Abstract
The applications of machine learning techniques to chemistry and materials science become more numerous by the day. The main challenge is to devise representations of atomic systems that are at the same time complete and concise, so as to reduce the number of reference calculations that are needed to predict the properties of different types of materials reliably. This has led to a proliferation of alternative ways to convert an atomic structure into an input for a machine-learning model. We introduce an abstract definition of chemical environments that is based on a smoothed atomic density, using a bra-ket notation to emphasize basis set independence and to highlight the connections with some popular choices of representations for describing atomic systems. The correlations between the spatial distribution of atoms and their chemical identities are computed as inner products between these feature kets, which can be given an explicit representation in terms of the expansion of the atom density on orthogonal basis functions, that is equivalent to the smooth overlap of atomic positions power spectrum, but also in real space, corresponding to n-body correlations of the atom density. This formalism lays the foundations for a more systematic tuning of the behavior of the representations, by introducing operators that represent the correlations between structure, composition, and the target properties. It provides a unifying picture of recent developments in the field and indicates a way forward toward more effective and computationally affordable machine-learning schemes for molecules and materials.
Introduction
To effectively deploy machine learning (ML) in chemistry and materials science, 3D atomic coordinates must be converted into a numerical "representation" or descriptor. A high-quality representation must satisfy three fundamental requirements: it must be invariant to rotations, translations, and permutations of identical atoms; it must be complete enough to distinguish between distinct chemical structures; and it must be concise to avoid the computational inefficiencies associated with high-dimensional data.
The field has historically seen a proliferation of diverse descriptors, ranging from Behler-Parrinello symmetry functions to the Smooth Overlap of Atomic Positions (SOAP). This diversity has made it challenging to compare methods or understand the underlying physical principles that drive their performance. This paper addresses this by providing a "Rosetta Stone" for atomic representations, grounding them in a unified mathematical framework based on smoothed atomic density correlations.
The Smoothed Density Framework
The authors utilize Dirac bra-ket notation to define an abstract "feature ket" $|A\rangle$ for an atomic configuration. In this approach, each atom is represented not as a point, but as a smoothed "cloud" (typically a Gaussian function). This smoothed density allows for a continuous representation of the atomic environment.
As shown in Figure 1, an atomic environment—such as that of an ethanol molecule—can be decomposed into densities associated with specific chemical species (C, O, H). By formally averaging these densities over rotation and translation groups (Haar integration), the representation becomes naturally invariant to these symmetries.
N-body Correlations and the Body-Order Hierarchy
To capture the complex spatial arrangements of atoms, the framework employs tensor products of the atomic density. This leads to a clear hierarchy of body-order correlations:
- $\nu=1$ (2-body): Captures information about atom types and radial distances.
- $\nu=2$ (3-body): Captures angular information (distances and angles between triplets of atoms).
- $\nu > 2$ (Higher-order): Captures many-body correlations like dihedrals.
Figure 2 illustrates how these 3-body correlations ($\nu=2$) appear in real space for an ethanol molecule, showing the density of triplets involving different chemical species.
Unification and Practical Insights
A key insight of this work is the mathematical proof that the widely used SOAP power spectrum is equivalent to a 3-body correlation of the atomic density expanded in a specific basis of spherical harmonics and radial functions. This unification shows that many seemingly different descriptors are actually specific instances of this general $n$-body correlation formalism.
Furthermore, the authors introduce a linear operator $\hat{U}$ that acts on the representation. This allows for "tuning" the descriptor for specific tasks. For instance, "alchemical kernels" can be used to project a large number of chemical elements into a lower-dimensional "chemical space." This is particularly useful for complex systems like high-entropy alloys, where the number of possible atomic combinations would otherwise lead to a feature explosion.
Conclusion
By providing a unified theoretical roof for atomic representations, this work simplifies the landscape of ML in chemistry. While challenges remain—such as the computational cost of higher-order correlations and the risk of overfitting when tuning the $\hat{U}$ operator—this framework points toward a more systematic design of descriptors. Instead of a "one-size-fits-all" approach, researchers can now tailor representations to the specific physical properties they aim to predict, leading to more efficient and accurate machine-learning models for molecular and materials discovery.
Summary updated