In July 02023, climate scientist Bjorn Stevens paused in the middle of a virtual talk about data and modeling to quote a short story by Argentine author Jorge Luis Borges. Borges’ work does not often come up in the context of climate modeling — but Stevens invoked him for a reason beyond idle literary fascination. Referencing “On Exactitude in Science,” Borges’ single-paragraph story in which an empire creates maps of ever-increasing size, until they perfectly match the territory they are mapping, Stevens warned of a looming crisis of scale occurring in the field of climate science. In Borges’ story, the map becomes useless in its magnitude; in climate science, the mad dash to reach kilometer scale climate simulations could make climate models too large to matter to the world outside of academia. How much data is too much data? What will it really mean to model the climate on 1-km scales?
Researchers are actively pursuing the answers to these questions, and CMIP7, the latest phase of the Coupled Model Intercomparison Project and the state-of-the-art for global climate models, will continue a trend of increasing resolution and dataset size. As new developments in computing and data storage advance the capabilities of the field to capture more detailed understandings of Earth systems, climate models like those comprising CMIP7 have the potential to become even larger leviathans: petabytes of data, tens of millions of lines of code, years of constantly running simulations, all to model one future pathway. Stevens’ reference makes clear a countercurrent running through the field, as urgency to measure, simulate, and respond to climate change compounds, while traditional methods to increase resolution offer diminishing returns. It’s a current that seeks to move climate science out of the stuffy halls of academia and into the hands of policymakers, businesses, journalists, and laypeople. Climate science goes public. But these changes to the field surface new entanglements with social and political forces that mark it as a contested discipline in which present developments will greatly influence future collective decision making.
Climate science is notoriously complex, with models containing upwards of the equivalent of 18,000 printed pages of code and consuming the energy equivalent of over 6,000 households per year to complete. CMIP7, the largest coordinated modeling project to date, consists of a multi-year effort to standardize, document, and coordinate across research areas including cloud patterns, various regional phenomena, land use scenarios, paleoclimate, and geoengineering — just to name a few. The data that scientists, researchers, and technicians feed into climate models adds up to quantities at the order of millions of gigabytes, costing hundreds of millions of dollars.
As a researcher of climate models without formal training in climate science, this means that I’ve spent months poring over the minutiae of atmospheric halocarbon concentrations, upwelling longwave radiation, climate reanalysis, and much more in order to grapple with the sheer volume and dizzying depth of the Earth System Models (ESMs) I work with every day. Increasingly, as data about the behavior of Earth systems increases and advances in the methods for synthesizing said data add nuance, it takes teams of experts to explain the details of how we know the extent to which climate change is occurring. Due to ESMs’ substantial run time and the limited availability of supercomputers in modeling centers, ESMs are typically only used to model a particular set of Shared Socioeconomic Pathways (SSPs), so as to be most useful for IPCC reports. But SSPs only account for a narrow view of potential societal and demographic changes, and ESMs typically do not include aspects of climate science such as pre-1850 anthropogenic experiments and methane emissions from thawing permafrost.
In July 02023, leaders and experts from the fields of climate science, technology, and sustainability convened in Berlin to address these issues, among others, for the Earth Virtualization Engines (EVE) Summit. It is EVE, a new international initiative coordinating and governing a digital commons of climate data and projects, that Stevens was positing as a solution to the crisis of scale in climate science, illuminated by his reference to Borges. Technology providers are interested, too, with Nvidia offering to integrate efforts by scientists and research institutions into their product Earth-2: an ambitious attempt to create digital twins of the planet. Digital twins virtually re-create a process or object such that its behavior can be simulated. Typically used to simulate consumer products or components of infrastructural systems, the digital twin ascends to new heights with Earth-2, a constantly running simulation of Earth systems distilled down into a single, unified data product.
These twins would differ from standard climate models in a few ways. The primary mode of interaction with a digital twin is via a graphical interface, rather than a code editor, utilizing personal computers rather than supercomputers. Only a few clicks on a digital twin interface, as previewed by Nvidia, will simulate weather conditions at 1.25km resolution on a global scale — a level of granularity rarely found outside regional climate models and ensembles. What’s more, Nvidia proposes to dynamically simulate Earth systems on temporal resolutions from the daily to the hourly, as opposed to the scenario-based modeling that shows up in the IPCC reports of today. In theory, an Earth digital twin will capture the entirety of interconnected atmospheric, oceanic, and geologic systems for ongoing manipulation and visualization — an ultimate simulation of our planet.
Climate change is everyone’s problem; a planetary conundrum rather than a scientific issue.
The lofty goals outlined in the EVE Summit statement offer an appealing view of a more public climate science. A dynamic visualization of the planet like the one showcased by Nvidia CEO Jensen Huang at the summit takes the immense complexity of climate data and makes it into something more intuitive to those not already steeped in the field. Whereas previously the user of climate modeling outputs might be a professor, graduate student, or author of an IPCC report, supporters of Earth digital twins purport that non-experts may soon be able to interact with and explore changes to the climate system, “on your PC, your tablet, or your phone.” The implication here is a deprofessionalization of climate science, a technological corollary to the idea that climate change is everyone’s problem; a planetary conundrum rather than a scientific issue.
Floating in a Sea of Data
A day in the life of a climate scientist differs depending on the person, but generally it looks like data: millions of rows of spreadsheets, Jupyter notebooks galore. It makes my head spin each time I encounter a new massive and inscrutable dataset, and learn how essential it is to making our models a little bit better. One colleague of mine described their work as “floating in a sea of data.” They meant it both metaphorically and literally, as they split their time between doing data analysis and studying the historical climate record by collecting sediment cores from the ocean floor. But opening up climate science to a broader audience means moving the visual interfaces of the field from the spreadsheet closer to the sci-fi interface — from staid rows and columns to an interactive, dynamic globe. This was on full display during the EVE Summit, when Huang demonstrated the capabilities of Earth-2 without displaying a line of code or spreadsheet entry. Instead, the high-production visuals showed swirling wind currents over satellite imagery of the earth. Then, they zoomed into Berlin, to show how those same wind currents flowed within a three-dimensional model of the city’s urban landscape. The sea of data is still there, of course. It’s going to take an immense amount of data and code written by climate scientists to power the sleek interfaces of an Earth digital twin, not to mention teams of designers and product managers to make the simulations come to life. But it’s obscured now; and in that obfuscation, the spinning globes of Earth-2 indicate a change in who climate science is for.
With developments such as Earth-2, the “user” of a climate simulation no longer needs a training in atmospheric physics or ocean currents, but can instead direct their attention towards implications for policy, urban planning, or infrastructure development. More public climate science means tools like Earth-2 that serve as a tool for action, for making decisions about where to build a direct-air capture plant, how to reduce emissions from public transportation, and when to anticipate the “unprecedented” weather that characterizes living in the midst of climate crisis. This move to deprofessionalize climate science, whether through a digital twin or otherwise, is a welcome one, as it signals an understanding that positioning ourselves as planetary inhabitants is an essential part of making responsible decisions on collective scales, from the local to the regional to the global.
It also matters what worlds make up our simulations, and what simulations shape how we live in our world.
Earth digital twins signal a future in which climate models become infrastructure for our collective knowledge — simulations serving as not just scientific tools but essential components of how we understand ourselves and act together. Said infrastructure will not only inform global challenges; local, regional and state decision making will rely upon simulations to ensure that the planetary system remains in homeostasis. Yet this shift in paradigm is far from a given. The EVE proposal will only succeed with billions of dollars of funding and an unprecedented level of international cooperation, bringing together industrial titans like NVIDIA and multiple national governments. Where the money comes from and which actors will step up to provide support is still unclear — EVE may very well turn out to be another failed proposal on the long road to understanding climate change. Regardless of whether or not EVE in particular succeeds, the ambition towards deprofessionalized climate science feels like something of a fait accompli — a socio-technological shift that promises to become the bedrock upon which we stand as planetary inhabitants.
But that’s exactly what this is: a promise. Becoming infrastructure marks a shift for climate science: from a particular and limited field within earth science to an increasingly contested zone of planetary knowledge-creation and knowledge-instantiation. Following feminist technoscientific philosopher Donna Haraway, “it matters what stories make worlds, what worlds make stories.” It also matters what worlds make up our simulations, and what simulations shape how we live in our world.
Agreements over what constitutes our world rarely come easily. The ways that we collectively resolve conflicts over emissions data, geopolitically-sensitive data, and statements of culpability will dictate whose Earth is being twinned, and who gets a say in the future of global decision making. Who gets to model the Earth? And what is done to resolve disagreements when governments refuse to release verifiable emissions information? What happens when the climate impact of running these simulations continues to grow? What will make an Earth digital twin spur more action than an already-constrained IPCC?
A Dance of Competing Models
There’s no easy answer to these questions, but the consequences of them will influence how we understand ourselves within a planetary system throughout the long now. To understand these potential consequences and the tensions we face now to resolve them, I posit three experiments in thinking about climate modeling: adversarial modeling, recursive modeling, and actionable modeling. In response to the contested nature of an expanded climate science, thinking through these experiments means pre-emptively addressing what will become dominant questions in how we know what we know about climate change.
Adversarial modeling questions how competing simulations of Earth systems change the dynamics of climate science. Climate science becoming the infrastructure of collective decision making means that governments and geopolitical entities have differing incentives to share or obfuscate data about their climate actions. Certain elements of assembling a climate model are outside the purview of state actors: carbon dioxide levels, remote sensing. But considering the economic benefits of obfuscating a coal power plant, for instance, from an aerial view, means considering that the factual basis upon which an Earth digital twin is constructed may never be assured. Sources of climate data range from ostensibly-neutral remote sensing datasets to highly-biased private sector disclosures to often-questionable state emissions numbers, all of which contain their own uncertainties and possibilities for obfuscation. As such, we can anticipate a future of adversarial climate modeling, in which differences in measurement are leveraged to de-legitimize other climate models, for the sake of avoiding culpability.
In this dance of competing models, a singular source of truth appears unattainable. A public, infrastructural climate science must contend with information asymmetries and contested sources of data. Identifying and verifying the data of climate science is sometimes possible, as coalitions such as Climate TRACE demonstrate, but it is never guaranteed and always ongoing. As STS scholar Bruno Latour has argued, we now live on different planets, and these changes are unlikely to reverse themselves. Earth digital twins will only be successful in becoming a trusted infrastructure of collective decision making if they allow for and clearly describe the differences in sources of data that undergird their simulations. Interpretability is paramount; for climate science to make the leap to a more public sphere, it must address the ambiguous nature of its data sources and uses. Adversarial climate modeling guides us as we consider what form(s) future knowledge on climate change will take, and as we address the challenge of tackling a planetary crisis with asymmetrical impacts.
But climate science is not a map, and this is no fantastical speculation.
Increasing the scope of climate science means increasing the resources the field consumes. As such, future climate models will encounter a challenge of recursive modeling — simulating the climate impacts of the model itself. Climate scientists gathering data on such a scale as to induce recursive modeling resemble the cartographers in “On Exactitude in Science”. But climate science is not a map, and this is no fantastical speculation. Digital technologies already have a growing carbon footprint larger than the airline industry. Water usage for cooling data centers threatens already-constrained ecosystems. In becoming infrastructure, climate science must contend with its own climate impact, which includes modeling the model itself. As a climate model simulates the world, it must simulate itself; an ongoing dynamic model, such as the one discussed at the EVE conference, must continually undergo this recursive action. This is an issue for complex systems of all sorts, as seen in a recent uptick in emissions as a result of the boom in large language models (LLMs).
While some recent developments in climate modeling, such as emulators, may serve to reduce the resource guzzling computational needs of a massive climate model, initiatives such as EVE must address their climate impact or risk further adding to the problem they hope to solve. Thinking through recursive modeling means considering how an infrastructural climate science will need to take measures to address its own climate impact even as it supports actions to decarbonize.
Moving climate science outside of academia and towards more public stakeholders presents a third proposition of actionable modeling. In this case, climate modeling undergoes a slight shift as it focuses more on the actions that are facilitated by a model, rather than the overall understanding of the system. While these goals are far from incompatible, a more actionable climate modeling is one in which outcomes of specific actions, such as the building of a new direct-air capture plant or an expansion of a city’s public transportation network, are expressly integrated into the system and interface. Huang notes that EVE “envisions a world where everyone knows how climate and climate change affect them, and where this knowledge empowers them to act.” But how does a climate model become empowering? The more a simulation is altered to become actionable, the more it reflects the desires and biases of its creators. Will an actionable climate model project the benefits of solar geoengineering, or will it demonstrate how reduced meat consumption would alter the air quality of a region? Of course, a climate model can do both, but the developers behind an Earth digital twin must contend with the fact that each design decision will have effects in how the product influences action. Actionable climate models have the power to expand the science behind a wider array of potential climate actions, but will always carry with them a bias in how and for whom the actions they model are displayed.
Simulating the deeply entangled Earth systems that comprise climate entails a gargantuan effort of coordination and computation. The warning of Borges’ overzealous cartographers and the promise of Huang’s Earth digital twins both speak to a tension in contemporary climate science. As urgency to take action on climate change accumulates, climate modeling will play an essential role in collective decision making. Considering how a deprofessionalization of climate science presents new provocations for adversarial, recursive, and actionable modeling entails an intentional contemplation of the future of our entanglement with Earth systems.
Like many of my colleagues, I was drawn to work in climate science by the urgency and immensity of these provocations. I don’t want to end up like one of Borges’ cartographers, pursuing something grandiose and ultimately ineffectual. Instead, I want a climate science that anticipates geopolitical differences, refuses to pursue scale for the sake of scale, and that facilitates more ambitious climate actions. The future of climate simulations is entangled with the future of our species’ relationship to Earth systems. Let this be an opportunity to take that entanglement seriously.
