Numerical Locking

In computational simulation, our goal is to create digital models that accurately reflect physical reality. However, a strange pathology known as ​​numerical locking​​ can emerge, where the digital model becomes artificially and excessively stiff, failing to capture the true behavior of the object it represents. This discrepancy is not a simple coding bug but a fundamental consequence of approximating complex physics with simple, discrete elements. This article addresses this critical challenge, explaining why a simulated rubber block might behave like diamond or a thin shell like a solid ingot. The reader will first journey through the core ​​Principles and Mechanisms​​, exploring the different forms of locking—volumetric, shear, and membrane—and the underlying mathematical conditions that govern them. Following this, the discussion expands to ​​Applications and Interdisciplinary Connections​​, revealing how this numerical ghost haunts fields from structural engineering to geomechanics and highlighting the ingenious solutions developed to exorcise it.

Imagine you are trying to build a perfect, smooth arch out of large, clumsy Lego bricks. No matter how you arrange them, the result is a jagged, stepped approximation. Worse, if you insist that every single point on the surface of each brick must lie on the perfect curve of the arch, you'll find it's impossible. The only way to satisfy such an overly strict set of rules is to not build the arch at all—to leave the bricks in a pile. The bricks become "locked" in place by an impossible constraint.

This is the essence of ​​numerical locking​​. In the world of computational simulation, our "Lego bricks" are small, simple shapes called ​​finite elements​​ that we use to approximate a complex, continuous reality. Locking is a fascinating, and at times frustrating, numerical pathology where the digital model becomes artificially and astronomically stiff. It's not a bug in the code, nor a mistake in the physics. It is a profound consequence of asking simple shapes to obey complex rules. It's a story of how an overabundance of constraints can paralyze a system.

Let's begin with the most common form of this ailment: ​​volumetric locking​​. Picture a block of rubber. A key property of rubber is that it's nearly ​​incompressible​​. You can twist it, stretch it, and bend it, but you can't easily squeeze it into a smaller volume. The total amount of space it takes up remains almost constant. In the language of physics, the constraint is that the volumetric strain, or the measure of volume change, must be close to zero.

When we model this rubber block on a computer, we chop it up into a grid of finite elements—perhaps tiny cubes or quadrilaterals. Inside each element, we calculate the forces and deformations. The physics of incompressibility is translated into a numerical command: at certain specific locations inside each element, known as ​​quadrature points​​, the calculated volume change must be zero.

Here's where the trouble starts. A simple, low-order element, like a bilinear quadrilateral ($Q_1$), has a very limited repertoire of shapes it can deform into. It can stretch and shear, but its internal motion is not very sophisticated. If we use a standard, "full" integration scheme, we are essentially placing four guards (the four Gauss quadrature points in a $2 \times 2$ rule) inside this simple element, each one shouting, "You shall not change volume at this spot!".

The poor element is caught in a bind. It finds that almost any interesting deformation—even a pure bend which shouldn't change the volume at all—causes some tiny, spurious volume change at one of these guard posts. Faced with an impossible set of commands, the element finds the only way to satisfy all of them is to simply not deform at all. The result? Our simulated rubber block, which should be soft and pliable, behaves as if it's made of diamond. It has "locked."

From an energy perspective, the story is even more dramatic. The strain energy within the material can be split into two parts: a ​​deviatoric​​ part that governs shape change (like shearing), and a ​​volumetric​​ part that governs volume change. The volumetric part is multiplied by a massive penalty number, the ​​bulk modulus​​ ($\kappa$), which approaches infinity for a perfectly incompressible material.

$W_{\text{internal}} = \underbrace{\int_{\Omega} 2\mu\,\boldsymbol{\varepsilon}^{\text{dev}} : \boldsymbol{\varepsilon}^{\text{dev}} \, dV}_{\text{Energy of shape change}} + \underbrace{\int_{\Omega} \kappa\,\text{tr}(\boldsymbol{\varepsilon})\,\text{tr}(\boldsymbol{\varepsilon}) \, dV}_{\text{Energy of volume change}}$

When our simple element creates a tiny, spurious volume change ($\text{tr}(\boldsymbol{\varepsilon}) \neq 0$), that term is multiplied by the enormous $\kappa$, leading to a catastrophic energy penalty. The computer, whose goal is always to find the state of minimum energy, avoids this penalty at all costs by suppressing any and all deformation. The system locks.

The principle of over-constraint extends beyond bulk materials into the delicate world of thin structures like plates and shells. Here, locking manifests in two related but distinct forms: shear locking and membrane locking. The culprit is the dramatic difference in how a thin structure resists different kinds of deformation.

Imagine a very thin steel plate. Bending it is relatively easy. The resistance to bending scales with the cube of its thickness, $t^3$. However, stretching it or shearing it through its thickness is much harder. These resistances scale linearly with its thickness, $t$. As the plate gets thinner and thinner, the difference between $t$ and $t^3$ becomes colossal. A sheet of paper is easy to bend ($t^3$ is tiny), but surprisingly strong if you try to pull it apart (resistance is proportional to $t$).

​​Shear locking​​ occurs when we try to simulate the bending of a thin plate. According to the refined theory of plates (Kirchhoff-Love theory), when a thin plate bends perfectly, there should be no transverse shear strain—no sliding of imaginary layers within the plate past one another. A low-order element, however, often doesn't know how to bend this "purely." Its simple geometry forces it to introduce spurious shear strains as it deforms. Since the resistance to shear ($\sim t$) is so much greater than the resistance to bending ($\sim t^3$), this spurious shear energy completely dominates the true bending energy. The element locks up to avoid the immense shear penalty, refusing to bend properly.

​​Membrane locking​​ is a similar villain that appears when we model curved shells, or even flat plates that are forced into complex curved shapes. A key aspect of many bending deformations is that they are ​​inextensional​​—the middle surface of the shell bends without being stretched, like rolling up that piece of paper. However, a grid of simple, flat-faced elements may find it impossible to follow a curved path without some of them having to stretch a little bit. This spurious stretching, or membrane strain, again brings in the enormous membrane stiffness ($\sim t$), which dwarfs the bending stiffness ($\sim t^3$). The result is the same: the system becomes artificially rigid and "locks" to avoid the parasitic membrane energy.

Are these three locking phenomena—volumetric, shear, and membrane—just a collection of unfortunate coincidences? Not at all. In physics, when we see the same pattern emerge in different contexts, it's often a clue to a deeper, more fundamental principle. And so it is here.

All forms of locking can be understood as a violation of a beautiful mathematical principle known as the ​​Ladyzhenskaya–Babuška–Brezzi (LBB) condition​​, or simply the ​​inf-sup condition​​.

Let's re-imagine our problem. Instead of a single displacement field, we can think of it as a game between two players: a ​​displacement field​​, $\boldsymbol{u}$, which proposes deformations, and a ​​constraint field​​ (like the pressure, $p$, in the incompressibility problem), which enforces the rules. The LBB condition is the rule that ensures a "fair game." It demands that the space of possible displacements is rich and sophisticated enough to respond to any constraint imposed by the pressure field.

When we use a low-order element for displacement but effectively use a very powerful, pointwise set of constraints (either through full integration in a penalty method or by picking an unstable pairing of elements in a mixed method), we violate the LBB condition. The constraint player becomes too powerful for the displacement player. The displacement field is so restricted that it has no good moves left, and the system freezes. Locking is the physical manifestation of this mathematical instability.

If the problem is too many rules, the solution, paradoxically, is to teach our simulation to be a little less rigorous—to practice some "strategic ignorance."

A popular and effective cure is ​​Selective Reduced Integration (SRI)​​. Instead of enforcing the constraint at all four guard posts (quadrature points) in our element, we tell the computer to only check it at a single point in the center. By doing this, we reduce the number of constraints from four to one. This single, weaker constraint is much easier for the simple element to satisfy, and it happily "unlocks" and deforms correctly. This trick works beautifully for both volumetric locking and shear locking. A more formal version of this idea is the ​​$\bar{B}$ method​​, where the problematic strain component is explicitly replaced by its average value over the element, effectively enforcing just one constraint.

For the particularly tricky membrane and shear locking in shells, engineers have developed even more elegant solutions like ​​Assumed Natural Strain (ANS)​​ and ​​Enhanced Assumed Strain (EAS)​​. These methods essentially tell the element, "I know your simple shape makes it hard to calculate strains correctly, so instead of using your calculation, we are going to assume a better, more physically realistic strain field for you to use.".

But this freedom comes at a price. When we tell the element to ignore what's happening at some of its internal points, it can sometimes get lazy. This can lead to a completely different numerical pathology called ​​hourglassing​​. This is a zero-energy deformation mode where the element can wiggle and contort in a bizarre, non-physical way, but because all the motion happens "between" the single quadrature point we are monitoring, the element reports zero strain and zero energy! While locking makes an element too stiff, hourglassing makes it too soft. The art of designing a good finite element is therefore a delicate balancing act: it must be sophisticated enough to avoid locking, but robust enough not to fall prey to hourglassing. This is where the true beauty of computational mechanics lies—not in blind calculation, but in the subtle and intuitive art of crafting numerical tools that are both simple and wise.

We have explored the intricate mechanics of numerical locking, this strange pathology where our digital models become stubbornly, artificially stiff. But this is not merely an abstract mathematical curiosity. It is a ghost in the machine of modern science and engineering, a subtle trap that can appear in the most unexpected places, with consequences ranging from the merely incorrect to the catastrophically misleading. Let us take a journey through some of these fields to see where this ghost lurks, and to appreciate the cleverness and beauty of the methods devised to exorcise it.

Imagine building a bridge or an airplane wing on a computer. Our primary tool is the Finite Element Method, which breaks down a complex structure into a mosaic of simpler pieces, or "elements." The most intuitive approach is to use simple shapes—like tiny straight beams or flat triangles—to approximate the elegant curves of the real object. And here, right at the outset, we encounter the tyranny of the constraint.

Consider a long, slender beam bending under a load. In the real world, as it bends, its cross-sections rotate, but they barely deform through their thickness—a property known as the Kirchhoff constraint. Now, if we model this with simple, linear elements that are only $C^0$ continuous (meaning they meet at points but their slopes can be discontinuous, like a chain of straight sticks), we run into a problem. For the element to bend without introducing spurious shear deformation, it must satisfy a kinematic constraint that its simple polynomial brain cannot handle. The result? The element "locks up." It becomes pathologically stiff against bending, refusing to deform properly. This is ​​shear locking​​, a classic ailment that plagues the modeling of thin structures like beams, plates, and shells,.

It gets worse. For a thin shell, like the wall of a soda can, there's another constraint: it should be able to bend without stretching its surface, a property called inextensional bending. Our simple, flat-faced elements often fail miserably at this, leading to ​​membrane locking​​. In a famous benchmark problem of a pinched cylinder, a naive model can be thousands of times stiffer than the real object, as if we tried to simulate a foil wrapper and ended up with a solid steel ingot.

Why does this matter? It's not just that our calculated deflections are wrong. The stability of the entire structure might be misjudged. Buckling—the sudden, catastrophic failure of a column under compression—is governed by the structure's stiffness. If our model is artificially stiff due to locking, it might predict that a column is perfectly safe, when in reality it is on the verge of collapse. The simulation cries "stable!" right up until the moment the real bridge gives way.

Locking is not just a problem for the civil engineer; it haunts the geoscientist as well. Imagine a water-saturated soil, like the clay beneath a dam. If a load is applied suddenly, the water has no time to escape. Since water is nearly incompressible, the entire soil-water mixture cannot change its volume. This is a physical constraint of incompressibility.

When we try to simulate this with simple, displacement-only finite elements, we hit a wall. These elements, defined only by the motion of their corner nodes, find it incredibly difficult to deform in a way that perfectly preserves their volume. To avoid violating this constraint, they simply refuse to deform at all. The result is ​​volumetric locking​​. The simulated soil becomes as rigid as diamond.

Again, the consequences are profound. A key parameter in soil mechanics is Skempton's B coefficient, which tells us how much the water pressure inside the soil increases when an external load is applied. This is critical for predicting landslides and foundation failures. A locked simulation, by artificially suppressing any volume change, will predict a pore pressure response that is close to zero—a complete and dangerous misrepresentation of the physics. To solve this, modelers use sophisticated "mixed" formulations, where pressure is introduced as an independent character in the play, a Lagrange multiplier that enforces the incompressibility constraint gracefully, without locking the system.

The specter of locking even appears when we model materials that flow, like soils, metals, or polymers. In geomechanics, the Drucker-Prager model describes how a material like sand or concrete yields and flows under pressure. Its "yield surface," which defines the boundary between elastic behavior and plastic flow, is shaped like a cone. A problem arises at the very tip, the apex of the cone, which corresponds to a state of pure hydrostatic pressure. At this singular point, the mathematics does not give a unique direction for the plastic flow.

A naive numerical algorithm trying to "return" a stress state to this surface can get stuck, oscillating or "locking up" at the apex, unable to make a decision. This is a form of ​​algorithmic locking​​, where the problem is not with the element itself, but with the step-by-step procedure for enforcing the material law at a mathematical singularity. The solution is to regularize the model, either by smoothing the tip of the cone with a hyperbola or by closing it with a smooth "cap."

An even more subtle connection emerges in viscoelasticity, the study of materials like plastics that exhibit both solid-like and fluid-like behavior. Consider a model where the material resists the rate of volume change, a property known as bulk viscosity. If we simulate this material's response using very small time steps, $\Delta t \to 0$, the resistance to any volume change in that tiny instant becomes enormous. The discrete time-stepping scheme itself makes the material appear nearly incompressible to the simulation. This, in turn, can induce severe volumetric locking, a pathology created by the interaction of our temporal and spatial discretizations.

For years, the battle against locking was fought with clever but sometimes patchwork fixes: "reduced integration" to relax constraints, "assumed strain fields" to enrich the element's behavior. But a more profound and elegant solution has emerged from a deeper insight into the problem's geometric roots.

The source of locking in standard finite elements is their lack of smoothness. They are built from polynomials that are only continuous ($C^0$), but their derivatives are discontinuous across element boundaries. This is like building a beautifully smooth car body out of faceted panels; the underlying description is fundamentally clunky.

​​Isogeometric Analysis (IGA)​​ proposes a revolutionary change: build your simulation using the same smooth spline functions (like NURBS) that are used in computer-aided design (CAD) to define the geometry in the first place. These functions have higher-order continuity ($C^p$ with $p \ge 1$). This seemingly simple change has a magical effect. Because the basis for displacements is smoother, its derivatives are also smoother. In the case of the Timoshenko beam, the derivative of the deflection field naturally lives in the same space as the rotation field. The constraint $\phi_h = w'_h$ is no longer a contradiction; it is a natural property of the spaces. The discrete kernel of the shear operator is perfectly represented. Locking vanishes, not because of a numerical trick, but because we chose a better mathematical language to describe the physics. This same principle beautifully mitigates membrane locking in shells, allowing them to bend without spurious stretching.

One might think that locking is purely the domain of continuum mechanics. But the ghost appears in the most surprising of places. Consider the world of stochastic differential equations (SDEs), which describe systems driven by random noise, like the jittery motion of a pollen grain in water or the fluctuations of a stock price.

Imagine we simulate two identical particles starting at different locations, but driven by the exact same sequence of random kicks. Intuitively, they should wander randomly but always remain distinct. Yet, when using a common numerical scheme called the Milstein method, a bizarre artifact can occur: for certain step sizes, the two paths suddenly snap together and move as one from that point forward. They become "locked".

This ​​artificial synchronization​​ is a numerical illusion. It happens because the Milstein method includes a correction term that is quadratic in the random noise increment. This quadratic term can, for a non-negligible fraction of the random kicks, create a situation where the effective force pulling the two particles apart momentarily vanishes or even becomes attractive, causing their separation to collapse to zero. It's a subtle conspiracy between the algorithm and the randomness, a purely numerical phenomenon that vanishes with smaller time steps but can completely corrupt a simulation at moderate ones.

From structural engineering to geomechanics, from plasticity to the abstract world of random processes, the theme is the same. When we translate the infinite subtlety of the continuous world into the finite language of a computer, we must be wary of the constraints we impose. Numerical locking is a profound reminder that our models are shadows of reality. Recognizing its signature, understanding its cause, and appreciating the mathematical beauty of its remedies is central to the art and science of computational discovery.