring these gaps identified in Table 5 is an area for future research.

4 Discussion and conclusions

Given the vast potential of AI tasks and methods to revolutionize all aspects of research and analysis, it is not surprising that they are being applied to one of today’s most pressing global challenges, addressing climate change. Our study contributes to the understanding of how AI is being used in climate related research with three key findings.

First, we find that articles at the intersection of climate and AI research include multiple disciplines from both the natural and social sciences. While the earth sciences dominate the research clusters identified, this is very closely followed by the social sciences. It is somewhat surprising that engineering and computer science do not show up in greater percentages in this area, likely because most climate related research is in fact not being done in these fields. A potential limitation of these categorizations however is that much of this work is interdisciplinary and may in fact span the natural and social sciences.

Second, we find that Chinese research institutions lead the world in publishing and funding research at the intersection of climate and AI, followed by the United States. In examining the research institutes that the study authors are associated with, we find that just as China is the leading country by author affiliation as presented above, many of leading research institutes at the intersection of climate and AI research are based in China. The Chinese Academy of Sciences, the largest research institute in China, is by far the dominant research institute where research at the intersection of climate and AI is being conducted. We also find that the leading funders associated with climate and AI publications are also based in China: The National Natural Science Foundation of China and China’s Ministry of Science and Technology. China’s dominance in AI applications has been well documented, and we show that China also leads the world in climate released research, as well as at the climate-AI interface. This is also reflected in Chinese government policy; for example, the Chinese government has issued explicit guidance on the use of AI in climate research in the “Meteorological Science and Technology Development Plan (2021–2035)” issued by the Ministry of Science and Technology and Chinese Academy of Sciences in March 2022 [9].

Third, by mapping the specific AI tasks or methods being applied to specific climate research fields, we find gaps and identify opportunities to expand the use of AI in climate research. While we believe this is the first study to examine this in a systematic way, we acknowledge some deficiencies in our methods, namely that we manually identified subfields in climate research using some keyword analysis as well as some subjective judgement, and that our pairing of AI-related tasks and methods to climate-related research areas represents the occurrence but not the frequency of these pairings. However, our findings raise multipole questions that present opportunities for future research and inquiry, including why certain tasks and methods are being used in specific fields, and what other fields might learn from applications to date.

Of course, any effort to make broad generalizations about fields as vast and complex as the fields of climate change and AI comes with some limitations. There are likely applications of AI to climate research that are not included here due to limitations in our original search terms or in the way in which we develop climate subfields in order to map them against AI tasks and methods. These are rapidly involving fields of research in which new methods and applications are being developed all the time. Furthermore, the field of research at the intersection of AI and climate change is growing very rapidly, so any attempt to assess the state of the field could be quickly outdated.

Yet given the tremendous opportunity that emerging AI tools provide in addressing a challenge so vast and multifaceted as climate change, the study of their application is no doubt of tremendous academic and practical importance. This paper allows for a more globally comprehensive and nuanced analysis of this relationship than past studies and consequently provides a tangible contribution to our broader understanding of the use of AI tasks and methods in climate change research.

This study also examines the role of specific countries and specific funding organizations in shaping the direction of climate and AI research which will be increasingly important to understand. Furthermore, tensions between China and the West are already shaping national decisions about investments in AI research and could influence future research directions.

Given the very limited time remaining to avoid even more dangerous impacts of climate change globally, the expanded use of AI tasks and methods presents the opportunity to transform our ability to understand and address climate change. This paper helps to identify opportunities to expand the use of AI tasks and methods in climate related research, and the predominance of China and the United States in this area raises important questions about national leadership and competitiveness.