Abstract: Evaluating the (dis)similarity of crystalline, disordered and molecular compounds is a critical step in the development of algorithms to navigate automatically the configuration space of complex materials. In the last few years several strategies have been designed to compare atomic coordination environments. In particular, the smooth overlap of atomic positions (SOAPs) has emerged as an elegant framework to obtain translation, rotation and permutation-invariant descriptors. Here we discuss how one can combine such local descriptors using a regularized entropy match (REMatch) approach to describe the similarity of both whole molecular and bulk periodic structures. Using this kernel and a ridge regression method we can predict atomization energies for a database of small organic molecules with a mean absolute error below 1 kcal mol(-1), reaching an important milestone in the application of machine-learning techniques for the evaluation of molecular properties.
Introduction
The search for new materials and drugs is often described as finding a needle in an astronomical haystack of possible atomic configurations. To navigate this vast space effectively, researchers require a robust mathematical "ruler" to quantify the similarity between different structures. Traditionally, methods have been bifurcated between molecular chemistry and materials science. This paper introduces a unified framework that bridges this gap, providing a universal language to describe everything from disordered glasses to flexible proteins.
Methodology: SOAP and REMatch
The proposed solution involves a two-step process. First, local atomic environments are characterized using Smooth Overlap of Atomic Positions (SOAP). SOAP descriptors represent the local neighbor density as a sum of Gaussians, ensuring the representation is invariant to rotation, translation, and the permutation of identical atoms.
Second, to compare entire structures, these local descriptors must be aggregated. The authors introduce the Regularized Entropy Match (REMatch) approach, which utilizes optimal transport theory (the Sinkhorn algorithm). Unlike simple averaging, which can be overly "blurry," or "best-match" algorithms, which are computationally expensive and numerically unstable, REMatch finds an optimal coupling with an entropy penalty. This results in a smooth, differentiable similarity metric that can be tuned to capture the desired level of structural detail.
The framework also incorporates "alchemical kernels," allowing the model to recognize that certain elements share chemical properties (e.g., based on electronegativity), rather than treating every element as a completely unique category.
Experimental Validation and Results
The framework was validated across four diverse datasets: $C_{60}$ clusters, silicon polymorphs, arginine dipeptide conformers, and the QM7b database of small organic molecules.
A standout result is the achievement of "chemical accuracy" in property prediction. Using Kernel Ridge Regression (KRR) with the REMatch-SOAP kernel, the authors predicted the atomization energies of the QM7b database with a Mean Absolute Error (MAE) of 0.04 eV (~0.92 kcal/mol). This represents a four-fold improvement over previous methods using Coulomb matrices or deep learning.
In the silicon dataset, the metric successfully clustered known phases and correctly placed hypothetical structures near their most similar known counterparts. Furthermore, the method proved sensitive enough to distinguish between dipeptide conformers differing only by a single proton transfer.
Conclusion
The REMatch-SOAP framework offers a powerful, unified approach to structural comparison across chemical and material space. By combining local geometric descriptors with a regularized global matching strategy, it enables both high-accuracy property prediction and intuitive visualization of complex chemical landscapes. Future work will focus on improving computational scaling for massive systems and automating hyperparameter selection.
Summary updated