Capturing long-range interaction with reciprocal space neural network
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "Capturing long-range interaction with reciprocal space neural network".
Paper summary: Kai: So, looking at the full paper "Capturing long-range interaction with reciprocal space neural network", the authors put forward this reciprocal space neural network to build potentials and descriptors that handle those long-range effects that are often neglected in current machine learning models <ref:2211.16684#pg1>.
Mira: They claim this method gives them a way to take structure information from real space, transform it into reciprocal space, and then encode it into a potential <ref:2211.16684#pg1>. This is important because long-range effects can dominate in ionic systems whose dynamics behavior is significantly influenced <ref:2211.16684#pg3>.
Lev: And the practical result they show is that Dimenet++ with this reciprocal descriptor achieves state-of-the-art accuracy on band gap prediction in the HOIP dataset Results Summary. That's a specific benchmark they used to prove its effectiveness.
Kai: It seems the core of what they accomplished is using this reciprocal space information to enhance both the crystal representation and the crystal atomic potential Conclusion. They successfully learn how to model those atomic long-range interactions beyond just the standard Coulomb interaction Conclusion.
Mira: So, for anyone listening who only cares about how this changes things, it means we can build better models for defect systems or any system where those long-range and full atomic interactions need to be taken into account Conclusion.
Lev: And from an engineering standpoint, it’s promising because they show the method improves stability in defect systems when you use that three-step training strategy Results Summary. That's a concrete improvement for simulation work.
Conclusion: Kai: So, to wrap up this piece, they're calling it "Capturing long-range interaction with reciprocal space neural network." It sounds like they’re trying to build a better model for crystals by looking at things in a different way.
Mira: Exactly. They take the atomic data from real space and project it into reciprocal space. This lets them see those long-range forces—the ones that aren't just simple point charges—more clearly than before.
Lev: From an error correction standpoint, if you have to calculate these long-range effects in a simulation, you usually have to sum over a lot of k vectors, which gets computationally expensive real fast.
Kai: Right. And this method seems to handle that by using the reciprocal space information to build the potential itself, not just adding a correction on top later. It’s built into the structure of how they learn it.
Mira: The authors claim this works for both local structure and long-range forces simultaneously, which is a big deal for describing complex materials with defects or in ionic systems where those long-range interactions matter most.
Lev: If the training strategy works as described—where you fit the real space first, then freeze it, then add the reciprocal space part—that sounds like it manages to keep the overall error manageable on a real piece of hardware.
Kai: The authors showed that this approach gets them state-of-the-art accuracy on band gap prediction in some datasets, which is a key property for materials. It suggests this representation isn't just theoretical; it’s actually useful for predicting what the material will do.
Mira: It means we can start building these neural network potentials that account for things like van der Waals forces or other long-range interactions, not just the basic Coulomb stuff.
Lev: Because if you can model those other interactions accurately, you might be able to make simulations more reliable for things like defect systems where those weaker forces become important.
Kai: So this work is promising for building models that really capture the full atomic picture of a crystal structure and its electronic properties. Next time, we’ll look at how these ideas apply specifically to those tricky defect systems.
Hongyu Yu, Liangliang Honga, Shiyou Chen, Xingao Gong, Hongjun Xianga
Key Laboratory of Computational Physical Sciences (Ministry of Education) · Shanghai Qi Zhi Institute · State Key Laboratory of ASIC and System
cond-mat.mtrl-sci, cs.LG, physics.chem-ph, physics.comp-ph
Submitted: 2022-11-30
Updated: 2022-11-30
Comments: 15 pages, 3 figures, 3 tables
Journal ref: Phys. Rev. B 2026
DOI: 10.1103/7yn2-22h6
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Machine Learning (ML) interatomic models and potentials have been widely employed in simulations of materials.
Key concepts
- Reciprocal Space Formulation
- This technique calculates the long-range part of the Ewald summation in reciprocal space using a generalized formalism involving a cutoff distance for k-vectors. It builds the long-range potential beyond explicit forms like Coulomb potential by using fully connected networks to process this information.
- Invariance under E(3) Symmetry
- The reciprocal space formulation ensures that the resulting potential remains invariant under all E(3) symmetry operations, including rotation, translation, and inversion. This invariance is maintained even when the choice of the simulation cell is changed, which is crucial for physically accurate descriptions.
- Three-Step Training Strategy
- The training process uses a three-stage approach: first fitting only the real space part; second freezing the real space network and adding long-range terms; and third jointly fitting both parts. This strategy improves performance and stability, particularly in defect systems where long-range interactions are important.
Terminology
Summary
Machine Learning (ML) interatomic models and potentials have been widely employed in simulations of materials. p. 0 1
Capturing long-range interaction with reciprocal space neural
network
Hongyu Yu,a,b,
*, Liangliang Honga,b,
*, Shiyou Chen,b,c Xingao Gong,a,b Hongjun
Xianga,b,†
aKey Laboratory of Computational Physical Sciences (Ministry of
Education), Institute of Computational Physical Sciences, and Department of Physics,
Fudan University, Shanghai 200433, China
b
Shanghai Qi Zhi Institute, Shanghai 200030, China
c
State Key Laboratory of ASIC and System, School of Microelectronics, Fudan
University, Shanghai 200433, China
Abstract:
Machine Learning (ML) interatomic models and potentials have been widely employed
in simulations of materials. Long-range interactions often dominate in some ionic
systems whose dynamics behavior is significantly influenced. However, the long-range
effect such as Coulomb and Van der Wales potential is not considered in most ML
interatomic potentials. To address this issue, we put forward a method that can take
long-range effects into account for most ML local interatomic models with the
reciprocal space neural network. The structure information in real space is firstly
transformed into reciprocal space and then encoded into a reciprocal space potential or
a global descriptor with full atomic interactions. The reciprocal space potential and
descriptor keep full invariance of Euclidean symmetry and choice of the cell. The reciprocal space potential and
descriptor keep full invariance of Euclidean symmetry and choice of the cell. Benefiting
from the reciprocal-space information, ML interatomic models can be extended to
describe the long-range potential including not only Coulomb but any other long-range
interaction.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Motivation and Limitations of Existing Models
Many types of ML potentials have gained much attention, such as high-dimensional neural network potentials (HDNNPs) [1], Gaussian approximation potentials (GAPs) [2], kernel ridge regression methods (KRR) [3], moment tensor potentials (MTPs) [4], spectral neighbor analysis potentials (SNAPs) [5] and message-passing neural networks (MPNN) [6]. The second-generation ML potentials have greatly overcome the shortcomings of the previous generation, but they tend to neglect all the long-range interactions, which limits their accuracy in some ionic systems. p. 2 Nearsightedness
means the total energy of a system can be expressed as a summation of atomic
contributions which depend on the local chemical environment inside a cutoff radius. As the price paid for nearsightedness, the second-generation ML potentials tend to
neglect all the long-range interactions, which limits their accuracy in some ionic
systems. As the price paid for nearsightedness, the second-generation ML potentials tend to
neglect all the long-range interactions, which limits their accuracy in some ionic
systems. However, the long-range interaction cannot be neglected when the potential decays
algebraically with a power smaller than the spatial dimension and influences the
dynamic behavior of the system [9,10]. There have been some ML potentials attempting to capture long-range
interaction [11–15]. Most of them are focused on the Coulomb interaction and neglect
other types of long-range interaction such as the Van der Waals interaction.Where the paper says it
Where the paper says it
Where the paper says it
Reciprocal Space Formulation for Long-Range Effects
The paper proposes capturing the long-range interaction in reciprocal space for systems with periodic boundary conditions (PBC). The long-range part of the Ewald summation is calculated in reciprocal space using the generalized formalism. p. 4 The short-range part of the Ewald
summation can be easily included by current models in the real space with a cutoff so
we are more interested in the long-range part of the Ewald summation. The long-range
part is calculated in reciprocal space with the generalized formalism 𝐸𝑙𝑟 =
1
Ω
∑ Φ(|𝑘⃗ |)|𝜌(𝑘⃗ , 𝑟 )|
2
𝑘 , 𝜌(𝑘⃗ , 𝑟 ) = exp(−𝑏
2
) , 𝑆(𝑘⃗ , 𝑟 ) = ∑ 𝑞𝑗 exp(−𝑖𝑘⃗
𝑗
∙ 𝑟 ) , 𝑏 =
|𝑘⃗ |
2𝛼
with a cutoff distance for integer k-vector in reciprocal space [41–43]. The short-range part of the Ewald
summation can be easily included by current models in the real space with a cutoff so
we are more interested in the long-range part of the Ewald summation. The long-range
part is calculated in reciprocal space with the generalized formalism 𝐸𝑙𝑟 =
1
Ω
∑ Φ(|𝑘⃗ |)|𝜌(𝑘⃗ , 𝑟 )|
2
𝑘 , 𝜌(𝑘⃗ , 𝑟 ) = exp(−𝑏
2
) , 𝑆(𝑘⃗ , 𝑟 ) = ∑ 𝑞𝑗 exp(−𝑖𝑘⃗
𝑗
∙ 𝑟 ) , 𝑏 =
|𝑘⃗ |
2𝛼
with a cutoff distance for integer k-vector in reciprocal space [41–43]. Based on the general formalism 𝐸𝑙𝑟
from the Ewald summation, we build our reciprocal space potential beyond an explicit
form such as Coulomb potential. With fully connected networks (FCN), it is constructed
as 𝐸𝑙𝑜𝑛𝑔 =
1
𝛺
∑ FCN(|𝑘⃗ |)|𝑆(𝑘⃗ , 𝑟 )|
2
𝑘 and the structural factor 𝑆(𝑘⃗ , 𝑟 ) =
∑ 𝑞𝑗 exp(−𝑖𝑘⃗
𝑗
∙ 𝑟 ) with the charge of the atom as 𝑞𝑗
. Based on the general formalism 𝐸𝑙𝑟
from the Ewald summation, we build our reciprocal space potential beyond an explicit
form such as Coulomb potential. With fully connected networks (FCN), it is constructed
as 𝐸𝑙𝑜𝑛𝑔 =
1
𝛺
∑ FCN(|𝑘⃗ |)|𝑆(𝑘⃗ , 𝑟 )|
2
𝑘 and the structural factor 𝑆(𝑘⃗ , 𝑟 ) =
∑ 𝑞𝑗 exp(−𝑖𝑘⃗
𝑗
∙ 𝑟 ) with the charge of the atom as 𝑞𝑗
.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Invariance and Training Strategy
The reciprocal space potential keeps invariant under operations of E(3) symmetry including rotation, translation, and inversion with different choices of the cell. For rotation operation, the length of the sampled k⃗ vectors keep invariant and (Rk⃗)∙ (Rr) = k⃗∙ r. p. 6 7
Similar to the usual local energy methods, our reciprocal space potential keeps
invariant under operations of E(3) symmetry including rotation, translation, and
inversion with different choices of the cell. As for rotation operation, the length of the
sampled 𝑘⃗ vectors keep invariant and (𝑅𝑘⃗ ) ∙ (𝑅𝑟 ) = 𝑘⃗ ∙ 𝑟 . During training, a three-step strategy is applied [45]. In the first stage, only the real
space part will be used to fit the potential of the datasets. In the first stage, only the real
space part will be used to fit the potential of the datasets. After the convergence is
achieved, it comes to the second stage in which the real space NN will be frozen and
the long-range part will be added to fit. The real space NN is not trained in this stage
but helps to fit the potential with the long-range part. Along with the second
convergence of the long-range NN, the third stage begins where the real space and the
reciprocal space NN are employed to fit the potential together.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Performance on Material Properties
The long-range descriptor is applied to the prediction of band gaps, an important intensive property, in the HOIP dataset. Dimenet++ with reciprocal descriptor achieves the best performance when compared to other configurations. p. 8 9
trained on the same training dataset and evaluated on the same testing dataset with
results shown in Table 1. Dimenet++ with reciprocal descriptor achieves the best
performance. Test results are shown in Table 2. The reciprocal space potential
achieves good accuracy in the experiments and shows superiority on the complex 6×6
×6 system with one point defect. The reciprocal space potential
achieves good accuracy in the experiments and shows superiority on the complex 6×6
×6 system with one point defect. Also, the 3-step strategy outperforms 1 step strategy
in all experiments, yielding the importance of this training strategy. The reciprocal space potential
achieves good accuracy in the experiments and shows superiority on the complex 6×6
×6 system with one point defect. Also, the 3-step strategy outperforms 1 step strategy
in all experiments, yielding the importance of this training strategy.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The reciprocal space descriptor with full atomic interaction and invariance helps to describe the crystal combined with local structure representation. p. 3 An ideal
reciprocal descriptor with full atomic information and invariance of E(3) symmetry and
choice of the cell is demanding for a better description of the crystal structure. In this work, we put forward the reciprocal space potential and the reciprocal space
descriptor that is invariant of E(3) symmetry and choice of the cell, which helps to
capture the long-range interatomic potential and the full atomic interaction information
for a better prediction of the global properties. Dimenet++ with reciprocal potential gives worse result than the pure
Dimenet++ as expected because it’s unreasonable to predict band gap with an addition
of the real space and reciprocal space. The ablation results of the pure Dimenet++ and
pure reciprocal descriptor shows that the reciprocal descriptor helps to encode the
crystal information and predict the band gap while the reciprocal descriptor alone is not
good enough for prediction. Dimenet++ with reciprocal potential gives worse result than the pure
Dimenet++ as expected because it’s unreasonable to predict band gap with an addition
of the real space and reciprocal space. The ablation results of the pure Dimenet++ and
pure reciprocal descriptor shows that the reciprocal descriptor helps to encode the
crystal information and predict the band gap while the reciprocal descriptor alone is not
good enough for prediction. Benefiting
from the reciprocal-space information, ML interatomic models can be extended to
describe the long-range potential including not only Coulomb but any other long-range
interaction. A model NaCl system considering Coulomb interaction and the GaxNy
system with defects are applied to illustrate the advantage of our approach. Benefiting
from the reciprocal-space information, ML interatomic models can be extended to
describe the long-range potential including not only Coulomb but any other long-range
interaction. A model NaCl system considering Coulomb interaction and the GaxNy
system with defects are applied to illustrate the advantage of our approach. It’s constructed by simulating the
scattering of an incident plane wave through the crystal as a 2D image with the
invariance of system size, atomic permutations, and the number of atoms but without
invariance of rotation. Also, Wang et al built the model with a reciprocal space
representation from elastic x-ray scattering theory without the information on the
chemical compositions. It’s constructed by simulating the
scattering of an incident plane wave through the crystal as a 2D image with the
invariance of system size, atomic permutations, and the number of atoms but without
invariance of rotation. Also, Wang et al built the model with a reciprocal space
representation from elastic x-ray scattering theory without the information on the
chemical compositions. To address this issue, we put forward a method that can take
long-range effects into account for most ML local interatomic models with the
reciprocal space neural network. The structure information in real space is firstly
transformed into reciprocal space and then encoded into a reciprocal space potential or
a global descriptor with full atomic interactions.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
[1] J. Behler and M. p. 1 So far, many types of ML potentials, such as high-dimensional neural network
potentials (HDNNPs) [1], Gaussian approximation potentials (GAPs) [2], kernel ridge
regression methods (KRR) [3], moment tensor potentials (MTPs) [4], spectral neighbor
analysis potentials (SNAPs) [5] and message-passing neural networks (MPNN) [6],
have gained much attention. Behler and his coworkers proposed a classification
scheme [7], which roughly classifies all the ML models into four generations based on
whether the short-range interaction, long-range interaction, and non-local effects can
be properly considered in the model. However, the long-range
effect such as Coulomb and Van der Wales potential is not considered in most ML
interatomic potentials. To address this issue, we put forward a method that can take
long-range effects into account for most ML local interatomic models with the
reciprocal space neural network.Where the paper says it
Where the paper says it
[2] A. P. Bartók, M. C. Payne, R. Kondor, and G. Csányi, Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons, Phys. p. 1 Therefore, more and more
researchers have tried to develop new machine-learning potential models and a series
of remarkable achievements have been made. So far, many types of ML potentials, such as high-dimensional neural network
potentials (HDNNPs) [1], Gaussian approximation potentials (GAPs) [2], kernel ridge
regression methods (KRR) [3], moment tensor potentials (MTPs) [4], spectral neighbor
analysis potentials (SNAPs) [5] and message-passing neural networks (MPNN) [6],
have gained much attention.Where the paper says it
[3] M. Rupp, A. Tkatchenko, K.-R. Müller, and O. A. von Lilienfeld, Fast and Accurate Modeling of Molecular Atomization Energies with Machine Learning, Phys. p. 0 1
Capturing long-range interaction with reciprocal space neural
network
Hongyu Yu,a,b,
*, Liangliang Honga,b,
*, Shiyou Chen,b,c Xingao Gong,a,b Hongjun
Xianga,b,†
aKey Laboratory of Computational Physical Sciences (Ministry of
Education), Institute of Computational Physical Sciences, and Department of Physics,
Fudan University, Shanghai 200433, China
b
Shanghai Qi Zhi Institute, Shanghai 200030, China
c
State Key Laboratory of ASIC and System, School of Microelectronics, Fudan
University, Shanghai 200433, China
Abstract:
Machine Learning (ML) interatomic models and potentials have been widely employed
in simulations of materials. Long-range interactions often dominate in some ionic
systems whose dynamics behavior is significantly influenced.Where the paper says it
[4] A. V. Shapeev, Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Potentials, Multiscale Model. p. 1 So far, many types of ML potentials, such as high-dimensional neural network
potentials (HDNNPs) [1], Gaussian approximation potentials (GAPs) [2], kernel ridge
regression methods (KRR) [3], moment tensor potentials (MTPs) [4], spectral neighbor
analysis potentials (SNAPs) [5] and message-passing neural networks (MPNN) [6],
have gained much attention. Behler and his coworkers proposed a classification
scheme [7], which roughly classifies all the ML models into four generations based on
whether the short-range interaction, long-range interaction, and non-local effects can
be properly considered in the model.Where the paper says it
[5] A. P. Thompson, L. P. Swiler, C. R. Trott, S. M. Foiles, and G. J. Tucker, Spectral Neighbor Analysis Method for Automated Generation of Quantum-Accurate Interatomic Potentials, J Comput Phys 285, 316 (2015)
[6] T. Xie and J. C. Grossman, Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties, Phys. p. 1 2
ability of current ML interatomic models and potentials when dealing with the longrange effect, hence paving a new way for accurate prediction of global properties and
large-scale dynamic simulations of systems with defects. Main text
Introduction.Where the paper says it
[7] J. Behler, Four Generations of High-Dimensional Neural Network Potentials, Chem. p. 1 So far, many types of ML potentials, such as high-dimensional neural network
potentials (HDNNPs) [1], Gaussian approximation potentials (GAPs) [2], kernel ridge
regression methods (KRR) [3], moment tensor potentials (MTPs) [4], spectral neighbor
analysis potentials (SNAPs) [5] and message-passing neural networks (MPNN) [6],
have gained much attention. Behler and his coworkers proposed a classification
scheme [7], which roughly classifies all the ML models into four generations based on
whether the short-range interaction, long-range interaction, and non-local effects can
be properly considered in the model.Where the paper says it
[8] T. B. Blank, S. D. Brown, A. W. Calhoun, and D. J. Doren, Neural Network Models of Potential Energy Surfaces, J Chem Phys 103, 4129 (1995)
[9] S. Gupta and D. Mukamel, Relaxation Dynamics of Stochastic Long-Range Interacting Systems, J Stat. p. 0 1
Capturing long-range interaction with reciprocal space neural
network
Hongyu Yu,a,b,
*, Liangliang Honga,b,
*, Shiyou Chen,b,c Xingao Gong,a,b Hongjun
Xianga,b,†
aKey Laboratory of Computational Physical Sciences (Ministry of
Education), Institute of Computational Physical Sciences, and Department of Physics,
Fudan University, Shanghai 200433, China
b
Shanghai Qi Zhi Institute, Shanghai 200030, China
c
State Key Laboratory of ASIC and System, School of Microelectronics, Fudan
University, Shanghai 200433, China
Abstract:
Machine Learning (ML) interatomic models and potentials have been widely employed
in simulations of materials. Long-range interactions often dominate in some ionic
systems whose dynamics behavior is significantly influenced. In general, there are two ways to construct the mapping
function, one taking the electronic structure of the system into account, while the other
only considers the atomic elements and positions as the independent variables. A typical
representative of the former is the density-functional theory (DFT), which tends to give
a high-precision result at the cost of computational efficiency.Where the paper says it
Where the paper says it
[10] S. Gupta and D. Mukamel, Slow Relaxation in Long-Range Interacting Systems with Stochastic Dynamics, Phys. p. 0 1
Capturing long-range interaction with reciprocal space neural
network
Hongyu Yu,a,b,
*, Liangliang Honga,b,
*, Shiyou Chen,b,c Xingao Gong,a,b Hongjun
Xianga,b,†
aKey Laboratory of Computational Physical Sciences (Ministry of
Education), Institute of Computational Physical Sciences, and Department of Physics,
Fudan University, Shanghai 200433, China
b
Shanghai Qi Zhi Institute, Shanghai 200030, China
c
State Key Laboratory of ASIC and System, School of Microelectronics, Fudan
University, Shanghai 200433, China
Abstract:
Machine Learning (ML) interatomic models and potentials have been widely employed
in simulations of materials. Long-range interactions often dominate in some ionic
systems whose dynamics behavior is significantly influenced.Where the paper says it
[11] N. Artrith, T. Morawietz, and J Behler, High-Dimensional Neural-Network Potentials for Multicomponent Systems: Applications to Zinc Oxide, Phys. p. 1 So far, many types of ML potentials, such as high-dimensional neural network
potentials (HDNNPs) [1], Gaussian approximation potentials (GAPs) [2], kernel ridge
regression methods (KRR) [3], moment tensor potentials (MTPs) [4], spectral neighbor
analysis potentials (SNAPs) [5] and message-passing neural networks (MPNN) [6],
have gained much attention. Behler and his coworkers proposed a classification
scheme [7], which roughly classifies all the ML models into four generations based on
whether the short-range interaction, long-range interaction, and non-local effects can
be properly considered in the model.Where the paper says it
[12] T. W. Ko, J A Finkler, S Goedecker, and J Behler, A Fourth-Generation High-Dimensional Neural Network Potential with Accurate Electrostatics Including Non-Local Charge Transfer, Nat Commun 12, 1 (2021)
[13] A. Gao and R. C. Remsing, Self-Consistent Determination of Long-Range Electrostatics in Neural Network Potentials, Nat Commun 13, 1 (2022)
[14] L Zhang, H Wang, M C Muniz, A Z Panagiotopoulos, R Car, and W E Deep Potential Model with Long-Range Electrostatic Interactions, J Chem Phys 156, 124107 (2022)
[15] E. R. Khajehpasha, J A Finkler, T D Kühne, and S A Ghasemi CENT2: Improved Charge Equilibration via Neural Network Technique, Phys. p. 2 To solve this issue, ML potentials of the third
generation have been developed in which the contribution of electrostatic interactions
is calculated analytically with the atomic partial charges predicted by the second set of
neural networks [11]. As a further improvement, the fourth-generation ML potentials
incorporate a global charge equilibration scheme into the previous framework, thus
being applicable to systems with non-local charge transfer or different charge
states [12].Where the paper says it
[16] C Ying, T Cai, S Luo, S Zheng, G Ke, D He, Y Shen, and T.-Y. Liu Do Transformers Really Perform Bad for Graph Representation? p. 1 Nowadays, molecular dynamics simulations and the Monte Carlo
method are widely used to study the properties of molecules or materials from a
microscopic perspective. To perform these numerical methods, a reliable mapping
function from the atomic configuration of the system to its energy (or forces on each
atom) is necessarily required.Where the paper says it
[17] F Wu, Q Zhang, D Radev, J Cui, W Zhang, H Xing, N Zhang, and H Chen 3DTransformer: Molecular Representation with Transformer in 3D Space, ArXiv211001191 Cs QBio (2021)
[18] Y Choukroun and L Wolf Geometric Transformer for End-to-End Molecule Properties Prediction, ArXiv21101372t ics (2021)
[19] S Batzner, A Musaelian, L Sun, M Geiger, J P Mailoa, M Kornbluth, N Molinari, T E Smidt, and B Kozinsky E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentials 13 Nat Commun 13 1 (2022)
[20] C Chen, W Ye, Y Zuo, C Zheng, and S P Ong Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals Chem. Mater. 31 3564 (2019)
[21] J Cheng, C Zhang, and L Dong A Geometric-Information-Enhanced Crystal Graph Network for Predicting Properties of Materials Commun. Mater. 2 1 (2021)
[25] S Li, Y Liu, D Chen, Y Jiang, Z Nie, and F Pan Encoding the Atomic Structure for Machine Learning in Materials Science WIREs Comput Mol Sci 12 2022
[30] R Drautz Atomic Cluster Expansion for Accurate and Transferable Interatomic Potentials Phys. Rev. B 99, 014104 (2019)
[31] L Himanen, M O J Jäger, E V Morooka, F Federici Canova, Y S Ranawat, D Z Gao, P Rinke, and A S Foster DScribe: Library of Descriptors for Machine Learning in Materials Science Comput. Phys Commun. 247 106949 (2020)
[35] M Pinheiro, F Ge, N Ferré, P O Dral, and M Barbatti Choosing the Right Molecular Machine Learning Potential Chem. Sci 12 14396 (2021)
[37] J Cheng, C Zhang, and L Dong A Geometric-Information-Enhanced Crystal Graph Network for Predicting Properties of Materials Commun. Mater. 2 1 (2021)
[38] J Klicpera, F Becker, and S Günnemann GemNet: Universal Directional Graph Neural Networks for Molecules ArXiv210608903 Phys Stat 2021
[40] K T Schütt, O T Unke, and M Gastegger Equivariant Message Passing for the Prediction of Tensorial Properties and Molecular Spectra ArXiv210203150 Phys. 2021
[47] C Kim, T D Huan, S Krishnan, and R Ramprasad A Hybrid Organic-Inorganic Perovskite Dataset Sci. Data 4 1 (2017)
[48] W Hu, B Liu, J Gomes, M Zitnik, P Liang, V Pande, and J Leskovec Strategies for PreTraining Graph Neural Networks Int Conf Learn Represent ICLR (2019)
[50] R Magar, Y Wang, and A Barati Farimani Crystal Twins: Self-Supervised Learning for Crystalline Material Property Prediction Npj Comput Mater 8 1 (2022)
[51] S P Ong, W D Richards, A Jain, G Hautier, M Kocher, S Cholia, D Gunter, V L Chevrier, K A Persson, and G Ceder Python Materials Genomics Pymatgen: A Robust Open-Source Python Library for Materials Analysis Comput. Mater Sci 68 314 (2013)
Results Summary
Dimenet++ with reciprocal descriptor achieves the best performance on the band gap prediction in the HOIP dataset. Dimenet++ with reciprocal potential shows superiority on the complex 6×6×6 system with one point defect. p. 9 Test results are shown in Table 2. The reciprocal space potential
achieves good accuracy in the experiments and shows superiority on the complex 6×6
×6 system with one point defect. Test results are shown in Table 2. The reciprocal space potential
achieves good accuracy in the experiments and shows superiority on the complex 6×6
×6 system with one point defect. The ablation results of the pure Dimenet++ and
pure reciprocal descriptor shows that the reciprocal descriptor helps to encode the
crystal information and predict the band gap while the reciprocal descriptor alone is not
good enough for prediction. As shown in results, band gap is predicted better with the
reciprocal descriptor. The best results are shown in boldface. The long-range interaction dominates in the defect system.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
In this article, we put forward the reciprocal space neural network for building long-range potential and descriptors. The reciprocal space descriptor with full atomic interaction and invariance helps to describe the crystal combined with local structure representation. Dimenet++ with the reciprocal space descriptor has achieved state-of-the-art accuracy on the band gap prediction. The reciprocal space potential is employed to describe the atomic long-range interactions in a neural-network style beyond the Coulomb interaction and successfully learn the model NaCl Coulomb interaction. With the reciprocal space information, we have enhanced the power of the crystal representation and the crystal atomic potential. It is promising in building potential models for defect systems and others in which long-range and full atomic interaction should be taken into account. p. 0 To address this issue, we put forward a method that can take
long-range effects into account for most ML local interatomic models with the
reciprocal space neural network. The structure information in real space is firstly
transformed into reciprocal space and then encoded into a reciprocal space potential or
a global descriptor with full atomic interactions.Where the paper says it
[50] R Magar, Y Wang, and A Barati Farimani Crystal Twins: Self-Supervised Learning for Crystalline Material Property Prediction Npj Comput Mater 8 1 (2022)
[51] S P Ong, W D Richards, A Jain, G Hautier, M Kocher, S Cholia, D Gunter, V L Chevrier, K A Persson, and G Ceder Python Materials Genomics Pymatgen: A Robust Open-Source Python Library for Materials Analysis Comput. Mater Sci 68 314 (2013)
Note
The work at Fudan is supported by NSFC (11825403, 11991067)
Reference List
[50] R Magar, Y Wang, and A Barati Farimani Crystal Twins: Self-Supervised Learning for Crystalline Material Property Prediction Npj Comput Mater 8 1 (2022)
[51] S P Ong, W D Richards, A Jain, G Hautier, M Kocher, S Cholia, D Gunter, V L Chevrier, K A Persson, and G Ceder Python Materials Genomics Pymatgen: A Robust Open-Source Python Library for Materials Analysis Comput.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Reciprocal space potential for long-range interactions allows machine learning interatomic models to describe Coulomb interaction and any other long-range interaction including Van der Waals interaction, which improves accuracy in systems where these effects are significant.
-
A learnable reciprocal space descriptor can be combined with local real space descriptors to predict global crystal properties, such as band gap, achieving state-of-the-art accuracy on the HOIP dataset when used with Dimenet++.
-
Dimenet++ with a reciprocal space potential and descriptor can be employed to describe atomic long-range interactions in a neural network style beyond the Coulomb interaction, successfully learning the model NaCl Coulomb interaction and helping to build a more accurate and stable potential for complex defect GaxNy systems.
Abstract
Machine Learning (ML) interatomic models and potentials have been widely employed in simulations of materials. Long-range interactions often dominate in some ionic systems whose dynamics behavior is significantly influenced. However, the long-range effect such as Coulomb and Van der Wales potential is not considered in most ML interatomic potentials. To address this issue, we put forward a method that can take long-range effects into account for most ML local interatomic models with the reciprocal space neural network. The structure information in real space is firstly transformed into reciprocal space and then encoded into a reciprocal space potential or a global descriptor with full atomic interactions. The reciprocal space potential and descriptor keep full invariance of Euclidean symmetry and choice of the cell. Benefiting from the reciprocal-space information, ML interatomic models can be extended to describe the long-range potential including not only Coulomb but any other long-range interaction. A model NaCl system considering Coulomb interaction and the GaxNy system with defects are applied to illustrate the advantage of our approach. At the same time, our approach helps to improve the prediction accuracy of some global properties such as the band gap where the full atomic interaction beyond local atomic environments plays a very important role. In summary, our work has expanded the ability of current ML interatomic models and potentials when dealing with the long-range effect, hence paving a new way for accurate prediction of global properties and large-scale dynamic simulations of systems with defects.
Sources
- MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields
- Examining graph neural networks for crystal structures: limitations and opportunities for capturing periodicity
Related papers
- Structural contribution to light-induced gap suppression in Ta 2 NiSe 5
- Anatomy of fast current-induced skyrmion motion in synthetic antiferromagnets
- Magnetic control of an exciton-polariton condensate in a van der Waals magnet
- High-pressure elastic properties of GeO2 polymorphs up to 120 GPa
- Decoupling Strain-Rate Sensitivity and Deformation Length Scale Effects in Neutron-Irradiated Tungsten: A Coupled Nano-Indentation, HR-EBSD and Crystal Plasticity Study
- Magnetism modulation in Co3Sn2S2 by current-assisted domain wall motion