Symmetry Functions

In the realm of computational science, one of the most fundamental challenges is teaching a computer to understand the structure of matter. A simple list of atomic coordinates is insufficient, as it fails to capture the underlying physical laws of the universe. The energy and properties of a molecule or material do not change if it is moved, rotated, or if two identical atoms are swapped. Any meaningful computational description must respect these fundamental invariances. This knowledge gap has historically limited our ability to accurately and efficiently simulate complex atomic systems.

This article introduces atom-centered symmetry functions, an elegant solution to this problem. They provide a mathematical "fingerprint" of each atom's local environment, a description that is invariant by design. By translating complex geometry into a fixed-size vector of numbers, symmetry functions build a bridge between the laws of physics and the algorithms of machine learning. In the following chapters, you will learn how these powerful descriptors are constructed and why they are so effective. "Principles and Mechanisms" will unpack the mathematical details of radial and angular functions, while "Applications and Interdisciplinary Connections" will explore how these fingerprints are used to build next-generation simulation tools and connect ideas across chemistry, materials science, and computer science.

Imagine you want to describe a cathedral to someone who has never seen one. You wouldn't just give them a list of GPS coordinates for every stone. That's a meaningless jumble of numbers. You would talk about the soaring arches, the layout of the columns, the patterns in the stained glass windows. You would describe the relationships between the parts, the geometry that gives the structure its form and function.

Our task in describing an atomic environment to a computer is much the same. A simple list of $x, y, z$ coordinates for each atom is useless on its own. Why? Because the universe has rules, fundamental symmetries that a list of coordinates doesn't respect. If you take a water molecule and move it across the room, its energy doesn't change. If you rotate it, its energy doesn't change. If you swap its two identical hydrogen atoms, its energy still doesn't change. The total energy of a system is ​​invariant​​ to translation, rotation, and the permutation of identical particles. Our description—our language for telling the computer what the molecule is—must have these same beautiful invariances baked into its very grammar.

This is the central challenge, and the solution is wonderfully elegant. Instead of absolute positions, we will build a "fingerprint" of each atom's local world using quantities that are naturally immune to these transformations: ​​distances​​ and ​​angles​​. This is the core idea behind ​​atom-centered symmetry functions​​.

First, we must be practical. An atom's energy is primarily influenced by its immediate neighbors. The pull of an atom a mile away is negligible. So, we make a simplifying assumption: we only care about atoms within a certain ​​cutoff radius​​, $R_c$. We draw a conceptual sphere around our central atom, and everything outside this sphere is ignored. This is our "local environment."

Of course, this is an approximation. Nature's forces, like electromagnetism, have infinite range. But for many chemical interactions, this local picture is an exceptionally good one. To ensure that atoms don't abruptly "pop" into or out of existence at the boundary—which would cause a disastrous, unphysical jump in energy and infinite forces—we use a clever mathematical trick: a ​​smooth cutoff function​​. This function ensures that an atom's influence gently fades to exactly zero as it approaches the edge of the sphere. This guarantees that our energy landscape is smooth and its derivatives (the forces) are well-behaved, a non-negotiable requirement for any physical model.

With our local environment defined, we can start building our invariant fingerprint. We do this with a set of mathematical probes called symmetry functions.

Two-Body Interactions: The Radial Scan

The simplest probe is the ​​radial symmetry function​​, often called $G^2$. Imagine you have a set of detectors you can place at various distances from your central atom. Each detector is tuned to a specific distance, say $R_s$, and it pings more strongly the closer a neighbor is to that exact distance. The mathematical form of such a detector is a Gaussian, $\exp(-\eta(R_{ij} - R_s)^2)$, where $R_{ij}$ is the distance to a neighbor $j$, and $\eta$ controls the sharpness of the detector.

To build the full $G^2$ function, we simply add up the signals from all neighbors within the cutoff sphere:

Here, $f_c(R_{ij})$ is our smooth cutoff function. Because we are summing over all neighbors, the order doesn't matter. If we swap two identical neighbors, the sum remains the same. Voila! ​​Permutation invariance​​ is achieved. And since the function only depends on distances $R_{ij}$, it is automatically invariant to translations and rotations. By using a whole set of these $G^2$ functions with different detector positions $R_s$ and widths $\eta$, we can build a detailed map of the radial distribution of atoms.

Three-Body Interactions: Capturing the Geometry

But distances are not the whole story. Consider a central carbon atom with two neighbors. Are they arranged in a straight line, as in carbon dioxide, or at an angle, as in a water molecule? The pairwise distances from the center could be identical in both cases. To distinguish these, we need to measure ​​angles​​.

This brings us to the ​​angular symmetry function​​, or $G^4$. This function looks at triplets of atoms: the central atom $i$ and two of its neighbors, $j$ and $k$. It is designed to capture the angular "texture" of the environment. A common form looks something like this:

Let's unpack this. The sum is over all unique pairs of neighbors $(j, k)$. The term $(1 + \lambda \cos \theta_{ijk})^{\zeta}$ is the heart of it all. It's a flexible function of the angle $\theta_{ijk}$ between the vectors $\mathbf{r}_{ij}$ and $\mathbf{r}_{ik}$. By choosing different values for the parameters $\zeta$ and $\lambda$ (which can be $+1$ or $-1$), we can create probes that are sensitive to different angular ranges—some might ping for acute angles, others for obtuse ones.

Just like the radial function, this angular function is built from scalars (distances and angles) and sums over neighbors, making it inherently invariant to translation, rotation, and permutation. By combining a library of both radial and angular symmetry functions, we construct a high-dimensional vector, $\mathbf{G}_i$, that serves as a rich, invariant, and differentiable fingerprint of atom $i$'s local world.

Is this fingerprint a perfect, one-to-one mapping of the local environment? The answer, in general, is no. This leads us to the crucial concept of the ​​information bottleneck​​.

We are compressing the infinitely complex geometry of an atomic neighborhood into a finite list of numbers. In this compression, some information is inevitably lost. It's possible for two geometrically distinct environments to accidentally produce the same fingerprint vector. If this happens, our model will be blind to the difference between them, forced to predict the same energy for both. This is a limitation of a finite descriptor set. While more sophisticated descriptors like SOAP (Smooth Overlap of Atomic Positions) are designed to be more "complete" and systematically improvable, the principle of the bottleneck remains.

Furthermore, the choice of where to center our functions is paramount. Why on atoms? Why not, say, on the midpoint of chemical bonds? A clever thought experiment reveals the wisdom of the atom-centric choice. Imagine a bond breaking. The "bond center" would suddenly vanish! The number of terms in our energy sum would change, causing a discontinuous jump in the potential energy. This is a mathematical catastrophe. Atoms, on the other hand, are persistent. Their number is constant. By centering our description on atoms, we build our model on a foundation that is stable and smooth, reflecting the continuous nature of physical reality.

Even with this beautiful theoretical framework, practical challenges abound. The world of computation is not the platonic realm of pure mathematics.

For instance, symmetry functions are perfectly rotationally invariant in theory. But a computer stores coordinates with finite precision. A rotation operation followed by rounding to, say, six decimal places, is not the same as rounding first and then rotating. This tiny discrepancy means that in a real-world implementation, a small "rotational discrepancy" can creep in, breaking the perfect symmetry we worked so hard to achieve.

A more profound issue arises when we build a large dataset. We might use dozens of symmetry functions to describe millions of atomic environments. This creates a massive data matrix. It often turns out that many of our carefully chosen symmetry functions are highly correlated. For example, a radial function looking for atoms at 1.5 Å will give very similar readings across the dataset as one looking for atoms at 1.51 Å. This is called ​​collinearity​​.

This creates a serious problem for the machine learning model. It's like trying to figure out the importance of salt and soy sauce in a recipe when you always add them together in the same ratio. You can't tell which one is responsible for the taste. A high degree of collinearity in our descriptors can destabilize the training process and make the resulting model unreliable. To combat this, practitioners use data science techniques like ​​standardization​​ (rescaling all features to have similar variance) or ​​whitening​​ (a more advanced transformation that removes all correlations). These are essential "preconditioning" steps that clean up the messy, real-world data, allowing the underlying physical relationships to be learned effectively.

In this journey from the fundamental symmetries of the universe to the practicalities of data science, we see the true nature of modern computational science. It is a dance between the elegant, continuous laws of physics and the discrete, finite, and sometimes noisy world of computation. The symmetry function is a bridge between these two worlds, a testament to the ingenuity required to translate the language of nature into a form that a machine can understand.

Having understood the principles behind symmetry functions—how we can distill the complex dance of atoms into a clear, invariant mathematical language—we are now ready to see what this new language allows us to do. The journey from an abstract principle to a practical tool is often where the true beauty of a scientific idea reveals itself. It's like learning the grammar of a new language; the real joy comes when you can finally read its poetry, tell its stories, and use it to communicate new ideas. Here, we will explore the remarkable applications of symmetry functions, seeing how they bridge the gap between the microscopic world of atoms and the macroscopic properties we can observe, and even how they connect seemingly disparate fields of science and engineering.

The primary and most transformative application of symmetry functions is in building machine learning interatomic potentials (MLIPs). For centuries, we have dreamed of a "digital laboratory" where we could simulate materials with perfect accuracy, predicting their properties before ever synthesizing them. This is the promise of molecular dynamics (MD), a method that simulates the motion of atoms over time. But there's a catch: to run an MD simulation, you need to know the forces acting on every single atom at every single instant.

Traditionally, these forces come from one of two sources: either from computationally expensive quantum mechanical calculations, which are too slow for large systems or long simulations, or from classical force fields, which are fast but often lack the accuracy and transferability to describe complex chemical events like bond breaking and forming. MLIPs, built upon symmetry functions, offer a revolutionary third way.

​​From Energy Landscapes to Atomic Motion​​

We have seen that the total energy of a system can be expressed as a sum of atomic energies, each predicted by a neural network that takes a vector of symmetry functions as its input. But how do we get from a static energy value to the dynamic movie of atoms in motion? The answer lies in a concept you learned in introductory physics: forces are simply the negative gradient (the "downhill slope") of the potential energy.

Because the Behler-Parrinello symmetry functions are constructed from smooth, continuous mathematical functions, the entire energy expression is differentiable. We can ask, "How does the total energy change if I give atom $k$ a tiny nudge in the $x$ direction?" The answer is the $x$-component of the force on that atom. Using the chain rule, we can analytically calculate the derivative of the total energy with respect to each atomic coordinate. This involves finding the derivative of the neural network's output with respect to its inputs (the symmetry functions), and the derivative of the symmetry functions with respect to the atomic positions. The smooth cosine cutoff function is absolutely critical here; its derivative goes to zero at the cutoff radius, ensuring that forces don't appear or disappear abruptly, which would be unphysical.

This ability to compute forces accurately and efficiently is the key that unlocks the door to large-scale, long-time molecular dynamics simulations with quantum accuracy. We can now simulate, with unprecedented fidelity, everything from the melting of a crystal to the intricate folding of a protein.

​​Probing the Macroscopic World​​

The power of this framework doesn't stop at forces. Many of the macroscopic properties we care about, like pressure and mechanical strength, are also related to derivatives of the energy. Imagine taking your simulation box and gently squeezing it with a uniform strain. The system's resistance to this deformation is related to its internal pressure and stress. This response is formally captured by the virial stress tensor, a quantity that can be calculated by taking the derivative of the total energy with respect to an applied strain.

Once again, because our entire model is differentiable, we can derive an analytical expression for the virial contribution from each atom. This provides a direct, rigorous link between the microscopic interactions described by the symmetry functions and the macroscopic, experimentally measurable properties of a material. We are no longer just watching atoms jiggle; we are predicting the very engineering properties that determine a material's usefulness.

While their genesis was in building potentials, the utility of symmetry functions extends far beyond. At their heart, they are a general-purpose tool for describing local geometric patterns in a way that is independent of orientation. This "universal language" for atomic patterns has profound connections to chemistry, machine learning, and computer science.

​​The Chemist's "Eye": Distinguishing Atomic Geometries​​

An experienced chemist can look at a molecular structure and instantly recognize key motifs: a tetrahedral carbon, a planar benzene ring, the ordered lattice of a crystal. Can our symmetry functions do the same? Can they learn to "see" like a chemist?

The answer is a resounding yes. Let's consider a simple, ideal crystal, like a simple cubic lattice. Every atom in the bulk has six nearest neighbors at a distance $a$, twelve second-nearest neighbors at $a\sqrt{2}$, and so on. A radial symmetry function centered near these distances will produce a distinct, quantitative "fingerprint" that uniquely identifies this structure based on its coordination shells.

But the real test is in distinguishing more subtle differences. Take carbon, the cornerstone of life, which can exist in different bonding environments, or hybridizations: $sp$ (linear), $sp^2$ (trigonal planar), and $sp^3$ (tetrahedral). These environments have different coordination numbers (2, 3, and 4) and different bond angles (180°, 120°, and 109.5°). By using a vector of symmetry functions—some radial functions to probe the different bond lengths, and some angular functions to probe the characteristic bond angles—we can create a unique fingerprint in a high-dimensional feature space for each hybridization state. Remarkably, it's possible to find a minimal set of these functions, perhaps just one or two, that is sufficient to tell these three fundamental chemical environments apart, even in the presence of thermal noise.

We can see this discriminative power in action with a simple thought experiment: consider a tiny molecule of three atoms. Are they arranged in a straight line or in an equilateral triangle? To our eye, they are obviously different. To a computer fed with raw coordinates, their difference is not so obvious, as it depends on their orientation. However, their symmetry function fingerprints are fundamentally different and this difference persists no matter how you rotate the molecules. We can even calculate the "distance" between their fingerprints to quantify how different their geometries are.

​​Connections Across Disciplines​​

This idea of creating a rotation-invariant description of a local pattern is not unique to chemistry. It's a fundamental problem in many fields, most notably in computer vision. This leads to a beautiful analogy between symmetry functions and the tools of modern artificial intelligence.

• ​​Analogy to Convolutional Neural Networks (CNNs):​​ In image recognition, a CNN uses "filters" or "kernels" that slide across an image to detect local features like edges, corners, or textures. An atom-centered symmetry function is conceptually analogous to one of these filters: it looks at a local neighborhood (defined by the cutoff radius) and outputs a value describing the pattern it sees. But there is a crucial and elegant difference. A standard CNN filter is translation-equivariant (if you shift the input, the output shifts), but it is not rotation-invariant (a rotated cat looks different to the filter). The network must learn to recognize rotated cats by seeing many examples. Behler-Parrinello symmetry functions, by contrast, are rotation-invariant by design. They build a fundamental law of physics—that physical reality doesn't depend on your point of view—directly into the representation. The final summation of atomic energies, $E = \sum_i E_i$, is then analogous to a global pooling layer in a CNN, which aggregates the features into a single, permutation-invariant output.

• ​​Beyond Energy Prediction:​​ Since symmetry functions provide a general way to describe atomic structures, why should we limit ourselves to predicting energy? Any property that depends on atomic structure and obeys the same physical symmetries—invariance to translation, rotation, and permutation—can be learned using this approach. For example, one could train a classifier to predict whether a given atomic configuration is thermodynamically stable, or whether it would be an effective catalyst for a particular reaction. The principle remains the same: represent the geometry in an invariant way, then feed this representation into a suitable machine learning model, be it a neural network or a support vector machine.

• ​​Connections to Kernel Methods:​​ The vector of symmetry functions, $\mathbf{G}_i$, serves as a fingerprint for the local environment of atom $i$. In another branch of machine learning, kernel methods (like Gaussian Process Regression or Kernel Ridge Regression) work by defining a "kernel function" $K(i, j)$ that measures the similarity between two objects. A simple and powerful way to do this for atoms is to use the dot product of their fingerprints: $K(i,j) = \mathbf{G}_i \cdot \mathbf{G}_j$. This shows that the descriptor-based approach of NNs and the kernel-based approach of GPR are deeply related. One can even construct a total kernel as a weighted sum of a radial kernel and an angular kernel, allowing the model to tune the relative importance of distance-based and angle-based information.

In our quest to build in physical symmetries, we must be careful not to enforce too much symmetry. Standard symmetry functions, based on distances and angles, are invariant under all rotations and reflections. They cannot tell the difference between a "left-handed" and a "right-handed" molecule (a property known as chirality), because the two are mirror images of each other. If the property we want to predict does depend on chirality—as is the case for many biological molecules—then using these standard functions would be a mistake, as they would map two different molecules to the exact same fingerprint. In such cases, the framework must be extended to include features that can detect chirality, such as those based on triple products of vectors, which change sign under reflection. This is a beautiful reminder that the art of feature engineering lies in precisely matching the symmetries of our representation to the symmetries of the problem we are trying to solve.

In the end, the story of symmetry functions is a powerful illustration of the importance of representation. The success of modern machine learning is not just about bigger networks or more data; it's about finding the right language to describe the world. By teaching a computer the fundamental symmetries of physics, we don't just make the learning problem easier; we create a tool of remarkable power and generality, a tool that is truly a digital alchemist's dream.