SciencePediaModeling the complex systems of our planet, such as the atmosphere and oceans, presents a fundamental computational challenge. The governing equations of fluid dynamics contain phenomena that evolve on vastly different schedules, from slow weather patterns that unfold over days to fast-moving sound and gravity waves that traverse a model grid in seconds. Standard explicit numerical methods are constrained by the fastest process, forcing the use of tiny time steps that make long-term simulations computationally prohibitive. This creates a significant knowledge gap, limiting our ability to efficiently model climate and weather over meaningful periods. This article addresses this problem by dissecting the split-explicit scheme, an elegant and powerful technique that circumvents this limitation. In the following chapters, we will explore the core principles and mechanisms that allow this scheme to work, and then survey its diverse applications and interdisciplinary connections, revealing how a single numerical idea unlocks a deeper understanding across multiple scientific fields.
To build a great skyscraper, you must lay a solid foundation. In the world of numerical modeling, that foundation is built upon understanding the core principles that govern the system you wish to simulate. The split-explicit scheme is a beautiful piece of computational architecture, designed to solve a very particular, and very frustrating, problem that lies at the heart of simulating our planet's atmosphere and oceans. Let's dig in and see how it works, starting from the ground up.
Imagine you are a project manager overseeing the construction of a new building. Your responsibilities include two very different tasks: supervising the welders who are assembling the steel frame, and checking the progress of the concrete foundation as it cures. The welders need constant supervision, perhaps a check-in every hour, to ensure every joint is perfect. The concrete, on the other hand, cures slowly and only needs to be inspected once a day. What would you do? It would be absurdly inefficient to check on the concrete every hour just because the welders need it. A sensible manager would handle the fast task (welding) on its rapid timescale, while dealing with the slow task (curing) on its own, much longer timescale.
This is precisely the dilemma faced by atmospheric and oceanic models. The governing equations of fluid dynamics, which describe the motion of air and water, are home to a zoo of phenomena that operate on vastly different schedules.
On one hand, we have the slow processes. These are the things we typically associate with "weather" or "ocean currents." They include advection, which is the bulk transport of air masses or water parcels by the prevailing winds or currents, and the effects of Earth's rotation, known as the Coriolis force. These processes evolve over timescales of hours, days, or even longer.
On the other hand, lurking within the same equations are the fast processes. These are waves that zip through the fluid at incredible speeds. In the atmosphere, the most famous of these are acoustic waves, or sound waves, which propagate at the speed of sound, roughly meters per second. In the ocean, the fastest signals are surface gravity waves that involve the entire water column moving in unison. This is called the external (or barotropic) mode, and for a typical ocean depth of meters, these waves travel at a blistering speed of m/s. Slower internal (or baroclinic) modes, which involve motion within the stratified layers of the ocean, travel at a comparatively sluggish pace of a few meters per second.
So why is this a problem? The trouble comes from a fundamental rule of explicit numerical simulations known as the Courant-Friedrichs-Lewy (CFL) condition. An explicit model advances time in discrete steps, or "snapshots," of size . The CFL condition is a rule of the road that says, to maintain numerical stability, your time step must be small enough that information doesn't jump over an entire grid cell of size in a single step. Mathematically, this is expressed as:
where is the speed of the fastest-moving wave in your system, and is a constant that depends on the specific numerical scheme (often close to 1). The dilemma is now clear: the incredibly fast acoustic or external gravity waves force us to use a tiny —perhaps just a few seconds. Yet, the large-scale weather patterns we actually want to predict evolve over hours. To run a massive, complex global model with a time step of a few seconds would be computationally ruinous, like checking on your slowly curing concrete every single minute.
Nature has handed us a difficult problem. The ingenuity of the split-explicit scheme is in how it sidesteps this problem. The core idea is simple and elegant: if the equations contain both fast and slow parts, let's split them apart and give each the attention it deserves.
We can formally write the governing equations for our system's state, , as a sum of two parts: a slow tendency, , and a fast tendency, .
The slow part, , includes terms like advection and the Coriolis force. The fast part, , contains the terms that generate sound and gravity waves, like the pressure gradient force and mass divergence.
The split-explicit strategy then proceeds just like our project manager:
The Outer Loop: We take one large time step, let's call it , for the slow dynamics. The size of this "slow" step is governed by the CFL condition for the slow-moving winds, , so we can choose based on the advective limit, . This step might be on the order of a minute.
The Inner Loop: Within that single large step, we perform a series of smaller "substeps," each of size , to accurately resolve the fast dynamics. The size of this "fast" step, , must satisfy the CFL condition for the fastest waves, . This step will be on the order of seconds.
The computational savings are immense. Let's take the realistic parameters from one of our case studies. For a model with a grid spacing of m, the advective speed m/s allows a slow time step of s. The sound speed m/s, however, demands a fast time step of s. This means we need , which rounds up to substeps. For every single time we compute the expensive slow physics, we perform 7 cheap updates of the fast physics. This is far better than computing everything 7 times. We've cleverly tailored our effort to the natural rhythm of the physics.
If it sounds too good to be true, you're right to be skeptical. Just running the fast and slow parts one after the other in a naive way is a recipe for disaster; it can lead to numerical instability and completely unphysical results. The true elegance of the split-explicit method lies in how the two loops "talk" to each other—a process called coupling.
During the inner loop, while the fast waves are zipping back and forth, the slow tendencies are essentially held constant, "frozen" at their values from the beginning of the outer step. The crucial question is: how does the slow step, at the end of its big leap forward in time, account for what the fast waves were doing?
The answer is that the slow dynamics should not react to the final, instantaneous state of the fast waves, but rather to their time-averaged effect over the entire outer step . Imagine the fast waves as a high-frequency vibration. You don't want to react to every single peak and trough; you want to respond to the net "push" they exerted over the whole interval.
This is achieved by accumulating the impulses from the fast steps. For example, the total change in the fluid's momentum isn't due to the pressure gradient at the very end of the time step. Instead, it's the sum of all the tiny pressure-gradient "pushes" from each of the inner substeps. We calculate the total fast impulse, which is an integral in time, by summing up the contributions from each small step:
This accumulated impulse is then used to update the momentum in the outer loop. This careful accounting prevents the slow evolution from being erratically "shocked" by the rapid oscillations of the fast waves, thereby suppressing spurious noise and maintaining stability.
This clever trick is not a free lunch. It's an approximation, and like all approximations, it has a cost. The primary cost is a phenomenon known as splitting error.
The true physics unfolds with all processes—advection, rotation, compression—happening simultaneously and continuously. Our split scheme treats them sequentially. The error arises because, in general, the order of operations matters. Imagine a parcel of a chemical tracer being carried by a river where the chemical reaction rate, , changes along the bank. Advecting the parcel downstream and then letting it react for a minute is not the same as letting it react for a minute while it is being advected through a region of changing reaction rates.
This difference is measured by a beautiful mathematical object called the commutator. For two operators, (advection) and (reaction), the commutator is defined as . If the operators commute, , the order doesn't matter, and the splitting would be exact. If they don't, the commutator gives us the leading error of the splitting scheme. For the advection-reaction problem, one can show that this error is:
This tells us, with beautiful clarity, that the splitting error is proportional to the speed of the flow () and the spatial gradient of the reaction rate (). The faster you move through a rapidly changing environment, the larger your splitting error.
Beyond accuracy, a robust scheme must obey the fundamental conservation laws of physics. It must not create or destroy mass, momentum, or energy from nothing. This requires careful design.
Conservation of Mass and Volume: In an ocean model, the total volume of water must be conserved. This means that the change in the sea surface height, , over a large time step must be perfectly consistent with the divergence of the time-averaged water transport, , that was accumulated during the fast barotropic substeps. Enforcing this link, , is essential.
Conservation of Energy: Energy can be exchanged between kinetic energy (motion) and potential energy (compression), but the total energy of the acoustic system should be conserved. This requires a subtle symmetry in the numerical operators. The discrete operator for the pressure gradient, , which converts potential energy to kinetic, and the operator for divergence, , which does the reverse, must be mathematical adjoints of one another, satisfying a property like . Using the same, consistent operators for both the fast inner loop and the final outer update ensures that this energy-exchange pathway is not corrupted.
Synchronization: Finally, after advancing the split components, the model's state must be made self-consistent. In the ocean model, for instance, we have a 3D velocity field and a 2D depth-averaged velocity field. After the time step, we must perform a synchronization step to ensure that the average of the new 3D velocity field exactly equals the new 2D velocity field.
The principle of splitting a problem based on its natural timescales is one of the most powerful ideas in computational science. The split-explicit scheme we've discussed is just one member of a large and versatile family.
It's possible to build higher-order accurate schemes, which reduce the splitting error. A second-order scheme, for instance, might use a leapfrog method for the slow step. This requires even more care in the coupling; to match the leapfrog stencil's centered interval, the fast-wave impulse calculated over one interval must be doubled.
The concept of stiffness isn't limited to waves. It can also arise from very fast chemical reactions or physical processes. For these "stiff source terms," an explicit time step might be limited to a tiny fraction of a second, a constraint completely independent of any advection speed. This leads to a cousin of the split-explicit method called IMEX (Implicit-Explicit) schemes. Here, instead of subcycling the stiff part explicitly, we solve it implicitly—a more computationally demanding approach that offers unconditional stability, allowing us to take large time steps no matter how stiff the process is.
And the split-explicit scheme is not the only way to tackle wave stiffness. An alternative approach is low-Mach preconditioning, which involves mathematically altering the governing equations themselves to artificially slow down the sound waves. This removes the stiffness, allowing a single large time step, but it comes at the cost of distorting the physics of the sound waves. [@problem__id:4070194]
Each of these methods represents a different choice in the fundamental trade-off that every modeler faces: the balance between computational cost, implementation complexity, and physical fidelity. The split-explicit scheme remains a workhorse in modern weather and climate models because it strikes a beautiful and effective balance, allowing us to compute the slow dance of weather systems without getting tripped up by the frantic jig of the sound waves.
Having journeyed through the principles and mechanisms of the split-explicit scheme, we might feel we have a solid grasp of how it works. We’ve dissected its logic, understood its gears and levers. But the true beauty of a great idea in physics or computation is not just in its internal elegance, but in the vastness of the world it unlocks. Why did we bother with all this splitting and subcycling? What does it do for us?
The answer is that it allows us to compute the seemingly incomputable. Nature is a symphony of motions playing out on outrageously different timescales. Imagine trying to listen to the slow, deep rumble of a cello while a piccolo shrieks a thousand notes per second right next to your ear. If you can only listen in tiny, piccolo-sized snippets of time, you will never hear the cello’s melody. The split-explicit scheme is our ticket to hearing the whole orchestra. It is a mathematical trick, to be sure, but it is a trick that allows our models to more faithfully mirror the multi-scaled reality of the world. Let's see where this trick takes us.
Nowhere is the drama of disparate timescales more pronounced than in the Earth's oceans and atmosphere. Consider modeling the vast, slow gyres of the ocean, currents that take years to circle a basin. These are the "cello notes" we are interested in. But the ocean also has a free surface, and ripples on this surface—gravity waves, not unlike those in a bathtub but on a planetary scale—are the "piccolo notes." The speed of these external, or barotropic, gravity waves is given by , where is the acceleration due to gravity and is the ocean depth. For a typical ocean depth of kilometers, these waves travel at about meters per second, or over kilometers per hour! In stark contrast, the internal, or baroclinic, motions associated with temperature and salinity variations, which drive the deep currents, meander along at a leisurely pace of perhaps one or two meters per second.
If we were to use a simple, unsplit explicit scheme, our time step would be dictated by the fastest thing in the model: those zippy surface waves. To keep the simulation stable, a wave cannot cross more than one grid cell per time step (the famous Courant-Friedrichs-Lewy or CFL condition). For a model with a 5-kilometer grid, this would demand a time step of mere seconds. Trying to simulate a century of climate change with time steps of a few seconds would be computationally impossible, even on the mightiest supercomputers.
This is where the split-explicit scheme makes its grand entrance. It recognizes that these two types of motion are physically distinct. It allows us to take one large time step for the slow, interesting baroclinic dynamics, and within that single large step, it subcycles the fast barotropic dynamics with many tiny time steps. The ratio of the wave speeds gives us the required number of subcycles. For our example ocean, the barotropic waves are over 100 times faster than the baroclinic currents, so we would need to take more than 100 small barotropic steps for every one large baroclinic step. The algorithm itself is an elegant dance of prediction and correction: we advance the slow internal motions over the large step, and then the fast surface motions are rapidly updated in their own loop, with the two modes constantly exchanging information about pressure gradients to stay consistent.
Of course, this is not the only way to tackle the problem. An older method, the rigid-lid approximation, simply gets rid of the problem by assuming the ocean surface is a flat, unmoving lid. This filters out the fast surface waves entirely, allowing a large time step. However, it's a brute-force physical approximation that sacrifices the real dynamics of sea-level change. The split-explicit scheme is a more sophisticated and physically faithful numerical solution, and understanding the trade-offs between these methods is a key part of the modeler's art.
The atmosphere presents a similar story, but with a different speed demon. In a non-hydrostatic atmospheric model, which can resolve thunderstorms and other vertical motions, the fastest signals are sound waves, which travel at around meters per second. The winds we want to predict—the advective motions—are much slower. Once again, a split-explicit scheme comes to the rescue, separating the equations into a "slow" part for advection and a "fast" part for acoustics. The fast part is subcycled, allowing the overall model to take large time steps limited by wind speed, not sound speed. The core logic is identical to the oceanographer's problem; only the names of the waves have changed. This is a beautiful example of how a single powerful idea can find a home in different branches of science. The actual implementation involves a careful decomposition of the governing Euler equations, splitting the flux of conserved quantities like momentum and energy into their advective and acoustic components.
Lest we think this separation is always clean, nature and mathematics conspire to introduce complications. When models use terrain-following coordinates to represent flow over mountains or undersea ridges, a subtle but venomous numerical error can arise: the pressure gradient error. In this scheme, the mathematical calculation of the horizontal pressure force involves subtracting two large, opposing terms. Over sloped terrain, small errors in each term fail to cancel, creating a spurious force that can wrongly excite the very fast waves we are trying to handle so carefully. This contaminates the clean separation of modes, causing the model to generate noise and currents where there should be none. Modern modeling has developed incredibly clever fixes for this, such as reformulating the pressure gradient calculation in a way that is guaranteed to be zero for an atmosphere at rest, thereby preserving the delicate balance. This is a testament to the fact that building a good model is not just about having a big idea, but also about the painstaking craft of getting the details right.
The power of the split-explicit idea extends far beyond geophysics. Its core principle—isolating and subcycling stiff terms—is a universal strategy in computational science. Consider a problem from a completely different domain: the simulation of friction in solid mechanics. Imagine a block sliding along a sloped surface. The friction force resisting the motion depends on the normal force pressing the block into the surface. In a typical explicit simulation, this coupling can create a numerical instability. If a small perturbation causes the block to dig into the surface slightly, the normal force increases. This, in turn, increases the friction, which can cause the block to "stick" and "slip" in a non-physical, numerically-induced vibration.
The stability analysis for this simple mechanical system reveals a time step limit that depends on, among other things, the friction coefficient and the stiffness of the contact. The equation looks remarkably similar to the CFL condition for waves. The "stiffness" of the frictional coupling plays the same role as the "speed" of the wave. A high friction coefficient or a stiff penalty contact creates a numerically "fast" process that requires a small time step if treated explicitly. Here again, one could envision a split-explicit scheme where the slow, bulk motion of the object is advanced with a large time step, while the stiff frictional forces are resolved in a subcycled loop. The underlying mathematical structure of the problem is the same.
This same conceptual framework finds one of its most advanced applications back in climate science, in an approach called superparameterization. One of the greatest challenges in climate modeling is representing clouds. They are too small and fast-evolving to be resolved by a global model's coarse grid. Superparameterization addresses this by embedding a small, detailed cloud-resolving model (CRM) inside each grid cell of the large-scale model (LSM).
You can think of this as a "model within a model." The LSM handles the slow, planetary-scale circulation, while the embedded CRM simulates the turbulent, fast-paced life of clouds and convection. How do they talk to each other? Through a split-explicit framework! The LSM takes a large time step, advancing the large-scale winds and temperatures. It then passes this updated large-scale state as a forcing to the CRM. The CRM then runs for many small time steps, simulating the birth and death of clouds and calculating their net effect on temperature, moisture, and momentum. Finally, the CRM passes these averaged effects back to the LSM, which incorporates them as a "convective adjustment" before beginning the next large time step. The LSM provides the "slow" advective tendency, and the CRM provides the "fast" convective tendency. This two-way coupling, mediated by operator splitting, is a revolutionary approach that allows for a much more physically realistic representation of clouds in our climate simulations.
In the world of modern science, an algorithm's elegance is judged not only by its mathematical beauty but also by how well it performs on a supercomputer. The split-explicit scheme, with its many small, explicit inner steps, has a distinct performance profile. Each of these small steps involves simple calculations on a local neighborhood of grid points. This structure is a double-edged sword on modern hardware like Graphics Processing Units (GPUs).
On the one hand, the local, independent nature of the calculations is a perfect match for the massively parallel architecture of GPUs. However, these simple calculations often require fetching more data from memory than they perform actual computations. Their arithmetic intensity is low. This means the speed of the simulation is not limited by the GPU's calculating power (its FLOP/s), but by the speed at which it can shuttle data to and from memory (its bandwidth). The many explicit substeps become memory-bound.
This has led to a fascinating co-evolution of algorithms and hardware optimization. One powerful strategy is kernel fusion. Instead of launching a separate computational job (a "kernel") for each of the, say, eight acoustic substeps, a programmer can write a single, larger kernel that performs all eight updates at once. The genius of this is that the intermediate results—the state of the atmosphere after step 1, step 2, and so on—can be kept in the GPU's ultra-fast on-chip memory (registers), rather than being written out to and read back from the much slower main memory after each tiny step. This dramatically reduces memory traffic, increases the arithmetic intensity, and allows the algorithm to better utilize the GPU's immense computational horsepower.
This dance extends to more complex hybrid schemes like HEVI (Horizontally Explicit Vertically Implicit) methods, which are common in weather forecasting models. These schemes treat the horizontal propagation of waves explicitly (and thus are well-suited for subcycling) but handle the vertical direction implicitly, which requires solving many independent systems of equations along each vertical column of the model. This structure is a perfect match for the GPU's architecture, as thousands of these vertical solves can be performed in parallel, a technique known as "batching".
The story of the split-explicit scheme is therefore not just a story of physics and mathematics, but also one of computer science. It teaches us that the path to better scientific prediction lies at the intersection of a deep understanding of the natural world, the creation of clever numerical methods, and a savvy appreciation for the architecture of our most powerful tools. From the vastness of the ocean to the heart of a silicon chip, it is a journey of discovery across scales, unified by a single, beautiful idea.