Convergence of body-orders in linear atomic cluster expansions
Apolinario Miguel Tan, Franco Pellegrini, Stefano de Gironcoli · 2025
Read Convergence of body-orders in linear atomic cluster expansions by Apolinario Miguel Tan, Franco Pellegrini, Stefano de Gironcoli (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-16125.txt.
Introduction
Apolinario Miguel Tan, Franco Pellegrini, and Stefano de Gironcoli test how a materials model grows with detail in this 2025 paper. The atomic cluster expansion builds interatomic potentials from clusters of two, three, or more atoms. Whether adding higher body orders steadily improves the result was unclear.
They trained expansions truncated at body orders two through five on carbon data and on small silicon clusters. Standard versions failed to recover even the simple two atom curves, let alone stable ones. Removing self interaction terms did not fix the fit. Only with carefully tailored datasets and basis sets did the curves show signs of settling. The authors read this as a route toward potentials whose parts stay interpretable.
Catalog record
| CREATOR | Apolinario Miguel Tan, Franco Pellegrini, Stefano de Gironcoli |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 20 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-16125
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Convergence of body-orders in linear atomic cluster expansions Apolinario Miguel Tan, Franco Pellegrini, Stefano de Gironcoli Abstract We study the convergence of a linear atomic cluster expansion (ACE) potential Drautz 2019 with respect to its basis functions, in terms of the effective two-body interactions of elemental Carbon systems. We build ACE potentials with descriptor sets truncated at body-orders K = 2 , 3 , 4 K=2,3,4 trained on pure dimers, or on large datasets of different diversities but without any dimers. Potentials trained on a more diverse dataset fare better in validation and result in a nontrivial dimer curve, but still very far from the theoretical two-body interaction calculated by DFT. Moreover, dimer curves between descriptor sets clipped at different K K do not seem to converge to a universal function for a given dataset. We conclude that machine learning potentials employing atomic cluster expansions optimize losses at low K K but fail to generalize and converge properties described by two-body interactions. I Introduction Decompositions of the total energy of atomic systems by N -body contributions, or cluster expansions, \displaystyle E\left(\{\mathbf{r}_{i}\}\right) \displaystyle=V^{(0)}+\sum_{i}V^{(1)}\left(\mathbf{r}_{i}\right)+\sum_{\begin{subarray}{c}i,j\\ i\neq j\end{subarray}}V^{(2)}\left(\mathbf{r}_{i},\mathbf{r}_{j}\right)+ \displaystyle\sum_{\begin{subarray}{c}i,j,k\\ i\neq j\neq k\neq i\end{subarray}}V^{(3)}\left(\mathbf{r}_{i},\mathbf{r}_{j},\mathbf{r}_{k}\right)+... (1) have been used as the framework to study material properties such as the thermodynamics of substitutional alloys Sanchez et al. 1984 ; Ferreira et al. 1989 and empirical potentials of condensed matter and molecular systems Stillinger and Weber 1985 ; Biswas and Hamann 1987 ; Tersoff 1989 . While the infinite expansion has been shown to be complete, its practical effectiveness is limited to cases where the total energy converges after a few body-orders \nu , as the number of terms increases combinatorially as higher orders are included. More recently, the expansion has been used in describing local chemical environments for machine learning potentials. The underlying assumption for these descriptors is that the energy of the system is a sum of its individual atomic contributions E_{i} Behler and Parrinello 2007 , \displaystyle E\left(\{\mathbf{r}_{i}\}\right) \displaystyle=\sum_{i=1}^{N}E_{i}\left(\{\mathbf{r}_{i}\}\right) (2) where N indicates the total number of atoms in the structure. The atomic energies may then be expressed as an expansion in the number of neighbors. Developments on improving the construction of E_{i} have been made in recent decades, with symmetry-respecting descriptors that use only a few terms in the expansion Behler and Parrinello 2007 ; Bartók et al. 2013 ; Zhang et al. 2018 ; Lot et al. 2020 or take advantage of tensorial operations to reach an arbitrary body order Shapeev 2016 ; Drautz 2019 ; Batatia et al. 2025 ; Batatia et al. 2023 . A detailed review on descriptors of local chemical environments has been done by Musil et al . Musil et al. 2021 While the completeness Sanchez et al. 1984 ; Ferreira et al. 1989 ; Dusson et al. 2022 of such expansions allow an apt description of the system, its form in Eq. 1 are considered to converge slowly 1 1 1 Counterarguments Thomas et al. 2022 against the speed of convergence of body-ordered expansions claim that the surrounding environments are typically ignored and are usually crucial in the convergence at low body-orders. Stillinger and Weber 1985 ; Biswas and Hamann 1987 ; Drautz et al. 2004 , where computing a term of order \nu would scale in powers of \nu-1 of the number of neighbors. In order to avoid this prohibitive scaling, it is possible to express higher body-order terms as the product of simple two-body terms, at the price of including lower order terms, and losing the orthogonality of the expansion. A popular version of this resummation is the Atomic Cluster Expansion proposed by Drautz Drautz 2019 . In this approach, each E_{i} is expanded with a non-orthogonal basis, \{A_{\nu}\} , with unrestricted sums, \displaystyle E_{i}\left(\{\mathbf{r}_{i}\}\right) \displaystyle=\sum_{\nu}c^{(1)}_{\nu}A_{i,\nu}+\sum_{\nu_{1}\nu_{2}}^{\nu_{1}\geq\nu_{2}}c_{\nu_{1}\nu_{2}}^{(2)}A_{i,\nu_{1}}A_{i,\nu_{2}}+... (3) with single sums over neighbors \displaystyle A_{i,\nu}=\sum_{j}f_{\nu}\left(\mathbf{r}_{i},\mathbf{r}_{j}\right), (4) and \nu indexing the terms of the expansion. The expansion has been shown to be complete Dusson et al. 2022 , and implementations of the code have shown promise in machine learning applications Kovács et al. 2021 ; Lysogorskiy et al. 2021 ; Bochkarev et al. 2022 ; Qamar et al. 2023 . One of the state-of-the-art potentials, MACE Batatia et al. 2023 , extends from the formalism of ACE and represents the state of the art in molecular dynamics applications for materials with multiple elements Batatia et al. 2024 . Whereas potentials adopting ACE descriptors have shown to converge fast in losses, a detailed study on the convergence of the body-orders with respect to properties described by lower orders in the original expansion in Eq. 1 have not been thoroughly investigated. In this work, we investigate the convergence of linear ACE at different body-orders K in terms of their projection on the lowest terms of the original cluster expansion Eq. 1 . In particular, we consider the convergence of the potential in energy and forces on different datasets, as a function of the number and type of basis functions utilized, and compare it to the stability of a real body ordered expansion. Our study focuses on the first nontrivial term of the expansion: the two-body term. This can be easily extracted from a given potential by studying the energy profile of two isolated atoms, a dimer, as a function of distance: \displaystyle E_{2b,i}\left(\mathbf{r}_{i},\mathbf{r}_{j}\right) \displaystyle=V^{(2)}\left(\mathbf{r}_{i}-\mathbf{r}_{j}\right). (5) It is thus natural to ask whether this function might converge to the real dimer dissociation curve, or to another arbitrary function for a dataset that does not contain dimers. Neural network potentials (NNPs) may also employ nonlinearities via functional embeddings Finnis and Sinclair 1984 ; Tersoff 1988 ; Tersoff 1989 , or are inherent from their activation layers Behler and Parrinello 2007 ; Bartók et al. 2010 ; Zhang et al. 2018 ; Lot et al. 2020 ; Musil et al. 2021 . While nonlinearities improve their performance by being able to represent higher-order terms when the expansion is truncated, one has to be careful in how they are applied as certain applications may invalidate the body-ordering of the descriptor Batatia et al. 2025 . We do not aim to build the most accurate potentials, but to have a more grounded understanding of their behavior in the ACE framework. As such, we limit our investigations to training simpler linear models as well as not going to too high of a body-order where returns are diminished. Linear ACE models perform comparably well relative to neural network potentials, and have been successfully employed to describe molecular Kovács et al. 2021 and condensed phases Lysogorskiy et al. 2021 . II Methods II.1 Datasets The datasets we use in the investigations consists of elemental Carbon structures. Three datasets were used: the first set consists of 247 dimer structures at logarithmically-spaced distance between 0.5 and 8 Angstrom. Half of the structures were randomly selected for the validation set. The remaining two are sets of 1000 structures coming from a collection of Carbon allotropes generated from Shaidu et al. 2021 . One subset contains a homogeneous collection of diamond-like structures clustered using a Euclidean distance measure Oganov and Valle 2009 . The diamond-like dataset is highlighted in the histogram of distances to a reference diamond structure in Fig. 1 . The last subset are structures obtained using farthest point