When modeling physical systems at the atomic scale, a fundamental challenge arises: physical laws do not care about our coordinate systems. If you translate a molecule in space, rotate a crystal lattice, or swap the arbitrary indices of two identical hydrogen atoms, the total potential energy of the system remains exactly the same.
Yet, if you feed raw Cartesian coordinates $(x, y, z)$ into a standard neural network, these trivial transformations produce entirely different input vectors. As noted by Wang et al., standard neural network architectures cannot directly process raw atomic coordinates because they fail to inherently respect these fundamental physical symmetries. Forcing a model to learn these invariants through data augmentation is computationally wasteful and mathematically imprecise. Instead, modern atomistic machine learning relies on symmetry-invariant local descriptors—mathematical representations that map 3D atomic configurations into numerical vectors that are natively immune to translation, rotation, and permutation.
The Locality Assumption
To represent a complex molecular or material system efficiently, models must break the global structure down into manageable pieces. This is justified by the "locality assumption." As discussed by Gubaev et al. and Wen et al., the total potential energy of a system can be modeled as the sum of individual atomic energies:
$$E_{\text{total}} = \sum_{i} E_i(\mathcal{N}_i)$$
where each local energy contribution $E_i$ is determined solely by the environment $\mathcal{N}_i$ of neighboring atoms within a specific cutoff radius around atom $i$. The core task of a descriptor is to encode this local neighborhood $\mathcal{N}_i$ such that the representation remains invariant under physical transformations.
Atom-Centered Symmetry Functions
One of the earliest and most intuitive ways to build an invariant representation is to transform raw coordinates into geometric invariants, such as distances and angles. Eshet et al., Kauffmann et al., and Tadmor et al. utilize sets of atom-centered symmetry functions (ACSFs) to achieve this.
Radial symmetry functions compute smooth, distance-dependent sums over neighboring atoms to capture the local density at various shells, while angular symmetry functions evaluate the cosines of angles between all pairs of neighbors. Because distances and angles do not change when a molecule is shifted or rotated, the resulting feature vector is perfectly invariant to translation and rotation. Furthermore, because the functions sum over all neighbors of a given chemical species, they are completely insensitive to how those neighbors are indexed, satisfying permutational symmetry.
Local Reference Frames
An alternative to hand-crafting geometric invariants is to dynamically redefine the coordinate system for each atom. Rather than looking at the system from a global origin, we can attach a unique, rigid coordinate frame to each individual atom based on the positions of its nearest neighbors.
Wang et al. demonstrate this mechanism in the Deep Potential method. By establishing a local reference frame for every atom, the relative coordinates of surrounding atoms are projected into this local coordinate system. If the entire system rotates globally, the local frames rotate with it, meaning the projected coordinates remain completely unchanged. This elegantly guarantees $SE(3)$ (translation and rotation) invariance. To handle permutational symmetry within this frame, the descriptors are sorted systematically according to chemical species and interatomic distances, ensuring that swapping identical atoms does not alter the network's input.
Global Frame (X, Y) Local Frame (x_local, y_local)
Y y_local
^ ^
| O (Atom i) | O (Atom i) ---> Neighbor 1
| +-------> x_local
| /
+-------------> X v Neighbor 2
(Global rotation changes (Local frame rotates *with* the atom;
coordinates of Atom i) relative coordinates stay invariant)
However, independently aligning every local neighborhood can sometimes discard vital global context. Hua et al. point out a subtle flaw in standard local frames: they can make atoms at distinct, symmetry-equivalent positions in a crystal lattice completely indistinguishable by forcing their local neighborhoods into identical canonical orientations. To preserve this global structural relationship without breaking invariance, they decompose the coordinate frame into the product of an equivariant global frame and an invariant local frame. For infinite periodic structures, Ito et al. construct dynamic local frames that evolve across layers, ensuring that predictions remain invariant to the specific mathematical unit-cell representation chosen to store the crystal.
Permutation Invariance Beyond Sorting
While sorting descriptors or summing over neighbors works well for many-body potentials, other approaches enforce permutational symmetry algebraically or through continuous fields.
- Permutationally Invariant Polynomials (PIPs): As implemented by Drehwald et al., inter-nuclei distances can be transformed into Morse variables and explicitly symmetrized into monomials and polynomials, creating a representation that is mathematically immune to the ordering of identical atoms.
- Sorted Atomic Contributions: Stefan Gugler achieves permutation invariance by decomposing electronic energy into atomic contributions and organizing them into sorted lists, bypassing the high computational cost of matrix diagonalizations.
- Continuous Density Maps: When dealing with highly fluid environments like solvent dynamics, tracking individual coordinates becomes impractical due to particle indistinguishability. Gu et al. solve this by placing a Gaussian kernel on each nearby solvent molecule and summing them to create a continuous density map. This eliminates the need for manual volume definitions or tracking individual particle indices.
Satisfying Quantum Gauge Symmetries
As machine learning expands from predicting classical potential energy surfaces to predicting quantum mechanical properties, descriptors must respect additional mathematical symmetries. For instance, electronic states depend on the choice of orbital basis sets, which introduce abstract mathematical rotations (gauge symmetries).
To bridge this gap, Chen et al. and the authors of DeePKS-kit developed descriptors based on the local electronic density matrix. By projecting the density matrix onto local atomic basis functions and extracting its eigenvalues, they create a representation of the local electronic environment. Because the eigenvalues of a matrix are fundamentally invariant to a change of basis (mathematical rotation of the orbital basis), this approach satisfies physical gauge symmetries naturally. This ensures that the machine learning model remains invariant not only to how the molecule is oriented in physical space, but also to how the underlying quantum mechanics are mathematically framed.
Key Foundational Work
Building Markov state models with solvent dynamics (2013) — contributed 2 concept elements to this direction.
Ab initioquality neural-network potential for sodium (2010) — contributed 1 concept elements to this direction.
Demonstrating the Transferability and the Descriptive Power of Sketch-Map (2013) — contributed 1 concept elements to this direction.