Machine Learning of Atomic-Scale Properties Based on Physical Principles

Michele Ceriotti, Michael J. Willatt, Gábor CśanyiOpenalex Unknown 2018 · DOI · arXiv · OpenAlex

Introduction

In the fields of materials science and chemistry, the ability to accurately predict atomic-scale properties is fundamental to discovering new materials and understanding chemical reactions. Historically, researchers have navigated a stark trade-off: computationally inexpensive empirical potentials, which often lack accuracy, or high-fidelity "first-principles" quantum mechanical calculations, which are computationally prohibitive for large systems. Machine learning (ML) has emerged as a promising bridge, yet standard "off-the-shelf" models frequently fail to respect the fundamental symmetries of physics—such as rotational and translational invariance. This paper presents a rigorous mathematical framework for regularizing ML models with physical principles, specifically through the Smooth Overlap of Atomic Positions (SOAP) representation and Kernel Ridge Regression (KRR).

The SOAP Representation and Physical Symmetries

A core challenge in modeling atomic systems is ensuring that the model's predictions remain invariant under physical transformations. For instance, the energy of a molecule should not change simply because the coordinate system is rotated. The authors address this by utilizing the SOAP representation, which describes the local environment of an atom by expanding the local atomic density in a basis of radial functions and spherical harmonics.

This density-based approach naturally embeds several critical symmetries:

  • Rotation and Translation Invariance: The representation is independent of the absolute position and orientation of the system.
  • Permutation Invariance: The model treats identical atoms (e.g., two hydrogen atoms in a water molecule) as indistinguishable.

By embedding these symmetries directly into the architecture, the model acts as a physically regularized learner, requiring significantly less data to achieve high generalization performance compared to non-symmetric models.

Mathematical Framework: KRR and Dirac Notation

The paper formalizes the relationship between atomic structures and properties using Kernel Ridge Regression. A notable innovation is the introduction of a "bra-ket" notation, analogous to quantum mechanics, to describe atomic configurations. This allows structural similarity to be treated as an overlap in a Hilbert space, providing a clean mathematical language for comparing complex atomic environments.

Furthermore, the framework extends beyond simple scalar properties. By treating forces (the negative gradient of energy) and virial stress as linear operators acting on the kernel, the authors demonstrate how models can be trained on derivative information. Since quantum mechanical calculations often provide forces at a marginal extra cost, training on these gradients provides $3N$ additional data points for a system of $N$ atoms, drastically improving the quality of the learned potential energy surfaces.

Covariant Properties and Alchemical Space

While many properties are invariant (scalars), others are covariant (tensors), such as polarizability or NMR shielding. The authors introduce $\lambda$-SOAP, an extension of the framework designed to predict tensorial properties that transform predictably under rotation.

Another key insight is the concept of "Alchemical Similarity." By treating different chemical elements as coordinates in a "species space," the model can learn relationships between elements. For example, it can recognize that Oxygen and Sulfur share similar chemical behaviors in certain environments, allowing the model to transfer knowledge across different parts of the periodic table. This is particularly useful for exploring the vast chemical space of alloys and organic molecules.

Efficiency through Sparsification

To make these models computationally viable for large-scale applications, the authors employ sparsification techniques. Using Farthest Point Sampling (FPS) and CUR matrix decomposition, they identify a minimal, representative set of atomic environments from the training data. This approach can reduce the number of required SOAP features by over 95% without sacrificing "chemical accuracy" (typically defined as errors below 1 kcal/mol), making the resulting potentials efficient enough for large-scale molecular dynamics simulations.

Conclusion and Future Directions

This work provides a unified mathematical umbrella for several prominent methods in the field, including Behler-Parrinello symmetry functions and Spectral Neighbor Analysis Potentials (SNAP). By leveraging the "nearsightedness" principle—the observation that most atomic properties are determined by their immediate local environment—the framework decomposes global properties into sums of local contributions.

However, challenges remain. The current SOAP framework is inherently local, defined by a cutoff radius. Capturing long-range effects, such as electrostatics or London dispersion forces, remains a challenge that requires augmenting these local kernels with global descriptors. Additionally, while the framework supports higher-order many-body interactions, the computational cost of calculating these descriptors increases significantly. Future work will likely focus on integrating these long-range interactions and optimizing the scaling for complex, non-additive systems like metallic alloys or excited electronic states.

Summary updated

BibTeX
@inproceedings{ceriotti2018machine,
  author = {Michele Ceriotti and Michael J. Willatt and Gábor Cśanyi},
  title = {{Machine Learning of Atomic-Scale Properties Based on Physical Principles}},
  booktitle = {Openalex Unknown},
  year = {2018},
  doi = {10.1007/978-3-319-42913-7\_68-1},
}

Loading related content...

Figure