Nonlocal Models

In classical physics and engineering, our understanding of the material world is built on the principle of locality: the properties at a point depend only on what is happening at that exact point. This powerful idea, known as the continuum hypothesis, allows us to describe everything from bridges to planets with elegant differential equations. However, this view is an approximation, one that breaks down dramatically when we examine materials at very small scales or under extreme conditions. When material failure concentrates into sharp cracks, or when components shrink to the nanoscale, local theories predict physical absurdities, revealing a fundamental gap in our classical models.

This article confronts this challenge by exploring the world of nonlocal models, a theoretical framework where a point's behavior is influenced by its entire neighborhood. First, under "Principles and Mechanisms," we will dissect the fundamental idea of nonlocality, contrasting its integral and differential formulations and examining the subtleties and paradoxes that arise. Following this, the "Applications and Interdisciplinary Connections" chapter will showcase the predictive power of these models, demonstrating how they resolve long-standing problems in fracture mechanics, explain the unique behavior of nanomaterials, and even bridge the gap between material science and theoretical chemistry.

Imagine you are trying to describe the temperature in a room. For most purposes, you can pick any point in the room, stick a thermometer there, and get the temperature of that point. The air molecule at that exact location determines the reading. This is the essence of a ​​local​​ theory. The properties at a point—stress, temperature, electric field—depend only on what is happening at that exact point. This idea, known as the ​​continuum hypothesis​​, is the bedrock of classical physics and engineering. It's an incredibly powerful approximation that allows us to use the elegant tools of calculus to describe the world. But it is, fundamentally, an approximation. And like all approximations, it has its limits.

When does this comfortable, local picture start to crumble? It happens when we "zoom in" too close, either by looking at very small things or at phenomena that change very rapidly in space. The continuum hypothesis rests on a principle called ​​scale separation​​. It assumes that the characteristic length of the material's microstructure, let's call it $l_m$ (think of grain size in a metal, or the spacing between fibers in a composite), is vastly smaller than the length scale over which the physical fields (like strain) are changing, let's call this $L_g$. In mathematical shorthand, the local world is valid when $l_m \ll L_g$.

Nature, however, frequently violates this condition.

Consider a piece of metal being pulled until it starts to fail. The deformation, which was once spread out evenly, will suddenly concentrate into a very narrow region called a ​​shear band​​. Inside this band, the strain changes dramatically over a distance comparable to the metal's grain size. Here, $L_g$ (the width of the band) is on the same order as $l_m$!. A classical, local theory trying to describe this situation runs into a catastrophic problem. It predicts that the band should be infinitely thin, leading to physically absurd results, like the energy required to break the material being zero. This isn't just a mathematical curiosity; in computer simulations using local models, the predicted failure zone shrinks with the size of the computational mesh, giving answers that depend on the simulation's setup, not the material's physics.

Or think of a tiny component in a micro-electromechanical system (MEMS). If the component itself is only a few dozen grains of material across, then its overall size $H$ becomes the characteristic length $L_g$. Again, the condition $l_m \ll L_g$ is no longer a given. Experiments confirm that in such cases, smaller can be stronger; the measured strength of a material starts to depend on the size of the specimen being tested, a ​​size effect​​ that local theories simply cannot explain.

Even in large objects, fast vibrations can reveal the breakdown of locality. When the wavelength $\lambda$ of a wave propagating through a material becomes comparable to the microstructural length $l_m$, the wave starts to "feel" the discrete nature of the material. This leads to ​​dispersion​​, where the wave's speed depends on its frequency, another phenomenon that purely local models miss. In all these cases—sharp gradients, small sizes, and short wavelengths—the comfortable assumption that a point only cares about itself is no longer tenable. We need a new idea.

The new idea is wonderfully intuitive: what happens at a point is determined not just by the state at that point, but by a weighted average of the states in its entire neighborhood. This is the principle of ​​nonlocality​​. A material point, before it decides how much stress to feel, effectively "polls its neighbors."

The most direct way to express this is with an integral. In classical local elasticity, the stress $\boldsymbol{\sigma}$ at a point $\mathbf{x}$ is directly proportional to the strain $\boldsymbol{\varepsilon}$ at that same point: $\boldsymbol{\sigma}(\mathbf{x}) = \mathbf{C}:\boldsymbol{\varepsilon}(\mathbf{x})$, where $\mathbf{C}$ is the stiffness tensor. The simplest nonlocal generalization of this idea states that the stress at $\mathbf{x}$ is a spatial average of the local elastic response from all other points $\mathbf{x}'$ in the body:

The magic is in the function $\alpha(r; \ell)$, called the ​​influence function​​ or ​​kernel​​. It acts as a weighting factor. Points $\mathbf{x}'$ that are very close to $\mathbf{x}$ have a strong influence, while points far away have a diminished influence. The function introduces a new fundamental material parameter, $\ell$, the ​​internal length scale​​. This length scale dictates the "range of conversation" between material points and is physically tied to the microstructural length $l_m$ we discussed earlier.

For this idea to be consistent, the kernel must satisfy a simple, common-sense rule: if the strain is uniform everywhere, $\boldsymbol{\varepsilon}(\mathbf{x}') = \boldsymbol{\varepsilon}_0$, the nonlocal model must give back the classical result. This means the nonlocal stress should also be uniform and equal to $\mathbf{C}:\boldsymbol{\varepsilon}_0$. This only works if the integral of the influence function over all space is equal to one:

This normalization ensures that the democratic polling process doesn't artificially inflate or deflate a consensus. If everyone in the neighborhood agrees, the average opinion is simply that opinion. In the limit where the internal length $\ell$ goes to zero, a properly defined kernel morphs into the Dirac delta function, $\delta(\mathbf{x}-\mathbf{x}')$, which is zero everywhere except at $\mathbf{x}' = \mathbf{x}$. In this limit, the integral collapses and we recover the purely local law. The classical theory is thus neatly contained within the nonlocal framework as a special case.

While the integral form is the most fundamental and physically intuitive picture of nonlocality, integrals can be notoriously difficult to work with. It would be wonderful if we could capture the same physics using a simpler differential equation.

Amazingly, for one very special choice of kernel, this is possible! If we are working in an infinite space (to avoid messy boundary effects) and we choose a specific exponentially decaying kernel, the so-called ​​Yukawa​​ or ​​Helmholtz kernel​​, the complex integral law becomes exactly equivalent to a much simpler-looking differential equation. In one dimension, this magical kernel is $\alpha(x) = \frac{1}{2\ell} \exp(-|x|/\ell)$. With this kernel, the integral law

is mathematically identical to the differential equation:

This is a beautiful result, a bridge between the world of integral convolutions and the world of differential operators. The differential form is often called a ​​gradient model​​, because if you rearrange it for strain, the strain depends on stress and its second derivative (a gradient of a gradient). The general form is often written as $(\mathbb{I} - \ell^2 \nabla^2) \boldsymbol{\sigma} = \mathbf{C} : \boldsymbol{\varepsilon}$.

This equivalence provides a powerful shortcut. However, the magic comes with fine print. The equivalence is exact only for this specific kernel and only in an infinite domain. Using a different kernel (like a Gaussian one) or, more importantly, dealing with a real-life object with boundaries, breaks the exact equivalence. The differential form then becomes an approximation, one that must be handled with great care.

The convenience of the differential formulation comes at a steep price, revealing deep subtleties in the transition from local to nonlocal physics. The trouble, as is often the case in physics, lies at the boundaries.

The differential form $(\mathbb{I} - \ell^2 \nabla^2) \boldsymbol{\sigma} = \mathbf{C} : \boldsymbol{\varepsilon}$ involves higher-order derivatives of the stress field compared to any equation in classical mechanics. Mathematically, a higher-order differential equation requires more boundary conditions to obtain a unique solution. But where do these extra boundary conditions come from? Classical mechanics only tells us how to specify forces or displacements on a boundary, not "hyper-tractions" or other higher-order quantities. The integral model, by contrast, needs no such extra conditions; the classical ones suffice. This discrepancy reveals that the differential form is not just a simple replacement for the integral; it's a different mathematical beast that poses new physical and mathematical questions.

The consequences of this can be dramatic, leading to what is known as the ​​cantilever paradox​​. Consider the textbook problem of a simple cantilever beam clamped at one end and subjected to a point force $P$ at the other. If you apply the differential nonlocal model, a bizarre thing happens. The equilibrium equations for a beam dictate that in the region without distributed load, the second derivative of the bending moment must be zero, $M''(x)=0$. When you plug this into the nonlocal constitutive law, $M(x) - \ell^2 M''(x) = EI \kappa(x)$, the nonlocal term $\ell^2 M''(x)$ vanishes completely! The model paradoxically predicts a response that is identical to the classical, local theory, showing no size effects whatsoever. The very nonlocality you sought to include has vanished into thin air. This cautionary tale teaches us a profound lesson: the physical laws of equilibrium and the constitutive laws of materials are deeply intertwined, and a seemingly sophisticated model can fail in unexpected ways if this interplay isn't fully respected.

The challenges posed by the simple differential model have spurred physicists and engineers to develop a rich family of nonlocal theories, each with its own strengths and subtleties.

• ​​Strain-Driven vs. Stress-Driven Models:​​ The original formulation we discussed, where strain determines the nonlocal stress, is called ​​strain-driven​​. Some of the paradoxes can be resolved by literally inverting the constitutive logic. In a ​​stress-driven​​ model, it is the stress field that is averaged to determine the local strain: $\varepsilon(\mathbf{x}) = \int \alpha(|\mathbf{x}-\mathbf{x}'|) \, (\mathbf{C}^{-1}:\boldsymbol{\sigma}(\mathbf{x}')) \, \mathrm{d}V_{\mathbf{x}'}$. This might seem like a minor algebraic rearrangement, but it leads to a model with different and sometimes more desirable mathematical properties.

• ​​Strongly vs. Weakly Nonlocal Models:​​ This brings us to a useful classification. Models based on an explicit integral convolution are often called ​​strongly nonlocal​​. They embody the "action at a distance" principle in its purest form. In contrast, models that build nonlocality by including gradients of strain (or other internal variables) directly into the material's energy function are called ​​weakly nonlocal​​ or ​​strain-gradient​​ theories. For instance, the energy density might depend on both $\boldsymbol{\varepsilon}$ and $\nabla\boldsymbol{\varepsilon}$. These models also introduce an internal length scale $\ell$ and can capture size effects, but they do so without an explicit integral. The mathematical price they pay is a demand for "smoother" displacement fields, as the energy now depends on second derivatives of displacement, whereas the integral model can get by with less stringent smoothness requirements.

This journey from the simple local world to the interconnected nonlocal one shows science in action. An elegant and powerful theory (the classical continuum) meets its limits when faced with new experiments. A new, intuitive idea (nonlocality) is proposed, leading to a flurry of different mathematical formulations, each with its own beauty, power, and unexpected pitfalls. The quest to understand how a material point communicates with its neighbors continues, pushing the frontiers of physics and engineering.

What does a point in space "know"? If you subscribe to the beautifully simple worldview of classical physics, a point knows only about the conditions—the fields, the forces, the strains—at its own infinitesimal location. This is the principle of locality. It's a tremendously powerful and successful idea, the bedrock upon which much of classical mechanics and field theory is built. It's like viewing the world through a pinhole; you can build up a complete picture, but only by assembling an infinite number of independent, localized snapshots.

But what happens when this approximation breaks down? What if a point isn't so ignorant of its surroundings? What if it feels the influence of its neighbors, not just infinitesimally close, but across a small yet finite distance? This is the world of nonlocal models. Stepping into this world is not a mere mathematical exercise; it is a necessary journey we must take to resolve paradoxes, explain new experiments, and unify our understanding of phenomena across a staggering range of disciplines. We find that the simple question, "What does a point know?", leads us to a deeper and more powerful description of nature.

Let's begin with something dramatic: the way things break. Imagine stretching a metal bar until it starts to fail. A classical, local model of the material assumes that each point in the bar decides whether to soften and yield based only on the strain at that exact point. If you use this model in a computer simulation, you run into a catastrophe. As the material begins to fail, the simulation predicts that all the deformation will concentrate into a crack of exactly zero thickness. This leads to absurdities like infinite strains and a predicted fracture energy that depends not on the material itself, but on the fineness of the computational mesh you chose! The model has lost its connection to physical reality.

The root of this "pathological" behavior is the local assumption. A real crack is not a mathematical line of zero width; it is a process zone where bonds are stretched and broken over a small but finite volume. Nonlocal models rescue us from the paradox by building this physical truth directly into the mathematics. They introduce an intrinsic material length scale, let's call it $\ell$. In these models, the stress or damage at a point is determined not by the local strain, but by a weighted average of the strain in a neighborhood of size $\ell$.

This single change has profound consequences. The mathematical problem becomes well-posed again, and the crack now has a natural, finite width related to $\ell$. The calculated energy to break the material now converges to a real, physical value. This length scale $\ell$ is not just a fudge factor; it is a genuine material property that can be connected to fundamental quantities like the material's stiffness $E$, its surface energy $\Gamma$, and its theoretical strength $\sigma_{\mathrm{th}}$ through relations like $\ell \propto E\Gamma/\sigma_{\mathrm{th}}^{2}$.

We see this principle in action when studying the fracture of thin, ductile metal sheets. Experiments show that the measured toughness of a sheet can depend on its thickness, which is puzzling if toughness is supposed to be an intrinsic material property. A nonlocal model resolves the puzzle beautifully. It recognizes that there are two competing length scales: the geometric thickness of the sheet, $t$, and the intrinsic material length scale of the fracture process, $\ell$. The observed behavior depends on their ratio, $t/\ell$. By using a nonlocal framework, we can untangle these effects and extract the true, size-independent fracture properties of the material, a task impossible for a purely local theory.

The need for an intrinsic length scale becomes even more apparent when we shrink our entire world down to the nanoscale. Consider a tiny cantilever beam, thousands of times thinner than a human hair, vibrating like a microscopic diving board in a vacuum. Such devices are the heart of nanoelectromechanical systems (NEMS). If we use classical, local elasticity to predict its resonant frequency, we get a certain number. But when we perform the experiment, we often measure a lower frequency. The beam appears to be "softer" or more flexible than our classical theory predicts.

Once again, nonlocality is the key. In a real crystal, atoms are connected by bonds. The stress at one point is not just a function of how the lattice is deformed there, but also how it's deformed several atoms away. A nonlocal elasticity model, like that proposed by Eringen, captures this by defining the stress as an integral of the strain field over a small neighborhood. This "action at a distance" between atoms provides an additional compliance mechanism that is absent in local theory. The result? The nonlocal model predicts a softening of the beam and a lower resonant frequency, precisely matching the experimental observations. What appears as an anomaly in the local worldview becomes a natural prediction in the nonlocal one.

Nonlocal thinking is a powerful tool for bridging the gap between different physical scales. How do we describe the properties of a bulk material, knowing that it's made of a complex, heterogeneous microstructure? Imagine a composite made of stiff ceramic fibers embedded in a soft polymer matrix. Far from any edge, we can use a local "homogenization" theory to find an effective, average stiffness for the composite. But what happens right at the surface? The repeating pattern of fibers and matrix is abruptly cut off.

A classical local model is blind to this truncation. A nonlocal model, on the other hand, intrinsically captures it. In a model like Peridynamics, where each point interacts with its neighbors within a finite "horizon" $\delta$, a point near the surface has fewer neighbors to interact with. Its "interaction neighborhood" is incomplete. This naturally and automatically changes the effective stiffness in a thin boundary layer near the surface, leading to size-dependent effects that local homogenization misses entirely.

This same idea of a spatially dependent response appears in a completely different field: theoretical chemistry. Consider an ion dissolved in water. The simplest "implicit solvent" model treats water as a uniform dielectric goo, characterized by a single number, the relative permittivity $\epsilon_s \approx 80$. This is a local model. But water is made of molecules, and its ability to screen an electric field depends on the distance from the ion. Close to the ion, where its electric field varies rapidly, the water molecules cannot orient themselves perfectly to screen the charge. Far away, where the field is weak and slowly varying, they can.

A nonlocal dielectric model captures this beautifully by making the permittivity a function of the wavevector, $\epsilon(\mathbf{k})$. The wavevector $\mathbf{k}$ is inversely related to wavelength; large $|\mathbf{k}|$ corresponds to rapidly varying fields (short distances), while small $|\mathbf{k}|$ corresponds to slowly varying fields (long distances). The nonlocal model correctly states that as $|\mathbf{k}| \to \infty$ (very close to the ion), the screening is weak and $\epsilon(\mathbf{k}) \to 1$. As $|\mathbf{k}| \to 0$ (far from the ion), the screening is strong and $\epsilon(\mathbf{k}) \to \epsilon_s$. This phenomenon, called "dielectric smoothing," leads to a more realistic, finite distribution of polarization charge around the ion and provides much more accurate calculations of solvation free energies, a cornerstone of computational biochemistry.

The power of the nonlocal framework extends even to the chaotic world of fluid dynamics. Turbulence, the swirling, unpredictable motion of fluids, is the epitome of a nonlocal phenomenon. An eddy at one location is influenced by the entire history and structure of the flow around it. Prandtl's famous "mixing length" hypothesis was a brilliant local approximation to model the Reynolds shear stress that drives turbulence. But it's still an approximation.

We can construct a more sophisticated model by defining a nonlocal mixing length, where the effective mixing at a point is a weighted average over a surrounding region. This integral formalism provides a systematic way to incorporate more physics into the model. For instance, we can use it to enforce consistency with the well-known "law of the wall"—the universal velocity profile observed near a solid boundary—by carefully choosing the form of our averaging kernel. This shows how the mathematical language of nonlocality serves as a powerful toolbox for building better physical theories.

Finally, having seen the predictive power of these models, we must ask a practical question: how do we compute with them? An integral that connects every point to a neighborhood of other points sounds like a computational nightmare, especially on a modern supercomputer with thousands of processors. Yet, here too, the physics of nonlocality guides us to an elegant solution.

In most physical applications, the nonlocal interaction, while finite, is not infinite in range. The averaging kernel has a "compact support" or a finite horizon $\delta$. A point only cares about its neighbors within this horizon. To parallelize the calculation, we can use a strategy called domain decomposition with "halo exchange." The computer divides the problem into subdomains, one for each processor. For a processor to compute the values in its own domain, it only needs to know the state of the system in a thin "halo" region of its neighbors' domains. Before each calculation step, the processors simply exchange this halo data with their immediate neighbors. Communication remains local, and the problem becomes tractable. The physical nature of the short-range nonlocality dictates the optimal parallel computing strategy.

From the catastrophic elegance of a crack propagating through steel, to the subtle dance of water molecules around an ion, to the swirling chaos of a turbulent river, the principle of nonlocality provides a unifying thread. It reminds us that sometimes, to truly understand a point, we must appreciate the neighborhood in which it lives. This shift in perspective is more than just a correction to our old theories; it is a doorway to a richer, more accurate, and more interconnected understanding of the physical world.