Coupled Model Intercomparison Project

Based on Wikipedia: Coupled Model Intercomparison Project

The Climate Models Fighting for Your Future

Thirty different computer programs are trying to predict your future. They disagree with each other—sometimes by a lot. And yet these disagreements might be the most valuable thing about them.

This is the story of the Coupled Model Intercomparison Project, known as CMIP. It's an awkward name for an elegant idea: what if, instead of letting each country's climate scientists work in isolation, we forced them all to answer the same questions with their models, then compared the results?

What Exactly Is a Coupled Climate Model?

Before we dive into the intercomparison part, let's talk about what's being compared.

A climate model is a computer program that simulates Earth's climate. Think of it as a virtual planet running on supercomputers. The program divides the atmosphere into a three-dimensional grid—imagine stacking thousands of boxes on top of each other around a digital Earth—and then calculates how energy, moisture, and momentum flow between those boxes over time.

The "coupled" part is crucial. Early climate models only simulated the atmosphere, treating the ocean as a static boundary condition—like modeling weather patterns while assuming the ocean is just painted on. But that's not how Earth works. The atmosphere and ocean are in constant conversation. Warm air heats ocean surfaces. Warm oceans pump moisture into the atmosphere. Ocean currents redistribute heat across the planet. Ice sheets reflect sunlight back to space, cooling the atmosphere, which affects the ice sheets.

A coupled model connects all these pieces: atmosphere, oceans, land surfaces, ice sheets, and sometimes vegetation and chemical cycles. Each component talks to the others. The atmosphere asks the ocean, "How warm are you right now?" The ocean responds, and then the atmosphere adjusts its calculations accordingly. This back-and-forth happens every simulated hour (or less) as the model steps forward through time.

Building one of these models is a monumental undertaking. A major modeling center might have dozens of scientists working for years to develop, test, and refine a single coupled model. By the early 1990s, perhaps a dozen groups around the world had working coupled models. Each group made different choices about how to represent clouds, ocean mixing, ice dynamics, and countless other processes. Each model embodied different theories about how the climate system works.

The Birth of Organized Comparison

In 1995, the World Climate Research Programme decided to do something unprecedented. Instead of letting each modeling center publish papers about their own model in isolation, they would organize a systematic comparison.

The Working Group on Coupled Modelling, the committee that oversees this effort, established what became the Coupled Model Intercomparison Project. Every participating modeling center would run the same experiments. They would all simulate the same time periods, with the same forcing scenarios, and submit their results to a central archive.

This central archive lives at Lawrence Livermore National Laboratory in California, managed by a group called the Program for Climate Model Diagnosis and Intercomparison. They store the outputs from all participating models and make them available to any researcher who wants to analyze them.

The first phase, CMIP1, attracted eighteen modeling groups—essentially every major center in the world with a working coupled model. Scientists were hungry for this kind of comparison.

Why Disagreement Is the Point

Here's something that might seem counterintuitive: the goal of CMIP is not to find the "right" model. The goal is to understand the full range of possible climate futures.

Consider a seemingly simple question: if we double the amount of carbon dioxide in the atmosphere, how much will Earth warm? This is called the climate sensitivity, and it's one of the most important numbers in all of climate science. The answer determines whether we're facing a manageable challenge or an existential crisis.

Different models give different answers. Some models, when you double their CO2, warm by about 2 degrees Celsius. Others warm by 4 or even 5 degrees. This isn't because some modelers are incompetent. It's because the climate system contains deeply uncertain processes—especially clouds—where honest scientists can make different defensible choices about how to represent them.

By running all the models through the same experiments and comparing their outputs, CMIP reveals where models agree (robust findings) and where they disagree (areas of genuine scientific uncertainty). When twenty-five out of thirty models show the same behavior, that's a pretty confident prediction. When the models split evenly, that's a warning sign—a known unknown that affects our ability to plan for the future.

The Evolution Through Phases

CMIP has evolved through distinct phases, each more ambitious than the last.

CMIP1 and CMIP2, in the late 1990s, focused on basic experiments: simulating the pre-industrial climate (a "control run" with no human influence), then seeing what happens when you gradually increase CO2. The standard experiment increased carbon dioxide by one percent per year for eighty years, which doubles the concentration around year seventy. It was an idealized scenario, not meant to match the real world, but perfect for comparing model responses.

CMIP3, which collected data during 2005 and 2006, stepped closer to reality. Models simulated the actual twentieth century, including historical changes in greenhouse gases, volcanic eruptions, and solar variations. Then they projected forward using various future scenarios. This phase directly supported the Intergovernmental Panel on Climate Change's Fourth Assessment Report, providing the scientific foundation for international climate negotiations.

CMIP5 (2010-2014) introduced more realistic scenarios called Representative Concentration Pathways, or RCPs. Instead of just ramping up CO2, these pathways described consistent futures including greenhouse gas concentrations, land use changes, and air pollution. RCP2.6 represented an aggressive mitigation scenario where emissions peak soon and decline rapidly. RCP8.5 represented continued fossil fuel expansion—sometimes called the "business as usual" pathway, though that characterization is now debated.

CMIP5 also pioneered better documentation. A project called METAFOR created exhaustive descriptions of exactly what each model did—which equations it solved, which approximations it made, which datasets it used for testing. This metadata, stored alongside the output data, allows researchers to understand not just what the models predicted, but why they predicted it.

CMIP6 and the Hot Model Problem

The sixth phase, which began planning in 2013 and started collecting data around 2018, is the most ambitious yet. CMIP6 includes thirty-three modeling groups from sixteen countries, each running dozens of experiments.

The organizing framework became more sophisticated. At the core is something called DECK—the Diagnostic, Evaluation and Characterization of Klima (using the Greek word for climate). DECK includes a small set of experiments that every participating model must run: a pre-industrial control, an abrupt quadrupling of CO2, a gradual one-percent-per-year CO2 increase, and simulations of recent decades using observed sea surface temperatures. These standardized experiments allow direct comparison across all models.

Beyond DECK, CMIP6 endorsed twenty-three specialized Model Intercomparison Projects, or MIPs. Some focus on specific time periods—paleoclimate simulations of past ice ages, for instance. Others focus on specific processes—cloud feedbacks, land-atmosphere interactions, ocean biogeochemistry. Some focus on specific applications—extreme weather, sea level rise, carbon cycle responses.

CMIP6 also introduced Shared Socioeconomic Pathways, or SSPs, which replaced the earlier RCPs. The SSPs describe different future worlds not just in terms of greenhouse gas concentrations, but in terms of human development: population growth, economic development, inequality, governance capacity. A world with high inequality might have the same emissions as a world with low inequality, but the challenges of adapting to climate change would be vastly different.

But CMIP6 also revealed a troubling development. Several of the newest, most sophisticated models showed higher climate sensitivity than their predecessors. Some projected warming of 5°C or more from doubled CO2—at the upper edge of scientific estimates and potentially implying catastrophic consequences. These "hot models" triggered intense scientific debate. Are they capturing real processes that older models missed? Or are they overreacting to improvements in cloud simulation that inadvertently amplified positive feedbacks?

This question—whether the latest models are too hot—directly relates to the Substack article that prompted this essay. When someone asks whether the world is "warming faster than expected," part of the answer depends on which models you use as your expectation. If you include the CMIP6 hot models, recent temperatures look less surprising. If you exclude them (as many scientists now argue is appropriate given other evidence), recent temperatures look more concerning.

What the Data Actually Looks Like

The CMIP archive is staggeringly large. We're talking petabytes of data—temperature, precipitation, wind, pressure, ocean currents, ice coverage, and hundreds of other variables, at locations around the virtual Earth, for every simulated month across centuries of model time, from dozens of models running dozens of experiments each.

To make this usable, the project developed strict data standards. Every modeling center must output their data in exactly the same format, with exactly the same variable names, at exactly specified grid resolutions. Without this standardization, cross-model analysis would be nearly impossible.

The data is freely available to any researcher through a network of nodes around the world. You can download outputs from the German Climate Computing Centre, or Lawrence Livermore, or the British Met Office, depending on which is closest or has the fastest connection. This openness has enabled thousands of studies that would otherwise be impossible.

The Forcing Datasets

One of CMIP6's innovations deserves special attention: the forcing datasets.

When climate models simulate the past, they need to know what actually happened to the factors that drive climate change. How much CO2 was in the atmosphere each year? How much methane? What was the sun doing? Which volcanoes erupted, and how much sulfur did they inject into the stratosphere?

Previously, different modeling centers used different datasets for these "forcings," making it harder to compare models fairly. Maybe Model A predicted more warming than Model B simply because it assumed higher historical emissions, not because it had different climate physics.

CMIP6 provides common forcing datasets for all models to use. There are datasets for historical greenhouse gas emissions and concentrations. There are biomass burning emissions—the smoke and gases released when forests burn. There are land use datasets showing how humans have transformed the landscape over time. There are ozone concentrations (crucial for stratospheric chemistry) and aerosol optical properties (how particles scatter and absorb sunlight). There are solar forcing datasets tracking the eleven-year sunspot cycle and longer variations. There are stratospheric aerosol datasets for major volcanic eruptions.

All of these forcing datasets are stored and coordinated by a sister project called input4MIPs (input datasets for Model Intercomparison Projects). When you download a forcing dataset from input4MIPs, you know that every modeling center participating in CMIP6 used exactly the same data. This common foundation makes the model comparisons much more meaningful.

Looking Ahead to CMIP7

The seventh phase is already in development, with first data expected by the end of 2025. CMIP7 will introduce a more continuous release approach—rather than waiting until all models complete all experiments before making data available, results will flow out as they're ready.

There will also be "fast track" experiments designed to address urgent questions quickly. If a new scientific question emerges—say, about the climate effects of a specific geoengineering proposal—CMIP7 can organize a rapid response without waiting for the next full phase.

The community is also grappling with how to handle model genealogy. Many of today's models share ancestors—code and ideas borrowed or adapted from other modeling centers over decades of development. How independent are the models, really? If twenty models agree but they're all descended from the same three ancestors, maybe that agreement is less meaningful than it appears. CMIP7 aims to better track these relationships.

Why This All Matters

CMIP might be the most successful international scientific collaboration you've never heard of. It has enabled rigorous comparison of climate projections, provided the foundation for IPCC assessments that shape global policy, and created an open archive that democratizes climate science—any researcher with internet access can download the same data that informed the Paris Agreement.

The project embodies a particular philosophy about scientific uncertainty. Rather than seeking a single authoritative model, it embraces diversity. Rather than hiding disagreements, it documents them. Rather than claiming false precision, it honestly reports the range of possibilities.

When you read headlines about climate projections—that global warming might reach 1.5°C by a certain date, or that Arctic ice might disappear by a certain decade—those numbers almost certainly come from CMIP. The models don't agree precisely, but they converge on certain fundamental findings: the climate is warming, humans are responsible, and continued emissions will bring more warming.

The disagreements between models aren't failures. They're honest acknowledgments of what we don't yet know—and guideposts for where research should focus next.