Machine-Learned Force Fields for Lattice Dynamics at Coupled-Cluster Level Accuracy

Sita Schönbauer, Johanna P. Carbone, Andreas Grüneis · 2025

Read Machine-Learned Force Fields for Lattice Dynamics at Coupled-Cluster Level Accuracy by Sita Schönbauer, Johanna P. Carbone, Andreas Grüneis (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-2507-06929.txt.

Introduction

Sita Schönbauer and colleagues train machine-learned force fields at coupled-cluster accuracy for crystal vibrations. Their paper appeared in two thousand twenty-five. Coupled-cluster theory is a gold standard, but far too slow for long simulations. Density functional theory is faster and less exact. The test solids are diamond and lithium hydride. Phonon patterns and vibrational spectra judge the results.

Delta learning bridges the two levels of theory. The learned field targets the gap between coupled-cluster and functional results. A charge-aware variant handles long-range effects and missing force data. Optical vibration frequencies rise toward the experimental values. Anharmonic effects in lithium hydride are estimated at the higher level too. The workflow brings top-tier accuracy within simulation reach.

Catalog record

CREATORSita Schönbauer, Johanna P. Carbone, Andreas Grüneis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-06929

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Machine-Learned Force Fields for Lattice Dynamics at Coupled-Cluster Level Accuracy Sita Schönbauer, Johanna P. Carbone, Andreas Grüneis Abstract We investigate Machine-Learned Force Fields (MLFFs) trained on approximate Density Functional Theory (DFT) and Coupled Cluster (CC) level potential energy surfaces for the carbon diamond and lithium hydride solids. We assess the accuracy and precision of the MLFFs by calculating phonon dispersions and vibrational densities of states (VDOS) that are compared to experiment and reference ab initio results. To overcome limitations from long-range effects and the lack of atomic forces in the CC training data, a delta-learning approach based on the difference between CC and DFT results is explored. Compared to DFT, MLFFs trained on CC theory yield higher vibrational frequencies for optical modes, agreeing better with experiment. Furthermore, the MLFFs are used to estimate anharmonic effects on the VDOS of lithium hydride at the level of CC theory. keywords Machine Learning, Phonons, Coupled Cluster, Density Functional Theory 1 Introduction Machine learning (ML) methods have become an invaluable asset in determining precise atomic force fields without having to explicitly keep track of the electronic problem Gastegger et al. 2017 ; John and Csányi 2017 ; Chmiela et al. 2017 ; Chmiela et al. 2019 ; Zhang et al. 2018 ; Unke et al. 2021 . This allows for increases in system size and, in the case of Molecular Dynamics (MDs), simulation times at comparatively negligible computational cost. However, the accuracy of the ML methods is limited by the accuracy of the training data. So far, most ML studies in the field of materials science are limited to data generated using approximate exchange and correlation density functionals. More accurate and systematically improvable Wave Function Theory (WFT) based methods such as the “gold standard” in quantum chemistry, Coupled Cluster (CC) theory at the level of single, double and perturbative triple particle-hole excitations (CCSD(T)) are computationally significantly more expensive, which severely limits the size of accessible training data Bartlett and Musiał 2007 ; Raghavachari et al. 1989 . Consequently, these methods have mostly been used to generate training data for molecules Ruth et al. 2023 . Recently, several studies have expanded their scope in the context of ML to calculate molecular adsorption energies in a zeolite structure and properties of water Daru et al. 2022 ; Chen et al. 2023 ; Herzog et al. 2024 . Here, we investigate the extension of ML approaches to study the lattice dynamics of solids from training data generated by periodic CCSD(T) theory. Phonons play a fundamental role in determining various materials properties, including heat capacity and electron-phonon coupling, which is closely related to the emergence of exotic states of matter such as superconductivity Bardeen 1973 ; Kulić 2000 and quantum paraelectricity Kim and Chung 2024 ; Ranalli et al. 2023 . Therefore, efficient and accurate computational methods for simulating phonons are essential for designing quantum materials with tailored and enhanced properties. In this context, DFT persuades with its great cost-accuracy trade-off for many materials. However, currently available density functionals sometimes fail to achieve the required level of accuracy. A specific example are optical phonon frequencies in diamond, which are underestimated by commonly used approximate density functionals based on the generalized gradient approximation Maezono et al. 2007 . Yet, highly accurate knowledge of the relationship between phonon frequency and pressure is needed for diamond anvil cell experiments Maezono et al. 2007 ; Occelli et al. 2003 . Energies and atomic forces predicted by MLFFs are only precise for configurations that are “sufficiently similar” to those included in the training data. Furthermore, training data containing energies and atomic forces substantially improves the trade-off between precision of the MLFF and the required training data size. However, there exist ab initio methods for which the calculation of atomic forces is difficult, because it requires the implementation of many terms, which might introduce excessive usage of memory. Therefore, it is also desirable to investigate approaches that are based on training data excluding atomic forces. Currently, \Delta - and transfer-learning approaches provide such an avenue Bogojeski et al. 2020 ; Qu et al. 2022 ; Ruth et al. 2023 ; Smith et al. 2019 ; Käser et al. 0 ; Radova et al. 2025 . Here, we restrict our investigation to a single \Delta -learning technique for periodic CCSD(T) theory because a larger number of different ML approaches would be beyond the scope of this work. Moreover, the number of systems studied is mostly limited by the relatively large computational cost of CCSD(T) calculations. Many successful approaches for MLFFs are described in literature. In particular, neural networks and regression models have become very popular in materials science during the last decades Behler and Parrinello 2007 ; Bartók et al. 2010 . We employ an equivariant message passing neural network referred to as MACE Batatia et al. 2023 . The MACE implementation requires few input parameters, enabling a straightforward application to training data generated by periodic CC theory. However, the training data generated for the present study is made publicly available to serve as input data for other techniques. The main goals of the present work are to: (i) further assess the potential of \Delta -learning approaches for periodic CCSD(T) calculations, (ii) contribute to directing future developments in periodic CC theory needed to take full advantage of MLFFs, and (iii) establish a workflow that can be used in future applications of periodic CC calculations, where MLFFs are needed to bridge time- and system-size scales currently limited by the computational cost. 2 Results Throughout this work we employ the following notation. ML( X ) refers to predictions made by a MLFF trained on data generated using method X . All MLFF are based on MACE . In the case of X=\text{DFT}_{\text{E}} , the training data includes energies computed using an approximate exchange and correlation density functional only. X=\text{DFT}_{\text{E,F}} also includes atomic forces in the training data. \Delta ML( X ) refers to results generated using a simple linear combination of two MLFFs. The first MLFF corresponds to ML(DFT {}_{\text{E,F}} ), whereas the second MLFF is trained on the difference in energy between WFT method X and DFT. This includes Hartree-Fock (HF), Møller-Plesset perturbation theory (MP2), CC singles and doubles (CCSD) and CC singles, doubles, and perturbative triples (CCSD(T)). More details are explained in Sec. 4 . 2.1 Diamond Figure 1: Phonon dispersions and DOS comparisons for diamond; neutron scattering data (black dots) from Warren et al. 1967 , Raman spectroscopy (black x) from Occelli et al. 2003 and quantum Monte Carlo data (turquoise plus) from Maezono et al. 2007 . a) ML(DFT E,F ) (green area) vs. DFT results (blue line). b) ML(DFT E ) (trained only on E) (cyan area) vs. DFT results (blue line). c) \Delta ML(HF) (yellow area) vs. HF results (blue line). d) Different \Delta ML(WFT) results (gold, red dashed, brown dotted areas) vs. DFT results (blue line). e) Brillouin zone with relevant high symmetry points for diamond. We start by discussing the phonon dispersion of carbon diamond. First, we focus on results generated using a MLFF trained on 200 training data points of DFT-PBE ground-state energies and forces. Fig. 1 a) shows that ML(DFT E,F ) (green area) is able to

Read the complete text