An Earth System Model, or ESM, is a computer simulation that represents not just the physics of the atmosphere and ocean but also the chemical and biological processes that shape the planet’s climate. Where an older-style climate model might track how air and water move and exchange heat, an ESM goes further: it simulates the carbon cycle, the growth and death of vegetation, the chemistry of aerosols, the flow of ice sheets, and sometimes even human land-use decisions. The result is a virtual Earth whose components can influence each other in ways that mirror the real planet’s complexity, and that complexity is what makes ESMs the primary tool behind the climate projections you see in international scientific assessments.
From Weather Simulators to a Virtual Planet
The ancestor of the ESM is the general circulation model, or GCM, which divides the atmosphere and ocean into a three-dimensional grid and solves the equations of fluid motion at each grid cell. GCMs are physics-based simulators that pair a numerical solver for large-scale dynamics with simplified stand-ins for small-scale processes like cloud formation.1PubMed Central. Neural general circulation models for weather and climate They do a good job tracking how heat, moisture, and momentum move through the atmosphere and ocean. But a GCM, on its own, treats important things like vegetation cover or the concentration of carbon dioxide as fixed inputs rather than variables that change in response to climate. That limitation matters because, in reality, a warming climate can alter how much CO₂ forests absorb, which in turn changes how fast the climate warms. An ESM closes that loop by allowing the biology and chemistry to evolve alongside the physics.
What Is Actually Inside an ESM
An ESM is not a single program. It is a collection of separate component models, each representing a different part of the Earth system, wired together through coupling software that passes information between them at regular time steps. The main components typically include an atmospheric model, an ocean circulation model, a land surface model (often with a dynamic vegetation module), a sea-ice model, and increasingly, an ice-sheet model and an atmospheric chemistry module. Some ESMs now also include ocean biogeochemistry and even a human-systems component.
The coupler is the piece that makes the whole assembly work as one system. It handles the exchange of fluxes: heat, freshwater, momentum, and chemical species crossing boundaries between the atmosphere and the ocean, or between the land and the atmosphere. The Community Earth System Model (CESM), one of the most widely used ESMs, introduced a coupling architecture in which a top-level driver calls each component’s initialize, run, and finalize steps through standardized interfaces, giving the system enough flexibility to add or swap components without rebuilding the whole model.2The International Journal of High Performance Computing Applications. A new flexible coupler for earth system modeling developed for CCSM4 and CESM1
The Land Surface and Living Vegetation
Land surface modeling originated as a way to represent the atmosphere’s lower boundary over continents in climate models. Energy, water, and momentum cross that boundary through processes that often involve plants.3WIREs Climate Change. Modeling vegetation and land use in models of the Earth System Early land models treated vegetation as a static map: forests stayed forests, grasslands stayed grasslands. Dynamic global vegetation models, or DGVMs, changed that by simulating how plant communities grow, compete, burn, and die in response to climate. These models cover carbon and water cycling, fire, and the effects of vegetation on atmospheric trace gases.4Encyclopedia of Biodiversity. Dynamic Global Vegetation Models
Embedding a DGVM into an ESM is not trivial. The vegetation model has to hand the atmospheric model realistic surface boundary conditions, including fluxes of energy, water, and CO₂, at every time step. Recent work on the LPJ-GUESS model, for instance, required introducing a sub-daily time step, a new scheme for how light passes through the canopy, improved soil physics, and full energy-balance closure before the model could simulate the diurnal exchange of energy and gases that an atmospheric model expects to receive.5Geoscientific Model Development. LPJ-GUESS/LSMv1.0: a next-generation land surface model with high ecological realism The payoff is that the ESM can now capture feedback loops like forests drying out under warming, releasing carbon, and amplifying the warming further.
Ocean Biogeochemistry
The ocean absorbs a large fraction of both the heat and the CO₂ that human activity adds to the climate system. In an ESM, a physical ocean model handles currents and temperature, but a separate biogeochemistry module tracks how nutrients, carbon, and oxygen move through the water column and interact with marine life. GFDL’s Earth System Model 4.1, for example, realistically captures large-scale nutrient distributions, plankton dynamics, and the biological pump that exports carbon from the surface to the deep ocean, including cumulative carbon uptake since preindustrial times consistent with observational estimates.6Journal of Advances in Modeling Earth Systems. Ocean Biogeochemistry in GFDL’s Earth System Model 4.1 and Its Response to Increasing Atmospheric CO2
Newer marine ecosystem modules go further, representing multiple types of phytoplankton and zooplankton to capture how climate-driven shifts in plankton communities change carbon export and energy transfer through the food web.7Journal of Advances in Modeling Earth Systems. Simulating Marine Ecosystem Dynamics and Biogeochemical Cycling With Multiple Plankton Functional Types Getting these details right matters because the ocean’s capacity to keep absorbing CO₂ depends on biology, not just physics: if warming shifts plankton communities in ways that reduce carbon export, the ocean becomes a weaker carbon sink.
Ice Sheets and Sea-Level Rise
For a long time, ice sheets were the missing piece in ESMs. The Greenland and Antarctic ice sheets change slowly by atmospheric standards, but their melting can raise sea levels by meters over centuries. Coupling an ice-sheet model to an ESM requires passing fluxes of energy and water, as well as updating the surface topography of the ice and the geometry of floating ice shelves, at the boundary between the components. The U.K. Earth System Model (UKESM1) achieved two-way coupling with both the Greenland and Antarctic ice sheets through substantial technical development of these interfaces.8Journal of Advances in Modeling Earth Systems. Coupling the U.K. Earth System Model to Dynamic Models of the Greenland and Antarctic Ice Sheets
CESM2 took a similar path with its ice-sheet component, CISM2, incorporating an energy-balance-based surface mass balance calculation with downscaling via elevation classes, a closed freshwater budget from the ice sheet through to the ocean, and dynamic updates to both land surface types and atmospheric topography as the ice changes shape.9PubMed Central. Description and Demonstration of the Coupled Community Earth System Model v2 – Community Ice Sheet Model v2 (CESM2-CISM2) Without this coupling, projections of future sea-level rise would miss the feedback between a melting ice sheet, the freshwater it dumps into the ocean, and the resulting changes in ocean circulation and regional climate.
Atmospheric Chemistry and Aerosols
The composition of the atmosphere is not just a backdrop for climate. It is an active participant. Ozone, methane, and aerosol particles all affect how much energy the planet absorbs and reflects. Atmospheric chemistry controls the radiative balance of the climate and interacts with the physical climate, biogeochemical cycles, and human emissions in tightly coupled ways.10Journal of Advances in Modeling Earth Systems. Interactive Gas Chemistry for Enhanced Science Capabilities of the Energy Exascale Earth System Model Version 3
The feedbacks here can be surprising and nonlinear. Analysis of the latest generation of ESMs used in CMIP6 found that the overall climate feedback through chemistry and aerosols is negative: a warmer world produces more sea salt spray and more volatile organic compounds from vegetation, both of which create aerosols that reflect sunlight and cool the surface. Increased chemical loss of ozone and methane also contributes a cooling effect. But the picture is messy. Methane’s atmospheric lifetime actually increases in a warmer climate because of those same plant-derived organic compounds, and wetland methane emissions rise with warming, partly offsetting the negative feedbacks.11Atmospheric Chemistry and Physics. Climate-driven chemistry and aerosol feedbacks in CMIP6 Earth system models Work with the U.K. ESM has shown that the internal mixing and chemical interactions among aerosol species mean that forcings do not simply add up: the total effect of all aerosols is less than the sum of the individual parts, and nitrogen oxide emissions can even flip the sign of their net climate effect once you account for how they alter oxidants and in turn alter aerosol formation.12Atmospheric Chemistry and Physics. Assessment of pre-industrial to present-day anthropogenic climate forcing in UKESM1
The Parameterization Problem
ESMs divide the planet into grid cells that are typically tens to hundreds of kilometers across. Anything smaller than one grid cell, like an individual thunderstorm or a turbulent eddy in the ocean, cannot be simulated directly. Instead, the model uses simplified formulas called parameterizations to estimate the collective effect of those small-scale processes. This is the single largest source of structural uncertainty in climate modeling: two ESMs using different parameterizations for clouds can give noticeably different projections of future warming, even when fed the same emissions scenario.
Machine learning is starting to change this. Researchers have trained deep neural networks on high-resolution simulations where convection is resolved explicitly, then dropped those trained networks into a global model to replace traditional parameterizations.13PubMed Central. Deep learning to represent subgrid processes in climate models A related technique, the multiscale modeling framework, embeds a kilometer-resolution cloud-resolving model within each column of the larger climate model. Machine learning can emulate that embedded model at a fraction of the computational cost.14Journal of Advances in Modeling Earth Systems. Stable Machine‐Learning Parameterization of Subgrid Processes in a Comprehensive Atmospheric Model Learned From Embedded Convection‐Permitting Simulations The goal is not to replace physics with a black box but to let the model learn more realistic small-scale behavior from data while still respecting the large-scale dynamics the physics equations handle well.
How Much Computing Power ESMs Demand
Running an ESM is expensive. The sheer number of grid cells, time steps, and interacting processes means that even with dedicated supercomputers, a high-resolution ESM might produce only about one simulated year per day of wall-clock time. Optimization work on a high-resolution version of CESM, for instance, improved throughput from about 1 simulated year per day to roughly 3.4 simulated years per day by refactoring key subroutines and exploiting the architecture of a many-core supercomputer, with individual computation kernels sped up by as much as 27 times.15Geoscientific Model Development. Optimizing high-resolution Community Earth System Model on a heterogeneous many-core supercomputing platform Those gains sound abstract until you consider that a climate projection spanning several centuries of simulated time could take half a year of continuous computation even at the faster speed. Multiply that by dozens of runs needed to explore different scenarios and quantify uncertainty, and the total cost in electricity and machine time becomes staggering.
Testing ESMs Against the Past
A model that cannot reproduce what already happened has no business predicting the future. Paleoclimate experiments are one of the most powerful tools for evaluating ESMs. By feeding a model the known conditions of a past period, like the concentrations of greenhouse gases during the last ice age, and checking whether it produces a climate that matches geological and ice-core records, researchers can test whether the model’s feedbacks and sensitivities are in the right ballpark.16Geoscientific Model Development. Set-up of the PMIP3 paleoclimate experiments conducted using an Earth system model, MIROC-ESM
A comparison of intermediate-complexity ESMs found that despite very different modeled preindustrial temperatures, the models reasonably reproduced observed 20th-century trends in surface temperature and carbon uptake. They did tend to underestimate the cooling between the Medieval Climate Anomaly and the Little Ice Age when compared to paleoclimate reconstructions, which hints at missing feedbacks or underestimated sensitivity to natural forcing in some models.17Climate of the Past. Historical and idealized climate model experiments: an intercomparison of Earth system models of intermediate complexity Modern evaluation efforts also focus on the biosphere, examining whether ESMs correctly simulate the carbon cycle and associated biological processes against both present-day observations and paleodata, with significant uncertainties remaining.18Biogeosciences. Evaluation of biospheric components in Earth system models using modern and palaeo-observations: the state-of-the-art
CMIP and Why Standardization Matters
No single ESM is authoritative. Different modeling centers around the world build their own ESMs with different component models, different parameterizations, and different resolutions. To make their outputs comparable and scientifically useful, the community organizes itself through the Coupled Model Intercomparison Project, or CMIP. The current phase, CMIP6, established a set of common experiments, shared standards for data output, and a federated structure of endorsed sub-projects that address specific scientific questions. The core includes a set of diagnostic experiments (called the DECK) and historical simulations from 1850 to the near present, which maintain continuity across CMIP phases and document basic model characteristics.19Geoscientific Model Development. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization When you see a headline about how much warming to expect by 2100, the underlying projections almost certainly come from a CMIP multi-model ensemble.
Where the Uncertainty Comes From
ESM projections carry three distinct types of uncertainty, and their relative importance shifts over time. In CMIP6 temperature projections over global land, model uncertainty (the spread between different ESMs) dominates in the near term, accounting for about 99% of the total spread early on, but drops to about 39% by late century. Scenario uncertainty, which reflects not knowing future emissions, starts negligible and rises to about 61% by late century.20Earth’s Future. Quantifying the Uncertainty Sources of Future Climate Projections and Narrowing Uncertainties With Bias Correction Techniques For precipitation, the picture is different: model uncertainty stays dominant at roughly 98% of total spread, with scenario uncertainty and natural internal variability contributing only small fractions.21Geophysical Research Letters. The Sources of Uncertainty in the Projection of Global Land Monsoon Precipitation The practical takeaway is that for temperature, reducing emissions ambiguity (through policy decisions) narrows long-term uncertainty more than improving models would. For rainfall, the models themselves need to get better.
Adding Humans to the Earth System
The “system” in Earth System Model has traditionally meant the natural system. Humans enter only as an external script: you tell the model how much CO₂ to emit and how much forest to clear, and the model runs from there. A newer approach tries to close that loop too. The Energy Exascale Earth System Model (E3SM) now includes a human component built around the Global Change Analysis Model, synchronously coupled with the land and atmosphere. Terrestrial productivity computed by the ESM is passed to the human model, which generates climate-responsive land-use changes and CO₂ emissions for the next five-year period, which are then fed back to the ESM annually. The coupling affects crop prices, terrestrial carbon storage, local surface temperature, and the frequency of climate extremes.22Journal of Advances in Modeling Earth Systems. E3SM‐GCAM: A Synchronously Coupled Human Component in the E3SM Earth System Model Enables Novel Human‐Earth Feedback Research This kind of two-way human-Earth coupling is still in its early stages, but it represents a conceptual shift: the model no longer treats human decisions as something outside the system.
Tipping Points and What ESMs Struggle With
Tipping points, such as the potential collapse of the Atlantic overturning circulation or the dieback of the Amazon rainforest, are among the most consequential risks in climate science and among the hardest for ESMs to capture. A recent review found that most proposed tipping elements may not possess the potential for abrupt change within years, and some may not exhibit classic tipping behavior at all, instead responding more predictably to the magnitude of forcing. But uncertainties remain large.23Reviews of Geophysics. Mechanisms and Impacts of Earth System Tipping Elements A new international effort, the Tipping Points Modelling Intercomparison Project (TIPMIP), aims to systematically assess these risks using state-of-the-art coupled ESMs alongside stand-alone ice-sheet and land-system models.24Earth System Dynamics Discussions. The Tipping Points Modelling Intercomparison Project (TIPMIP): Assessing tipping point risks in the Earth system The challenge is that tipping dynamics often depend on processes at scales or timescales that current ESMs resolve poorly, making this an area where model development and scientific understanding need to advance together.
Storm-Resolving Models and AI Integration
One of the most exciting frontiers is the move toward global storm-resolving models, which push atmospheric grid spacing down to around 5 kilometers or finer. At that resolution, deep convection, the towering storms that drive much of the tropics’ weather and energy transport, can be simulated explicitly rather than parameterized. Fully coupled global storm-resolving models run at roughly 5-kilometer atmospheric resolution produce far more realistic tropical cyclone activity, peak intensity, and rapid intensification than coarser models. Rapid intensification, which is absent at typical climate-model resolutions, is captured at storm-resolving scales.25Geophysical Research Letters. On the Realism of Tropical Cyclone Intensification in Global Storm‐Resolving Climate Models These models also offer enhanced detail of boundary-layer cloud processes.26Journal of Geophysical Research: Atmospheres. Boundary‐Layer‐Coupled and Decoupled Clouds in Global Storm‐Resolving Models: Comparisons With the ARM Observations The trade-off is cost: running a global model at 5 km for century-length simulations remains beyond current budgets for routine use.
Artificial intelligence is being woven into ESMs in other ways too. Researchers have replaced individual parameterization schemes with trained neural networks, such as a transformer-based AI model that handles ocean vertical mixing in CESM, achieving roughly a three-to-five-times improvement in computational efficiency for that component while maintaining stable operation for multi-year runs.27Ocean Modelling. Developing Intelligent Earth System Models: An AI scheme of K-profile parameterization and stable coupling into CESM with FTA New software frameworks now provide a Fortran-Python bridge so that machine-learned parameterizations for convection, radiation, and even wildfire can be plugged into existing ESM code and interact bidirectionally with the rest of the model.28Geoscientific Model Development. A Fortran–Python interface for integrating machine learning parameterization into earth system models
From Global Projections to Local Decisions
ESMs run on grids that are too coarse to tell you what will happen in a specific valley or coastal city. Bridging that gap requires downscaling: feeding ESM output into a regional climate model or statistical method to generate high-resolution local projections. Dynamical downscaling, which nests a finer-resolution regional model inside the ESM, is the gold standard for localized future climate information but is computationally expensive enough that it can only be applied to a handful of ESM runs at a time. A newer hybrid approach combines a regional model that downscales to an intermediate resolution with a generative AI diffusion model that refines the output further, reducing cost while improving uncertainty estimates across large climate-projection ensembles.29PubMed Central. Dynamical-generative downscaling of climate model ensembles These downscaled products are what farmers, water managers, and city planners actually use when they make adaptation decisions, making ESMs the upstream engine behind a remarkably wide range of practical choices about infrastructure, agriculture, and disaster preparedness.