Tracking artificial intelligence in climate inventions with patent data

Artificial intelligence (AI) is spreading rapidly in many technology areas, and AI inventions may help climate change mitigation and adaptation. Previous studies of climate-related AI mainly rely on expert studies of literature, not large-scale data. Here I present an approach to track the relation between AI and climate inventions on an economy-wide scale. Analysis of over 6 million US patents, 1976 to 2019, shows that within climate patents, AI is referred to most often in transportation, energy and industrial production technologies. In highly cited patents, AI occurs more frequently in adaptation and transport than in other climate mitigation areas. AI in climate patents was associated with around 30–100% more subsequent inventions when counting all technologies. Yet AI-climate patents led to a greater share of citations from outside the climate field than non-AI-climate patents. This suggests the importance of tracking both increased invention activity and the areas where subsequent inventions emerge. Rapid growth of AI could lead to more inventions and innovations in climate actions, yet evidence of this connection is lacking. The use of large-scale patent data and automated techniques helps elucidate trends in climate-related artificial intelligence inventions for different technology areas.

A range of artificial intelligence (AI) technologies are rapidly being developed with high expectations of technological innovation and economic growth [1][2][3] . AI could contribute to increasingly effective climate change mitigation and adaptation technologies in multiple areas [4][5][6] . However, an increasing capability to automate and transform production, equip industries with new tools and draw increasing commercial support also means that AI technologies could lead to a higher demand for computing power, larger carbon footprints, shifts in patterns of electricity demand and an accelerated depletion of natural resources [7][8][9][10] . High expectations of new technologies with limited experience suggest a risk of unjustified techno-optimism, which could delay effective climate policy 11 . Whether the net effect of AI on the climate system will be ameliorative or detrimental is currently an open question, and concerns about the impact of AI have been followed by calls for new regulations and increased international oversight [12][13][14][15][16] . This suggests a need for improved capabilities to track, examine and analyse these emerging technologies. Here I use large-scale patent data to track AI inventions in technologies that can contribute to climate adaptation and mitigation.
The initial research into the connection between AI and climate change has often been framed in terms of the United Nations Sustainable Development Goals and conducted as expert studies. These have indicated both positive and negative effects of AI 4,[17][18][19] . For climate change, it has been suggested that machine learning could have broad potential in both mitigation and adaptation strategies, with a mixed message regarding the potential net effect on the climate system [20][21][22][23][24] . An advantage of expert-based reviews is the possibility to integrate knowledge from different domains, even when data are scarce. However, experts often find it challenging to unpack and fully explain their partially automatic judgement processes 25 . Moreover, expertise tends to be difficult to translate from one domain into another 26 . Scaling up to cover a larger literature is a challenge for any team of experts, and an interesting option would be to complement the analysis with other data sources. Here I investigate how to use large data sources from national patent offices and intellectual property organizations regularly used to monitor inventions and innovation in large economies.
Patents are possibly the most detailed track record of modern technological inventions [27][28][29] , allowing individuals and organizations to protect the use of their patented inventions typically for years. National patent offices have organized and classified millions of patents using international classification systems. The resulting classification codes provide the primary means to group patents and make them searchable: patent offices need to examine the prior art to judge whether claimed inventions are sufficiently novel before granting patents. Patents have previously been used as a data source and a proxy to study trends in both AI 1,2,30 and climate technologies [31][32][33][34] separately. Here I combined classification data for climate inventions and AI technologies to find patents that are both. First, I used the Y02 States, patent claims that include abstract inventions for algorithms and computer software require a link to a practical application 37 . In the European Union, patents with abstract invention claims need to have a technical character-for example, controlling some physical process or providing an implementation or function that solves a particular technical problem 38 . The differences between what patent laws permit can sometimes be subtle, so patterns based on patentability might be specific to a country under consideration. Here I used data from the United States, which should be seen as one case study, although an interesting one. The approach used here is possible to extend for analysing patents from other countries. Besides variation in national patent laws, other factors could need to be controlled when investigating and comparing patenting trends. Such differences include technical expertise and specialization in countries and industries 33 .
Second, inventors do not always apply for patents, even when patenting is possible. Some AI inventions are being made available through an open-source culture where inventions are neither patented nor protected by secrecy, which can be seen in academia and public-interest AI research. However, an initial investigation of patented classification system initiated by the European Patent Office 35 to monitor selected technologies related to mitigating or adapting to climate change 31 . Second, I found AI patents with a recent method developed by the World Intellectual Property Organization (WIPO) that can be automated computationally 1,36 . The WIPO method classifies patents as AI on the basis of patent classification codes and by matching certain keywords from key sections in the patent texts, including terms such as 'machine learning', 'deep learning' and 'natural language processing'; more details can be found in the Supplementary Information. Third, I combined both of these classification approaches to find patents that are labelled as both. A few example patents that are classified as both AI and climate inventions are referred to and presented in the Supplementary Information.
As useful as patent data can be, it is also essential to understand some of the limitations of using patents and avoid unwarranted generalization from patented inventions to the population of all inventions, for reasons that follow. First, certain types of inventions may not even be possible to patent: it is currently not possible to patent entirely abstract inventions (for example, pure mathematical results). In the United  in a can be seen as exponential growth in climate AI patents (linear on a log scale), and this holds for climate patents and AI patents separately. Within climate patents, however, AI patents are not growing exponentially: AI is associated with an approximately linear growth in shares (Supplementary Information). c,d, Transportation, energy and industrial production mitigation technologies have accumulated the most AI patents, while the smaller classes of energyefficient ICT and adaptation patents have larger shares of AI inventions. Energyefficient ICT cover inventions that reduce energy use within ICT equipment, but not ICT used to reduce energy use in a further piece of equipment. For readability, I have shortened the official names 35 . The official names and classification codes can be found in the Supplementary Information.

Article
https://doi.org/10.1038/s41558-022-01536-w AI inventions shows that these are increasingly related to commercial patent rights, and at the same time the share of patents that depend on public government support is getting lower (Supplementary Information). Furthermore, firms and individuals also choose to protect some inventions with secrecy instead of filing patents. The incentives for secrecy vary between technological fields. For example, in the United States, the pharmaceutical and biomedical industries with high cost, high uncertainty and long innovation cycles rely more on patents than the software industry 39 . Software inventions often have lower costs and can have innovation cycles on a timescale of days or weeks rather than months or years. When patenting processes take longer than the innovation cycle, patenting might lose some of the appeal of getting expected rewards from innovation. For these reasons, a share of AI inventions can be expected not to be found in patents. To the best of my knowledge, this share is unknown and is a knowledge gap in the literature. AI technologies are being invented and used in various industries 1,2 , so this share for AI probably depends on the incentives in several technological domains that may differ. That we do not know the share of inventions that are protected by secrecy suggests a need to be cautious about generalizing from patents to other non-patented inventions. The data are as follows. First, I collected historical data on over six million granted patents publicly available 40,41 from the US Patent and Trademark Office for the period from 1976 to 2019, up to when the WIPO method for finding AI patents was developed and evaluated. I worked with data from the United States because it is a leading economy and because the US institutions have made patent full-text data readily available 41 . Previous work 42 indicates that US patents have been found to well represent the frontier of technological innovation in low-carbon energy innovation, which is part of the scope. Second, I extracted technology classification data for the patents, including current labels for climate inventions 43 : the Cooperative Patent Classification (CPC Y02) code "covers selected technologies, which control, reduce or prevent anthropogenic emissions of greenhouse gases in the framework of the Kyoto Protocol and the Paris Agreement, and also technologies which allow adapting to the adverse effects of climate change". 35,44 Third, I applied the WIPO method 1,36 described above to label the same patents as AI or non-AI for further analysis. For details about the classifications, see the Methods and the Supplementary Information. Precise summary metrics for AI and climate patents are also found in the Supplementary Information.
It is natural to first look at the data by aggregating mitigation and adaptation technologies, but then later disaggregate these into separate groups. Both climate and AI patents have seen clear growth in the past decade (Fig. 1a). AI and climate invention patent counts, as well

Fig. 2 | Comparing AI and non-AI citation counts in climate patents (citations to patents granted between 2010 and 2017)
. AI patents are on average associated with more subsequent inventions when counting citations from all subsequent technologies. The total count of AI breakthroughs is smaller, as expected because of fewer AI patents overall. In each plot, the centre line is the median, and the red square shows the mean. The dashed vertical lines show the boundaries for breakthroughs (for forward citation counts above the 99th percentile) over all years. The box limits show the 25th and 75th percentiles, with whiskers at the 5th and 95th percentiles. For plotting, a log-plus-one transformation was used.
Article https://doi.org/10.1038/s41558-022-01536-w as the climate AI patents, have undergone exponential growth during the past decade (Fig. 1b). However, note that this does not mean that the share of AI within climate patents is growing exponentially: the growth of the share of AI within climate patents has been approximately linear, and AI climate patent counts are actually lower than expected if AI and climate innovations had been statistically independent (Supplementary Information). More than half of all AI inventions in climate patents since 1976 are found in technologies for transportation, energy and production (Fig. 1c). Climate adaptation and building/housing mitigation technology patents involving AI are somewhat lower in absolute numbers. For waste management and carbon capture/storage, there are very little data on AI. Energy-efficient information and communication technologies (ICT) and adaptation patents are areas where AI has had larger overall shares in the past few years (Fig. 1d). The number of unique patents found to involve both climate and AI between 1976 and 2019 is 4,390. This is around 1.5% of the total climate patents and 2.7% of the AI patents.
To examine whether AI makes a difference in climate inventions, I chose to analyse the number of citations from subsequent patents that cite back to previously granted patents. For a given patent, the forward citation count reflects the number of subsequent patents that relate to or build on it-or, put differently, cite it. Forward citation counts have been considered to be important indicators of the technological impact of a patented invention 45 . Harhoff et al. 46 found that the economic value of individual patents, measured through a survey with patent assignees, is positively correlated with the number of forward citations. Hall et al. 47 also showed that the number of forward citations per patent correlated positively with the market value of firms, and they estimated that if a firm's quality of patents increases so that the patents receive one additional citation, on average, the firm's market value increases by 3%. Moreover, forward citations are positively correlated with patent assignees' willingness to pay renewal fees 48 , which indicates the economic value of cited patents. Furthermore, forward citations can also be used to investigate knowledge spillovers, or how knowledge from technologies in one area is useful in different areas 49,50 . In the analysis that follows, I distinguished between technological domains that cite back to previously granted AI and non-AI patents. Finally, forward citations have been used to investigate highly cited technological breakthroughs by using the accumulated forward citations in the years after which a patent was granted. Squicciarini et al. 51 define breakthrough inventions as the top 1% cited documents for each year, and they use a three-year window from the patent grant date to accumulate forward citation counts. Benson and Magee 27 constructed a metric that they term "immediate importance" as the average number of citations that a patent receives within three years of publication. Consistent with the previous literature, I examined the predictive difference associated with AI on a three-year horizon after patents have been granted. The breakthrough inventions were defined to be the top 1% cited patents in a technical domain per year.
To examine whether AI is associated with a difference in forward citations, it is natural to distinguish between AI and non-AI in the groups of climate patents described in Fig. 1. An initial exploration revealed two aspects about the target variable, shown in Fig. 2. First, the average   forward citation count for climate AI inventions is greater than for climate non-AI inventions. Second, zooming in on the highly cited breakthroughs (the highest counts in Fig. 2), a large majority of the most highly cited breakthroughs appear to be non-AI inventions. In other words, AI patents are related to more subsequent inventions on average but seemingly fewer highly cited breakthroughs. However, this does not address the fact that AI has a much smaller share of the total patents. This suggests estimating the predictive difference of AI on the average forward citation counts by including controls and testing for differences in breakthroughs after adjusting for the group size. However, carbon capture/storage and waste technology patents were left out, as the number of AI patents is too small for reasonable statistical analysis.
To estimate the predictive difference of AI on patent forward citation counts, I used count regression modelling of the forward citations on a three-year horizon, limited to patents granted in the previous decade (a three-year horizon, so patents from 2010 to 2017). I controlled for the difference in year, technological areas that patents are from and other factors in line with previous work on modelling forward citations for patents 29 (for details, see the Methods and Supplementary Information). To use control groups for wider context, I repeated the regression modelling for the climate invention areas and control groups given by similar technological domains (based on CPC classifications). The control groups have related technological functions but broader than climate inventions: buildings, electricity, smart grids and transport technologies in general, not restricted to climate relevance.
For the climate patents as well as the control groups, AI seems associated with more subsequent inventions even after controlling for other factors (Fig. 3). In groups of climate inventions, AI was associated with a 30-100% increase in forward citations, with predictive differences being statistically significant. Among the technologies with climate inventions, buildings and energy-efficient ICT showed the greatest increases related to AI, with adaptation and energy technologies on average showing a weaker difference. For the control groups, the ranking and effect sizes are similar to the groups with climate inventions: the coefficients in Fig. 3 are in a similar range as those found for related technologies such as electricity, transport and building technologies in general. An analysis for other control groups less related to climate inventions showed that differences between AI and climate inventions can be stronger or weaker than in other technological domains (Supplementary Information). So far the results show that AI inventions in climate patents were on average cited more from the group of all subsequent patents. However, this does not show us in which technological domains these subsequent inventions emerge. The idea that inventions can benefit some technologies by supporting subsequent inventions in certain technological areas more than in others can be framed in terms of knowledge spillovers 50,52,53 . In this case, one can examine whether AI patents get cited from technology domains that are either climate inventions or not. I examined whether AI and non-AI patents get cited to any different degree by climate patents as follows. For cited patents, I distinguished whether these are AI or non-AI climate innovations. For citing patents (the spillovers), I distinguished between climate and non-climate inventions. Aggregate patent citation count networks between technology domains are known to depend on several factors, such as technology domain size and average age 54 . Recent work 53,54 has shown that it is possible to control for domain size and average patent age and to test whether differences would arise just because of random patent citations with a null model based on the hypergeometric distribution. 53 The results show that AI in climate patents is associated with a smaller share of spillovers to climate inventions than spillovers from other non-AI technologies (Fig. 4a). Within patents in general, it would be reasonable to expect AI spillovers to primarily be related to AI technologies rather than climate inventions (Fig. 4b). However, within climate patents, a larger share of spillovers than non-AI technologies are non-climate, suggesting that knowledge from the climate AI patents was more useful in other areas. Similar results hold when disaggregating patents into groups of climate adaptation and mitigation (Fig. 5). Taken together with the results above, this suggests a double association for AI in climate inventions: AI has been related to increased activity in subsequent inventions but also a smaller share of spillovers to climate patents than from non-AI technologies. Citation counts cannot be expected to always reflect the actual usefulness of individual technologies in practice, and citations are one of several ways to track the overall importance of new technologies. Therefore, tracking AI in climate patents will require distinguishing between the direct impact from more subsequent inventions and how knowledge spillovers are distributed between areas to improve our knowledge about the net effects of AI inventions.
Finally, I examined whether AI has any relation to highly cited patents, also termed as breakthroughs. An association between AI and highly cited breakthroughs would indicate where applications of AI have been more interesting or particularly valuable. Figure 2 shows fewer highly cited AI breakthroughs in total, possibly because of the smaller number of AI patents overall. I considered in each group of climate inventions the 1% patents per year with the most forward citations in the three years following publication. I then took the cumulative experience of AI inventions as the total count of AI patents in the area to compare technological domains of different sizes and the accumulated experience of patenting with AI.
Most groups of climate mitigation technologies have been associated with similar AI breakthrough shares (defined as the number of AI-related breakthroughs per AI invention in that technology) as other technologies (Fig. 6). For adaptation technologies, AI breakthroughs were initially higher compared with other technologies; among mitigation technologies, transport is clearly leading in shares. A quantile test (Methods and Supplementary Information) also suggests that the share of AI breakthroughs has been higher in climate adaptation and transport technologies. Estimates for the other groups are more uncertain: wide confidence intervals (Supplementary Information) suggest that the current evidence is too weak to strongly rule one way or the other about AI breakthroughs in most areas of climate mitigation, besides transport patents and in contrast to climate adaptation technologies. For most groups of climate mitigation technologies, the uncertainty means that the jury is still out with respect to the role of AI in breakthroughs.
Taken together, the results show that AI in climate patents is associated with more subsequent patents but also a larger share of knowledge spillovers to non-climate technical domains than non-AI technologies. This suggests that the analysis of AI in climate inventions should also consider the impact on other technological fields that may benefit more from these inventions than from non-AI climate inventions. AI has been associated with a higher share of breakthroughs in climate adaptation and transportation patents than in other groups of technologies with a potential for climate mitigation. These results are for one country, but the approach can be used to study other countries and regions. Caution is also needed to avoid unjustified generalization to inventions and innovations beyond those covered by patents and being aware that the criteria for patentability can vary between countries.
New inventions and technological breakthroughs may meaningfully contribute to addressing climate change. More capabilities are necessary to track the emerging technologies for which both risks and promises exists but where the use of large-scale data is still scarce. Using patent data, we can better track AI in technologies to adapt and mitigate climate change.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41558-022-01536-w.