Parameterization in Climate Models

Climate models are humanity's most powerful tools for understanding our planet, built upon the fundamental laws of physics that govern the motion of air and water. In an ideal world, these models would simulate every gust of wind and every cloud droplet. However, computational limitations force scientists to simplify this intricate reality, laying a coarse grid over the globe and solving for the average conditions within each grid box. This act of averaging creates a profound challenge: what happens to the crucial physical processes that are smaller than a single grid box? This is the knowledge gap that the science of parameterization seeks to fill.

This article delves into the art and science of parameterization, the method by which climate models account for the statistical effects of these unresolved, subgrid-scale processes. First, the "Principles and Mechanisms" chapter will explain why parameterization is necessary, how it works by "closing" the governing equations, and the different philosophical approaches to its construction. Following this, the "Applications and Interdisciplinary Connections" chapter will explore how these schemes are the practical heart of weather and climate prediction, governing everything from cloud formation and radiation to the grand exchanges of energy between the ocean, land, and atmosphere.

Imagine you want to build a model of our planet's climate. You begin, as any good physicist would, with the fundamental laws of nature: the conservation of mass, momentum, and energy. These are the beautiful, immutable rules, expressed as partial differential equations, that govern the motion of every parcel of air and every drop of water. In a perfect world, we would solve these equations for the entire atmosphere and ocean, tracking every gust of wind and every tiny cloud droplet. But here we hit a wall—a wall of scale.

Our planet is vast and intricate. The climate system is a turbulent symphony of motions, from continent-spanning jet streams down to the microscopic dance of water molecules condensing on a speck of dust. A computer, no matter how powerful, cannot possibly track it all. We are forced to simplify. We lay a grid over the globe, with each grid box perhaps 100 kilometers on a side, and we try to solve our equations for the average state of the atmosphere within each box.

But what does this averaging do to our beautiful equations? Consider a thunderstorm, a magnificent engine of weather, maybe 10 kilometers across. To a model with a 100-kilometer grid, this entire storm lives and dies entirely inside a single grid box. The model, solving for the average wind and temperature in that box, is completely blind to the furious updrafts and swirling motions of the storm within.

This blindness is not just an inconvenience; it creates a fundamental mathematical hole in our model. The laws of physics are nonlinear. For instance, the way heat is moved around depends on the product of velocity and temperature ($\mathbf{u}T$). When we average this product over a grid box, we get something tricky. The average of the product is not the same as the product of the averages. An extra term appears, a term that represents the transport of heat by the small-scale, unresolved motions—the correlation between the fluctuations in wind and temperature, written schematically as $\overline{\mathbf{u}' T'}$.

Suddenly, our equations for the resolved, large-scale flow (the averages) are contaminated by terms that depend on the unresolved, small-scale flow (the fluctuations). Our system of equations is no longer self-contained. It is "unclosed." This is the famous ​​closure problem​​. To make our model predictive, we must find a way to represent the statistical effects of all the unresolved processes—the thunderstorms, the wispy clouds, the turbulent eddies—in terms of the large-scale, resolved variables that our model actually knows about. This act of "closing" the equations is the science and art of ​​parameterization​​. A parameterization is a physically and statistically informed recipe that approximates the effects of the unresolved world on the resolved one.

To understand how this works in practice, it helps to think of a climate model as having two distinct souls: the ​​dynamical core​​ and the ​​physical parameterizations​​.

The ​​dynamical core​​ is the powerful, relentless engine of the model. Its job is to solve the discretized equations of fluid motion on a rotating sphere. It is the part of the code that handles the grand, resolved-scale waltz of the atmosphere: the Coriolis force, the pressure gradients, and the advection that moves air masses across continents. It is pure, resolved fluid dynamics.

But an engine needs forces to act upon. The ​​physical parameterizations​​ are the collection of modules that provide these forces. They are the sub-models that calculate the tendencies—the heating, cooling, moistening, and dragging—that arise from the unresolved subgrid world. For example, a radiation parameterization calculates how much the sun warms the ground and how much heat escapes to space. A turbulence parameterization calculates the frictional drag at the Earth's surface. And a convection parameterization calculates the powerful vertical transport of heat and moisture by thunderstorms. The dynamical core takes these tendencies and integrates them forward in time, evolving the state of the climate.

How does one build a recipe for something one cannot see? This is the central challenge of parameterization, and it has given rise to different schools of thought, broadly falling into two camps.

The first is the ​​physically-based​​ approach. Here, we try to build a simplified, miniature theory of the subgrid process. Consider the formation of rain. We can't model every single cloud droplet, but we can write down prognostic equations for the bulk properties of the grid box, like the total mass of cloud water ($q_c$) and the total mass of rain water ($q_r$). We then use simplified physical laws—like how quickly cloud droplets collide to form raindrops (a process called autoconversion)—to create a set of equations that relate these bulk quantities. These are known as bulk microphysics schemes, and they are built on a foundation of mechanistic laws, even if they are heavily simplified.

The second approach is ​​statistical​​. Instead of modeling a simplified physical entity, it tries to represent the subgrid world as a statistical ensemble. For example, instead of modeling one idealized convective plume, one might assume a probability distribution for the vertical velocity within a grid box. The total effect of convection is then found by integrating over this distribution. This approach acknowledges our ignorance of the precise subgrid state and instead aims to capture its statistical character. Modern machine learning techniques, which learn statistical relationships from vast amounts of data (often from high-resolution simulations), represent a powerful new frontier for this statistical approach.

Let's make this more concrete by dissecting a parameterization for deep convection—thunderstorms. This process is so important for the Earth's energy budget that it has been a focus of intense research for decades.

First, the model needs to decide when convection should occur. Convection is a conditional process; it requires an unstable atmosphere and a lifting mechanism. Parameterizations encode this logic in a ​​trigger function​​. This is essentially an IF statement in the code: IF the Convective Available Potential Energy (CAPE) is large enough, and IF the Convective Inhibition (CIN) is small enough, THEN call the convection scheme. The trigger function is a dimensionless gate, a mapping from the resolved state of the model to a simple {0, 1} or [0, 1] decision: "on" or "off".

Once triggered, what does the scheme do? Early, simple schemes used a ​​convective adjustment​​ principle. They observed that the end result of convection is to stabilize the atmosphere. So, the scheme's logic was simple: if the atmosphere is unstable, just mix the temperature and moisture vertically until it reaches a neutral, moist-adiabatic profile. This is done over a specified relaxation timescale, $\tau$. It's a phenomenological, brute-force approach that says, "I don't know the details of the journey, but I know the destination is stability."

A more physically sophisticated approach is the ​​mass-flux scheme​​. Instead of just adjusting the whole grid box, it attempts to model the key actors: an ensemble of rising convective updrafts (plumes) and sinking downdrafts. It solves a separate budget equation for the properties of the air inside the idealized plume as it rises, entraining air from the environment and detraining its own air at the top. The "closure" for this type of scheme is to determine the total strength of the convection, typically by relating the cloud-base mass flux, $M_b$, to some property of the large-scale flow, such as the rate at which CAPE is being generated. This is a far more mechanistic view, representing the subgrid world as a collection of physical objects with their own dynamics.

The science of parameterization is continually evolving, pushing into fascinating new territory that reveals the deep, interconnected nature of physics.

​​Memory:​​ Does the effect of a subgrid process have to be instantaneous? Think about a forest fire. The smoke it produces today will affect the sunlight and temperature tomorrow. Subgrid processes can have memory. In some advanced parameterizations, the tendency at time $t$ is not just a function of the state at time $t$, but is represented as a convolution integral over the entire history of the resolved state: $T(t) = \int_0^\infty K(\tau) X(t-\tau) d\tau$. The function $K(\tau)$ is a ​​memory kernel​​, describing how much the state of the system $\tau$ seconds ago influences the tendency now. For a process with a fast response, $K(\tau)$ will be sharply peaked near $\tau=0$; for a process with long memory, it will have a long tail. This is a profound idea borrowed from statistical mechanics, acknowledging that the present is shaped by the integrated influence of the past.

​​Scale-Awareness:​​ Traditional parameterizations were designed for a single, coarse grid resolution. But what happens as our computers get faster and we can afford to run models with a 50 km grid, then a 10 km grid, then a 2 km grid? At some point, our model's dynamical core will begin to explicitly resolve the very processes our parameterizations were designed to represent. A naive parameterization would continue to add its own convective heating, even as the model's own resolved updrafts are generating heating, leading to a "double counting" of the effect. A modern, ​​scale-aware parameterization​​ is designed to recognize this. It monitors the state of the resolved flow (for instance, the variance of the vertical velocity) and gracefully "tapers" its own contribution as the process becomes more resolved. This ensures a smooth transition from a fully parameterized regime to a fully resolved one, a critical challenge for the next generation of climate models.

​​Inherent Uncertainty:​​ No matter how sophisticated, a parameterization is an approximation of a complex reality. Can we quantify the error this introduces? The answer, beautifully, is yes. Imagine we parameterize a cloud property as a simple polynomial approximation of the true, complex physical function. This is equivalent to writing down a Taylor series for the true function. The error we make by truncating this series at, say, the cubic term is nothing more than the ​​remainder term​​ from Taylor's theorem. Mathematics gives us an exact expression for this error, $R_n(x) = f^{(n+1)}(\xi) (x-x_0)^{n+1} / (n+1)!$, which depends on a higher-order derivative of the true function. If we can put a physical bound on that derivative, we can put a rigorous bound on the "structural uncertainty" of our parameterization. This provides a powerful, fundamental link between pure mathematics and the practical limits of our knowledge.

Why do we obsess over these seemingly esoteric details of subgrid models? Because the climate system is a tightly coupled, nonlinear machine. Small, persistent errors in the representation of the small scales can cascade upwards, leading to large, systematic biases in our simulations of the entire planet.

Many of the most stubborn biases in modern climate models can be traced directly back to flaws in parameterization, particularly for convection.

• The ​​"double ITCZ"​​ problem, where models produce two bands of tropical rain straddling the equator instead of one, has been linked to convection schemes that are insensitive to the humidity of the surrounding air (a consequence of low entrainment rates).
• The tendency for models to produce rain over land in the early afternoon, hours before it is observed in reality, is often caused by trigger functions and closure assumptions that initiate convection too readily, as soon as surface heating creates a little bit of instability.
• The "too frequent, too light" rain problem—the perpetual, boring drizzle seen in many models—is a direct consequence of schemes that lack the mechanisms (like prognostic memory or mesoscale organization) to store energy and release it in less frequent, more intense bursts.

Parameterization, therefore, is not just a technical fix for a computational limitation. It is the science of representing the physics of our multiscale world. It is a frontier where fluid dynamics, thermodynamics, statistical mechanics, and computer science meet. Getting it right is one of the grandest intellectual challenges we face, and it is absolutely essential for building the tools we need to understand and predict the future of our climate.

Having peered into the inner workings of parameterization, we might be tempted to think of it as a collection of necessary but perhaps unglamorous mathematical tricks—the scaffolding that holds up the grand cathedral of a climate model. But to do so would be to miss the forest for the trees. Parameterizations are not just computational conveniences; they are the very language through which a model speaks about the complex, multi-scale processes that govern our world. They are the bridges connecting disciplines, the tools for exploring Earth’s deep past, and the lenses through which we view its potential futures. It is here, in their application, that the true power and beauty of this scientific art form are revealed.

Let us start with something we can all see: a cloud. To a climate model, a grid box might be a hundred kilometers on a side, a vast territory. Inside this box, a real thunderstorm might live and die, a majestic, turbulent tower of air and water—a process far too small and swift for the model to "see" directly. So, how does the model know a storm is brewing? It relies on a parameterization, a sort of statistical weather forecaster for the sub-grid world. These schemes look at the large-scale conditions—the temperature, the humidity, the upward motion of air—and make an educated guess about the collective effect of the unseen storms. They answer questions like: How much heat and moisture will they pump into the upper atmosphere?

This works beautifully, up to a point. The core assumption is one of ​​scale separation​​: that the small, fast world of convection is distinct from the large, slow world of the model's grid. But nature loves to break rules. In the tropics, thunderstorms often organize themselves into colossal weather systems, like the Madden-Julian Oscillation (MJO), that can span thousands of kilometers and influence weather globally for weeks. Here, the scales are no longer separate; the "sub-grid" process has become a planetary-scale phenomenon. The failure of classic parameterizations in this regime is one of the grand challenges in climate modeling, pushing scientists to develop new ideas like "superparameterization," where a tiny, high-resolution cloud model is run inside each and every grid box of the global model—a brute-force, but stunningly effective, approach to honoring the complex reality of organized convection.

Even once a model decides a cloud exists, the story is just beginning. What is it made of? How does it produce rain? Imagine a tiny cloud droplet, a microscopic sphere of water. As it falls, the air feels thick and syrupy, like honey. For such a small object, the drag force is governed by the simple, elegant physics of Stokes' law, where drag is directly proportional to speed. The droplet quickly reaches a terminal velocity, a gentle drift downwards. But this is not the world of a raindrop. A millimeter-sized raindrop is a behemoth by comparison. It falls so fast that the air is no longer a syrupy fluid but a rushing wind. The drag becomes turbulent and chaotic, proportional to the square of its velocity. The raindrop itself deforms, flattened by air pressure. The simple physics breaks down completely. No clean, simple formula can describe its fall. So, what does a model do? It uses an ​​empirical parameterization​​—a relationship, like $V_t = a D^{\nu}$, derived not from first principles but from painstakingly collected laboratory data on falling drops. It's a pragmatic, effective admission that sometimes the best way to describe a complex reality is to measure it and fit a curve.

This complexity deepens when we consider that clouds are not always made of liquid water. High in the cold atmosphere, they are often composed of ice crystals. The difference is profound. Ice clouds are typically thinner, have different optical properties, and are less likely to rain, so they live longer. A model's decision to form a liquid or an ice cloud can dramatically alter its prediction of Earth's energy balance. This decision hinges on the process of freezing, which comes in two main flavors: ​​homogeneous​​, the spontaneous freezing of pure water at frigid temperatures (below about $-38^{\circ}\mathrm{C}$), and ​​heterogeneous​​, where freezing is catalyzed by a tiny speck of dust or aerosol, an Ice Nucleating Particle (INP), at much warmer temperatures. Parameterizations must therefore act as microphysical referees, tracking the population of INPs and deciding which pathway—immersion freezing from within a droplet, contact freezing from a collision, or direct deposition of vapor onto a particle—is most likely under the given conditions. This intricate dance of aerosol, water, and ice is a major source of uncertainty in climate projections, and it all plays out within the logic of a parameterization scheme.

Every parameterization we've discussed—convection, rain, ice formation—ultimately feeds into the most fundamental question for climate: the energy budget. Clouds are the planet's great regulators of temperature, reflecting sunlight back to space and trapping heat radiating from the surface. But how do models calculate this? The interaction of radiation with matter is fiendishly complex. A greenhouse gas like carbon dioxide or water vapor doesn't just absorb light at one frequency; it absorbs at thousands of individual, sharp spectral lines. To calculate the radiative transfer line-by-line across the whole spectrum for the entire atmosphere would bring the world's largest supercomputers to their knees.

Instead, models use a brilliant parameterization known as the ​​correlated-k method​​. The core idea is to recognize that from a radiative point of view, it doesn't matter what the exact frequency of light is, only how strongly it is absorbed. The method cleverly re-sorts the entire spectrum not by frequency, but by absorption strength. By calculating the radiative transfer for just a few representative "bins" of absorption strength, it can reproduce the total effect with astonishing accuracy and a tiny fraction of the computational cost. It is a beautiful example of how a deep physical insight can be translated into an elegant and powerful mathematical shortcut, enabling climate models to simulate the greenhouse effect efficiently.

Our planet is an interconnected system. The atmosphere, ocean, land, and biosphere are constantly exchanging energy and matter. Parameterizations form the bridges that allow these components to "talk" to each other in a model.

Consider the wind blowing over the ocean. This is not just a weather phenomenon; it is the primary way the atmosphere transfers momentum to the ocean, driving currents that move heat around the globe. The amount of momentum transferred depends on the "roughness" of the sea surface. But what determines this roughness? A simple parameterization might just assume a constant value. A more sophisticated one recognizes that the sea surface is a dynamic interface. A young, choppy sea, whipped up by a strong local wind, is much "rougher" to the wind than a sea with long, rolling swells that have traveled from a distant storm. Modern models therefore use sea-state-dependent parameterizations, often using a wave model to tell the atmospheric model how rough the sea is. This coupling is vital for accurately predicting the intensity of hurricanes, which feed on the transfer of energy from the warm ocean surface—a transfer governed by these very roughness parameterizations.

The ocean and atmosphere also exchange matter, most critically for us, carbon dioxide. The rate at which the ocean can absorb the $\text{CO}_2$ we emit is a crucial regulator of our planet's climate. This process is governed by the thin, turbulent boundary layer at the top of the ocean. The gas transfer velocity, a key parameter in this exchange, is known to depend strongly on wind speed. Decades of research and observation have led to parameterizations, often quadratic in wind speed ($k \propto U^2$), that allow Earth System Models to calculate this vital carbon flux. It is through such a parameterization that a model can connect the physics of wind and waves to the global carbon cycle, helping us understand one of the most pressing questions of our time: where does all the carbon go?.

With these tools in hand, we can do more than just forecast the weather. We can become time travelers. How was the climate different during the Last Glacial Maximum (LGM), 20,000 years ago, when vast ice sheets covered North America? Paleoclimate records tell us the atmosphere was colder and, in many places, drier. Using an entraining plume parameterization for convection, we can ask the model: what does a drier atmosphere do to thunderstorms? The parameterization provides a clear physical answer. Entraining dry air into a rising plume of moist air saps its buoyancy. In the drier world of the LGM, a convective parcel needed a much stronger initial "kick" of energy at its base to survive the ascent and grow into a deep storm. Parameterizations thus provide a physical mechanism to connect the large-scale environmental changes of the past to the behavior of weather systems.

These same tools allow us to explore potential futures, including those involving geoengineering. One proposed idea, Marine Cloud Brightening (MCB), suggests spraying sea salt aerosols into marine air to make clouds brighter and more reflective, thereby cooling the planet. How could we even begin to evaluate such a radical idea? The answer lies in the aerosol-cloud interaction parameterizations we've already discussed. These schemes link the number of aerosol particles (CCN) to the number of cloud droplets ($N_d$). Adding more CCN creates more, smaller droplets. This has two effects: the ​​Twomey effect​​, where a cloud with more, smaller droplets becomes optically thicker and reflects more sunlight for the same amount of water; and the ​​Albrecht effect​​, where the smaller droplets are less efficient at forming rain, making the cloud live longer and hold more water. Both effects, born from parameterizations, are at the heart of the MCB hypothesis, making these schemes indispensable tools for assessing the potential efficacy and unintended consequences of such technologies.

If parameterizations are so central, it is also true that they are one of the largest sources of uncertainty in climate projections. Different modeling groups make different choices, leading to different results. This is not a failure of science, but a reflection of it. Scientists tackle this head-on using ​​ensemble modeling​​. They create ​​structural ensembles​​, collections of different models, to map out the uncertainty in our physical understanding ($s$). They run ​​perturbed-physics ensembles​​, where they vary the values of parameters within a single model, to explore the uncertainty in our measurements ($\theta$). And they run ​​initial-condition ensembles​​, where they start the same model with minutely different conditions, to quantify the irreducible uncertainty from chaos itself. This framework, used in projects like the Coupled Model Intercomparison Project (CMIP), allows scientists to attach confidence levels to their projections and understand what aspects are robust and what are still uncertain.

This brings us to the bleeding edge of the field. If deriving parameterizations from simplified theory is so hard, could we teach a machine to do it for us? This is the revolutionary idea behind using machine learning for parameterization discovery. The goal is profound. It's not about teaching a neural network to predict the "weather" — the exact trajectory of the system — which is impossible due to chaos. Instead, the goal is to teach it to reproduce the system's long-term statistical behavior, its "climate." In the language of dynamical systems, the aim is to create a reduced model whose ​​invariant measure​​—the mathematical object describing the probability of finding the system in any given state over a long time—matches that of the true, high-resolution system. By framing the problem this way, using loss functions like the Kullback-Leibler divergence to compare probability distributions, scientists are training AI to discover the statistical "ghosts" of the unresolved physics. This is more than just curve-fitting; it is a deep and beautiful quest to learn the fundamental statistical laws that govern our climate system, opening a new chapter in our quest to understand and predict the fate of our planet.