Global Universal Scaling and Ultra-Small Parameterization in Machine Learning Interatomic Potentials with Super-Linearity

Caichao Ye; Jing Huang; Jiong Yang; Mingqiang Zhang; Wenqing Zhang; Xiaoxin Xu; Yabei Wu; Yanxiao Hu; Ye Sheng; Yuyan Yang

REVIEW 4 major objections 5 minor 52 references

Global Universal Scaling and Ultra-Small Parameterization in Machine Learning Interatomic Potentials with Super-Linearity

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An element-independent universal radial function can replace element-pair-specific fits in machine-learned interatomic potentials, collapsing parameter counts by two to three orders of magnitude while retaining comparable accuracy across…

desk verdict A compact MLIP that works better than its own derivation justifies; the universal-scaling step is an ansatz, not a theorem, but the empirical case is strong enough to referee. read the letter →

arxiv 2502.07293 v1 pith:C2B4M6XG submitted 2025-02-11 cond-mat.mtrl-sci cs.LG

Yanxiao Hu , Ye Sheng , Jing Huang , Xiaoxin Xu , Yuyan Yang , Mingqiang Zhang , Yabei Wu , Caichao Ye

show 2 more authors

Jiong Yang Wenqing Zhang

This is my paper · ORCID

classification cond-mat.mtrl-scics.LG

keywords machinelearninginteratomicpotentialsuniversalequationofstatescalingpotentialenergysurfaceultra-smallparameterizationphonontransportionicdiffusionsuper-linearexpressivity

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a machine-learning interatomic potential can keep full many-body expressive power while replacing the usual element-pair-specific radial functions with a single element-independent universal radial function. The replacement succeeds because each pair distance $r_{IJ}$ is first rescaled by two element-pair parameters into a scaled coordinate $r^*_{IJ}$, and the paper derives this scaling from the universal equation of state. The consequence, if the derivation holds, is that the parameter count stops growing with the square of the number of elements and shrinks by two to three orders of magnitude relative to large neural-network potentials, while energy, force, and stress accuracy remain comparable. The paper demonstrates this compact model on alloys, cathodes, perovskites, semiconductors, half-Heusler thermoelectrics, and superionic conductors, including predictions of phonons, lattice thermal conductivity, Cu-ion diffusion, and Li-ion migration.

What carries the argument

The load-bearing object is the universal radial function $\tilde R_l(r^*_{IJ})$, defined by $r^*_{IJ} = \alpha_{Z_I Z_J}(r_{IJ}/r_0^{Z_I Z_J} - 1)$, where $\alpha$ and $r_0$ are the two element-pair scaling parameters inherited from the universal equation of state. The cluster property field $P_I(\boldsymbol r_I;\{\boldsymbol r_j\}) = p_I \prod_{j\neq I}(\varphi_{Ij}+1)$ supplies the many-body expansion, with all $n$-body terms generated from the same pair function $\varphi_I = \sum_j \varphi_{Ij}$. Moment-tensor angular descriptors $r^{\otimes l}$ provide rotational features, while the nonlinear map $x = \tanh(r^*)$ sends the scaled distance into the Chebyshev domain $(-1,1)$ and gives the basis super-linear expressive capacity. Together these pieces decouple the element space from the coordinate space, so the learning parameters are a small fixed set of radial coefficients and linear expansion coefficients rather than element-pair-indexed lookup tables.

What would settle it

Train the model on one-element crystals only, read out the learned scaled radial functions $\tilde R_l(r^*)$ for different element pairs, and check whether they overlap on a common curve; if they do not, the universal-radial claim is refuted. A complementary test is to train only near equilibrium and then evaluate on strongly compressed or expanded configurations: if energy and force errors grow as fast as an unconstrained linear model, then the equation-of-state constraint is not doing the extrapolation work claimed.

Watch

Extended reading notes

Core claim

The central claim is that the element-dependent radial function $R_l(r_{IJ}, Z_I, Z_J)$ in the cluster expansion of the potential energy surface can be replaced by $R_l \to \tilde R_l(r^*_{IJ})$ with $r^*_{IJ} = \alpha_{Z_I Z_J}(r_{IJ}/r_0^{Z_I Z_J} - 1)$, so the radial basis is no longer indexed by element pairs. The justification is that the cluster property $P_I(r_{\mathrm{nn}})$ of an isotropic one-element crystal is an equation of state, so it must obey the universal equation of state $E^* = -(r^*+1)e^{-r^*}$, and the paper asserts this forces every pair interaction function $\varphi_{IJ}$ to live in the same scaled space. With this step, the radial expansion coefficient space shrinks from $\mathbb{R}^{n\times l\times Z^2}$ to $\mathbb{R}^{n\times l}$, which is the ultra-small parameterization. The paper then feeds the scaled coordinate through a $\tanh$ map into Chebyshev polynomials, a nonlinearity-embedded radial function claimed to give super-linear expressive capacity, and shows the resulting model matches larger neural-network potentials in accuracy while using far fewer parameters.

Load-bearing premise

The load-bearing premise is that the universal equation of state for the total cluster energy transfers to each individual pair interaction $\varphi_{IJ}$, so all element-pair radial functions collapse onto one universal curve after rescaling by $\alpha$ and $r_0$; a constraint on total energy versus volume does not by itself force every pair basis function to follow that same curve.

Editorial extensions

If this is right

Adding a new element to a multi-element model does not add new radial basis parameters; only the new pair-specific scaling constants $\alpha$ and $r_0$ need to be supplied.
A model trained mostly near equilibrium can extrapolate to strongly compressed or expanded states because the universal equation of state imposes the correct global volume dependence.
Phonon frequencies, third-order force constants, and lattice thermal conductivity can be obtained from the compact model's potential energy surface with errors comparable to DFT-level references.
Transfer learning works with very few structures: adding 45 configurations of a previously absent element pair corrected the phonon dispersion of a solid solution.
The light parameterization permits molecular dynamics on systems exceeding one million atoms on conventional CPU resources, where a larger neural-network baseline ran out of GPU memory below one hundred thousand atoms.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: if the universal radial claim is exact, then pair-dimer binding curves of diverse elements should collapse onto one master curve after the two-parameter rescaling; checking this directly would separate a physical law from a flexible fitting ansatz.
Editorial inference: the same scaling argument could be imported into graph-neural-network potentials as a physics prior that initializes or constrains their radial embeddings, reducing their parameter count without changing their update rules.
Editorial inference: a practical testable extension is to train the model on a handful of single-element crystals and then predict multi-element alloys with no additional training; success would demonstrate that the element-pair scaling parameters alone carry the chemical transferability.
Editorial inference: because the model is analytic and linear in its coefficients, its scaled radial functions can be read out directly, offering a way to compare learned interatomic interactions against measured pair potentials across the periodic table.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know: this is a serious empirical paper with a shaky theoretical center. The authors build a generalized linear MLIP (ACE/MTP-style) and claim the universal equation of state lets them replace element-pair radial functions with one universal radial function after per-pair scaling. That step, between Eq. (5) and Eq. (6), is not proven—UEOS constrains total energy vs volume, not each pair interaction. The stress-test note is right. But the paper still does a lot well.

First, the empirical work is substantial. They train on alloy, cathode, perovskite, semiconductor, half-Heusler, Cu2Se, and sulfide electrolyte datasets, and report accuracy comparable to DPA-2, MACE, Nequip, etc., with 2–3 orders of magnitude fewer parameters (17k–29k vs. millions). They also demonstrate physical scalability: phonon dispersions and lattice thermal conductivity for 130 half-Heusler compounds, MD-based thermal transport in beta-Cu2Se, and Li diffusion in LGPS-type electrolytes, with comparisons to DFT and experiment. The open-source implementation is a plus. These are genuine out-of-sample tests, not just fits.

The soft spots are real but proportionate. The central theoretical claim is overreaching: the UEOS does not force every pair function onto one universal curve. The per-pair scaling constants alpha and r0 are themselves fitted parameters, so part of the apparent collapse is an affine transformation absorbing chemical variation. That makes Eq. (6) a compact fitting ansatz rather than a physics-derived law. The benchmark on semiconductors also filters frames with max force < 5 eV/Å, which could bias the comparison toward smaller forces; that’s a minor concern given the other benchmarks, but worth flagging. The super-linearity argument leans on Hopfield/KAN analogies and is mostly hand-waving, though it’s not load-bearing for the empirical claims.

Who is this for? People working on universal or low-parameter MLIPs, and anyone who cares about whether physics-inspired constraints can reduce model size. The paper deserves a serious referee. I’d recommend major revision: reframe the universality claim as an inspired ansatz rather than a derivation, add a test that checks whether the universal radial function actually transfers to element pairs unseen in training, and be clearer about what the fitted alpha/r0 absorb. That would make the contribution both honest and stronger.

Referee Report

4 major / 5 minor

Summary. The paper proposes SUS2-MLIP, a low-parameter machine learning interatomic potential built around a generalized-linear cluster expansion. Its central theoretical claim is that the universal equation of state (UEOS) justifies replacing element-pair-specific radial functions with a single element-independent universal radial function R̃_l(r*_IJ) after scaling r*_IJ = α_{ZI,ZJ}(r_IJ/r0_{ZI,ZJ} − 1), reducing the radial parameter space from R^{n×l×Z^2} to R^{n×l}. The model is benchmarked against DPA-2, MACE, Nequip, Allegro, and other models on alloy, cathode, perovskite, and semiconductor datasets, and is applied to phonon/thermal conductivity calculations for half-Heuslers, Cu2Se molecular dynamics, and Li-ion diffusion in sulfide electrolytes. An open-source implementation is provided.

Significance. If the universal scaling of Eq. (6) were established, the parameter reduction and out-of-domain transferability would be a significant practical contribution: the reported parameter counts are genuinely small and the property predictions (phonons, thermal conductivity, ionic diffusivity) are broad and encouraging. The paper also ships code and performs genuine out-of-sample benchmarks. However, the central derivation is not currently established: the UEOS constrains total energy versus volume, not individual pair interaction functions, and the fitted per-pair scaling parameters α and r0 absorb part of the claimed universality. The practical model may still be a useful compact ansatz, but the physics-derived law claim needs substantial revision or supporting evidence.

major comments (4)

[II.2, Eq. (5)–Eq. (6)] The inference that UEOS implies every pair interaction function φ_IJ inherits the two-parameter universal scaling is not proved and is not a logical consequence of Eq. (5). For a one-element BCC/FCC crystal, Eq. (1) gives P_I(r_nn) = p_I ∏_j(1+φ_IJ(r_nn)), so the requirement that P_I satisfy the universal Rydberg form constrains only a specific scalar combination of pair functions, mainly the isotropic l=0 channel under cubic symmetry. It does not determine the angular decomposition, the l>0 radial channels, or the heteronuclear pair functions. Eq. (6) is therefore an ansatz, and the central parameter reduction from R^{n×l×Z^2} to R^{n×l} rests on this unproved universality.
[Eq. (6) and Fig. 4] The parameter-count comparison omits the fitted scaling parameters α_{ZI,ZJ,η} and r0_{ZI,ZJ,η} from the claimed reduction. These parameters are element-pair- and channel-dependent and grow as O(Z^2), so the statement that the model 'decouples the element space from coordinate space' is overstated. To support the universality claim, the authors should either report the total number of fitted parameters including α and r0 or provide a direct test that pair-specific fitted radial functions collapse onto a common R̃_l(r*) after scaling, rather than inferring the collapse from aggregate benchmark accuracy.
[III.1 and Methods: semiconductor benchmark] The semiconductor benchmark is not a controlled comparison. The caption of Fig. 3 and the Methods state that the training and validation sets 'were chosen based on a criterion of max atomic force < 5 eV/Å', while the results for the other MLIP models are taken from ref. 40. If the reference models were trained and evaluated on the full semiconductor dataset, the force RMSE comparison is not apples-to-apples, and the lower-force filtered test set may explain part of the reported accuracy. The baseline models should be retrained on the same filtered set, or the comparison should be reported both on the filtered and full datasets.
[II.3, Eq. (8)] The claim of 'super-linear expressive capacity' is not established. The paper cites the Hopfield-network memory-capacity results of Krotov and Demircigil et al., but those theorems concern the storage capacity of associative memory models for random patterns, not the approximation power of a tanh-composed Chebyshev radial basis for interatomic potential energy surfaces. The statement that 'tanh enables R̃ to possess a super-polynomial expansion form' does not by itself imply super-linear approximation capacity. The authors should either provide a concrete approximation-theoretic statement for the class of radial functions relevant to PES fitting or soften the super-linearity claim to a phenomenological one.

minor comments (5)

[Eq. (6) vs. Supplementary S2] The definition of the scaled coordinate is inconsistent between the main text and the supplementary information: Eq. (6) defines r*_IJ = α_{ZI,ZJ}(r_IJ/r0_{ZI,ZJ} − 1), while Eq. (S2) in Supplementary S2 defines r*_{IJ,η} = α_{ZI,ZJ,η}(r_IJ − r0_{ZI,ZJ,η}). The authors should use one definition throughout and clarify the units of r0.
[III.1, Methods] There are numerical inconsistencies in the reported parameter counts: the text lists 2,130 parameters for the cathode model and 6,092 for the perovskite model, whereas the Methods lists 2,066 and 6,028, respectively. These should be reconciled.
[Fig. 5 and general text] The labels 'DTF' in Fig. 5(b)–(c) and the associated text should read 'DFT', and the abstract contains typographical errors such as 'outcomes' (for 'overcomes' or 'avoids') and 'enbeding' (for 'embedding').
[II.1, Eq. (4)] The quantities τ_I and τ_j are introduced as element-specific parameters but their contribution to the total parameter count is not discussed. Since the paper is explicitly about ultra-small parameterization, the authors should state whether these parameters are fitted and include them in the parameter counts.
[III.1, Fig. 3] The benchmark figure reports single error values without repeated-seed statistics or error bars. Reporting the mean and standard deviation over multiple training runs would strengthen the comparison, especially given the very small model sizes involved.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: UEOS is an external physical prior, the benchmarks are independent held-out tests, and the contested Eq. (5)-to-Eq. (6) step is an unproved ansatz rather than a result equivalent to its input.

full rationale

The derivation chain is not circular. The UEOS relation is imported from external literature (Rose et al., Phys. Rev. B 1984; Banerjea & Smith, Phys. Rev. B 1988), not from the authors' own prior work, and it is used as a physical prior rather than as a fitted output. The contested inference in Section II.2 — that because the total cluster property P_I(r_nn) for an elemental FCC/BCC crystal must satisfy UEOS, 'all phi_IJ should also possess the universal representation within the scaling physical space as in equation (5)' — is an asserted extension or modeling assumption, not a circular reduction: a constraint on a total cohesive-energy curve does not, by itself, force each individual pair basis function or each many-body channel to collapse onto the same universal radial function. That is a legitimate logical-gap/correctness concern, but it is not a case of the prediction being equivalent to its input by construction. The model's accuracy claims are supported by genuine out-of-sample benchmarks against DFT, AIMD, and experimental data, and the phonon, thermal-conductivity, and diffusion results are not re-statements of fitted scaling parameters. One transparency issue is that the parameter-count reduction from R^{n*l*Z^2} to R^{n*l} in Eq. (6) omits the per-element-pair, per-channel fitted scaling factors alpha and r0; this weakens the 'ultra-small parameterization' claim as stated, but it is a reporting/counting concern, not a circular prediction. The only apparent self-citation (the HH130 dataset, Ref. 42, with a co-author) is a standard dataset citation and is not load-bearing for the universal-scaling argument. No circular step meeting the quoted-evidence standard is exhibited.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the universal EOS as an external physical law, on the mean-field factorization of the property field, and on the assertion that pair interactions inherit the scaling. The pair-specific scaling parameters and reference energies are fitted, so the model's parameter parsimony is real but not parameter-free.

free parameters (4)

alpha_{ZI ZJ, eta} = unknown; per-pair and per-channel
Element-pair and representation-channel scaling factor in r* = alpha(r - r0). The paper states scaling factors depend on element and crystal structure and provides no first-principles formula, so they appear to be fitted to data.
r0_{ZI ZJ, eta} = unknown; per-pair and per-channel
Equilibrium reference distance per element pair and channel. Used in the universal scaling transformation; same fitting caveat as alpha.
epsilon_{Z_I} and epsilon~_{Z_I} = unknown; per element
Cluster energy scale and reference energy in Eq. (3). The SI says reference energies are initialized with a least squares solution and then optimized.
tau_I and tau_j = unknown; element-specific
Element-specific node factors introduced in Eq. (4). If implemented as fitted element embeddings, they add to the parameter count.

assumptions (6)

domain assumption Universal equation of state applies to all atomic systems
The paper relies on Rose et al. (refs 21 and 31) to assert that all interatomic interactions collapse onto a universal curve after two-parameter scaling. This is cited external physics, not proven here.
domain assumption Mean-field factorization of cluster property: P_I = p_I * prod(phi_IJ + 1)
Eq. (1) and SI S1 assume indirect many-body interactions can be absorbed into the direct pair term, which is the mean-field approximation.
ad hoc to paper Each pair interaction inherits the universal scaling of the total-energy EOS
The leap from P_I(r_nn) satisfying UEOS to every phi_IJ satisfying Eq. (6) is asserted without proof. This is the weakest premise in the paper.
ad hoc to paper tanh followed by Chebyshev expansion yields super-linear expressive capacity
The Hopfield/KAN analogy is used to claim super-linear capacity, but no rigorous bound is given for this specific architecture; tanh is bounded, not a super-polynomial growing function.
standard math E(3) invariance and permutation invariance
The energy is required to be invariant under rotations, translations, and permutations; handled through moment-tensor contractions and full-index summation.
standard math Chebyshev polynomials form a complete basis for continuous functions on [-1,1]
Used in Eq. (S2) to expand the universal radial function in the transformed variable x in (-1,1).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Global Universal Scaling and Ultra-Small Parameterization in Machine Learning Interatomic Potentials with Super-Linearity." pith.science (2026). https://pith.science/paper/C2B4M6XG

@misc{pith2026250207293,
  author       = {Pith},
  title        = {Pith review of: Global Universal Scaling and Ultra-Small Parameterization in Machine Learning Interatomic Potentials with Super-Linearity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2B4M6XG}},
  note         = {Machine review of arXiv:2502.07293}
}

read the original abstract

Using machine learning (ML) to construct interatomic interactions and thus potential energy surface (PES) has become a common strategy for materials design and simulations. However, those current models of machine learning interatomic potential (MLIP) provide no relevant physical constrains, and thus may owe intrinsic out-of-domain difficulty which underlies the challenges of model generalizability and physical scalability. Here, by incorporating physics-informed Universal-Scaling law and nonlinearity-embedded interaction function, we develop a Super-linear MLIP with both Ultra-Small parameterization and greatly expanded expressive capability, named SUS2-MLIP. Due to the global scaling rooting in universal equation of state (UEOS), SUS2-MLIP not only has significantly-reduced parameters by decoupling the element space from coordinate space, but also naturally outcomes the out-of-domain difficulty and endows the potentials with inherent generalizability and scalability even with relatively small training dataset. The nonlinearity-enbeding transformation for interaction function expands the expressive capability and make the potentials super-linear. The SUS2-MLIP outperforms the state-of-the-art MLIP models with its exceptional computational efficiency especially for multiple-element materials and physical scalability in property prediction. This work not only presents a highly-efficient universal MLIP model but also sheds light on incorporating physical constraints into artificial-intelligence-aided materials simulation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 41 canonical work pages

[1]

Four Generations of High -Dimensional Neural Network Potentials

Behler, J. Four Generations of High -Dimensional Neural Network Potentials. Chem. Rev. 121, 10037– 10072 (2021)

work page 2021
[2]

Perspective: Machine learning potentials for atomistic simulations

Behler, J. Perspective: Machine learning potentials for atomistic simulations. J. Chem. Phys. 145, 170901 (2016)

work page 2016
[3]

Machine-learning interatomic potentials for materials science

Mishin, Y . Machine-learning interatomic potentials for materials science. Acta Mate r. 214, 116980 (2021)

work page 2021
[4]

L., Caro, M

Deringer, V . L., Caro, M. A. & Csányi, G. Machine Learning Interatomic Potentials as Emerging Tools for Materials Science. Adv. Mater. 31, 1902765 (2019)

work page 2019
[5]

Kocer, E., Ko, T. W. & Behler, J. Neural Network Potentials: A Concise Overview of Methods. Annu. Rev. Phys. Chem. 73, 163–186 (2022)

work page 2022
[6]

P., Batatia, I., Arany, E

Kovács, D. P., Batatia, I., Arany, E. S. & Csányi, G. Evaluation of the MACE force field architecture: From medicinal chemistry to materials science. J. Chem. Phys. 159, 044118 (2023)

work page 2023
[7]

Musaelian, A. et al. Learning local equivariant representations for large-scale atomistic dynamics. Nat. Commun. 14, 579 (2023)

work page 2023
[8]

Deng, B. et al. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nat. Mach. Intell. 5, 1031–1041 (2023)

work page 2023

Show all 52 references

[9]

Kolluru, A. et al. Open Challenges in Developing Generalizable Large-Scale Machine-Learning Models for Catalyst Discovery. ACS Catal. 12, 8572–8581 (2022)

2022
[10]

Lan, J. et al. AdsorbML: a leap in efficiency for adsorption energy calculations using generalizable machine learning potentials. npj Comput. Mater. 9, 172 (2023). 27

2023
[11]

Hu, T. et al. Impact of Amorphous LiF Coating Layers on Cathode‐Electrolyte Interfaces in Solid‐State Batteries. Adv. Funct. Mater. 34, 2402993 (2024)

2024
[12]

Lee, T. et al. Atomic-scale origin of the low grain -boundary resistance in perovskite solid electrolyte Li0.375Sr0.4375Ta0.75Zr0.25O3. Nat. Commun. 14, 1940 (2023)

2023
[13]

Hu, T. et al. Impact of the Local Environment on Li Ion Transport in Inorganic Components of Solid Electrolyte Interphases. J. Am. Chem. Soc. 145, 1327–1333 (2023)

2023
[14]

Wang, Y ., Lv, J., Gao, P. & Ma, Y . Crystal Structure Prediction via Efficient Sampling of the Potential Energy Surface. Acc. Chem. Res. 55, 2068–2076 (2022)

2022
[15]

Wang, J. et al. MAGUS: machine learning and graph theory assisted universal structure searcher. Nal. Sci. Rev. 10, nwad128 (2023)

2023
[16]

Kruglov, I. A. et al. Crystal structure prediction at finite temperatures. npj Comput. Mater. 9, 197 (2023)

2023
[17]

Fu, X. et al. Forces are not Enough: Benchmark and Critical Evaluation for Machine Learning Force Fields with Molecular Simulations. Preprint at http://arxiv.org/abs/2210.07237 (2023)

2023 arXiv
[18]

D., Gardner, J

Morrow, J. D., Gardner, J. L. A. & Deringer, V . L. How to validate machine -learned interatomic potentials. J. Chem. Phys. 158, 121501 (2023)

2023
[19]

Statistical Modeling: The Two Cultures

Breiman, L. Statistical Modeling: The Two Cultures. Stat. Sci. 16, 199–231 (2001)

2001
[20]

Kutz, J. N. & Brunton, S. L. Parsimony as the ultimate regularizer for physics -informed machine learning. Nonlinear Dynam. 107, 1801–1817 (2022)

2022
[21]

H., Smith, J

Rose, J. H., Smith, J. R., Guinea, F. & Ferrante, J. Universal features of the equation of state of metals. Phys. Rev. B 26, 2963–2969 (1984)

1984
[22]

Atomic cluster expansion for accurate and transferable interatomic potentials

Drautz, R. Atomic cluster expansion for accurate and transferable interatomic potentials. Phys. Rev. B 28 99, 014104 (2019)

2019
[23]

Atomic cluster expansion of scalar, vectorial, and tensorial properties including magnetism and charge transfer

Drautz, R. Atomic cluster expansion of scalar, vectorial, and tensorial properties including magnetism and charge transfer. Phys. Rev. B 102, 024104 (2020)

2020
[24]

Shapeev, A. V . Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Potentials. Multiscale Model. Simul. 14, 1153–1173 (2016)

2016
[25]

& Ong, S

Chen, C. & Ong, S. P. A universal graph deep learning interatomic potential for the periodic table. Nat. Comput. Sci. 2, 718–728 (2022)

2022
[26]

Takamoto, S. et al. Towards universal neural network potential for material discovery applicable to arbitrary combination of 45 elements. Nat. Commun. 13, 2991 (2022)

2022
[27]

W., Wood, B

Qi, J., Ko, T. W., Wood, B. C., Pham, T. A. & Ong, S. P. Robust training of machine learning interatomic potentials with dimensionality reduction and stratified sampling. npj Comput. Mater. 10, 43 (2024)

2024
[28]

Choudhary, K. et al. Unified graph neural network force -field for the periodic table: solid state applications. Digital Discovery 2, 346–355 (2023)

2023
[29]

Freitas, L

Focassio, B., M. Freitas, L. P. & Schleder, G. R. Performance Assessment of Universal Machine Learning Interatomic Potentials: Challenges and Directions for Materials’ Surfaces. ACS Appl. Mater. Interfaces acsami.4c03815 (2024) doi:10.1021/acsami.4c03815

2024 doi
[30]

Batatia, I. et al. A foundation model for atomistic materials chemistry. Preprint at https://doi.org/10.48550/arXiv.2401.00096 (2024)

2024 doi arXiv
[31]

& Smith, J

Banerjea, A. & Smith, J. R. Origins of the universal binding -energy relation. Phys. Rev. B 37, 6632– 6645 (1988)

1988
[32]

S., Gubaev, K., Podryabinkin, E

Novikov, I. S., Gubaev, K., Podryabinkin, E. V . & Shapeev, A. V . The MLIP package: moment tensor 29 potentials with MPI and active learning. Mach. Learn.: Sci. Technol. 2, 025002 (2021)

2021
[33]

A new frontier for Hopfield networks

Krotov, D. A new frontier for Hopfield networks. Nat. Rev. Phys. 5, 366–367 (2023)

2023
[34]

J., Gutfreund, H

Amit, D. J., Gutfreund, H. & Sompolinsky, H. Storing Infinite Numbers of Patterns in a Spin -Glass Model of Neural Networks. Phys. Rev. Lett. 55, 2554–2558 (1985)

1985
[35]

Liu, Z. et al. KAN: Kolmogorov-Arnold Networks. Preprint at http://arxiv.org/abs/2404.19756 (2024)

2024 arXiv
[36]

& Vermet, F

Demircigil, M., Heusel, J., Löwe, M., Upgang, S. & Vermet, F. On a Model of Associative Memory with Huge Storage Capacity. J. Stat. Phys. 168, 288–299 (2017)

2017
[37]

Gasteiger, J. et al. GemNet-OC: Developing Graph Neural Networks for Large and Diverse Molecular Simulation Datasets. Preprint at http://arxiv.org/abs/2204.02782 (2022)

2022 arXiv
[38]

& Smidt, T

Liao, Y .-L., Wood, B., Das, A. & Smidt, T. EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations. Preprint at http://arxiv.org/abs/2306.12059 (2024)

2024 arXiv
[39]

Batzner, S. et al. E(3)-equivariant graph neural networks for data -efficient and accurate interatomic potentials. Nat. Commun. 13, 2453 (2022)

2022
[40]

Zhang, D. et al. DPA-2: Towards a universal large atomic model for molecular and material simulation. Preprint at http://arxiv.org/abs/2312.15492 (2023)

2023 arXiv
[41]

Wang, H., Zhang, L., Han, J. & E, W. DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics. Comput. Phys. Commun. 228, 178–184 (2018)

2018
[42]

Yang, Y . et al. HH130: a standardized database of machine learning interatomic potentials, datasets, and its applications in the thermal transport of half -Heusler thermoelectrics. Digital Discovery 3, 2201 (2024)

2024
[43]

Dai, S. et al. Revealing the decisive factors of the lattice thermal conductivity reduction by electron - 30 phonon interactions in half-Heusler semiconductors. Mater. Today Phys. 31, 100993 (2023)

2023
[44]

Copper ion liquid-like thermoelectrics

Liu, H. Copper ion liquid-like thermoelectrics. Nat. Mater. 11, 422–425 (2012)

2012
[45]

& Chou, M.-Y

Zhuo, K., Wang, J., Gao, J., Landman, U. & Chou, M.-Y . Liquidlike Cu atom diffusion in weakly ionic compounds Cu2S and Cu2Se. Phys. Rev. B 102, 064201 (2020)

2020
[46]

Zhu, Y . et al. Atomic potential energy uncertainty in machine-learning interatomic potentials and thermal transport in solids with atomic diffusion. Phys. Rev. B 108, 014108 (2023)

2023
[47]

Liang, F. et al. Designing inorganic electrolytes for solid-state Li-ion batteries: A perspective of LGPS and garnet. Mater. Today 50, 418–441 (2021)

2021
[48]

Wang, R. et al. A Pre-trained Deep Potential Model for Sulfide Solid Electrolytes with Broad Coverage and High Accuracy. Preprint at http://arxiv.org/abs/2406.18263 (2024). 31 Acknowledgements: This project is supported by the National Natural Science Foundation of China (9216...

2024 arXiv
[49]

S., Gubaev, K., Podryabinkin, E

Novikov, I. S., Gubaev, K., Podryabinkin, E. V . & Shapeev, A. V . The MLIP package: moment tensor potentials with MPI and active learning. Mach. Learn.: Sci. Technol. 2, 025002 (2021)

2021
[50]

First-principles Phonon Calculations with Phonopy and Phono3py

Togo, A. First-principles Phonon Calculations with Phonopy and Phono3py. J. Phys. Soc. Jpn. 92, 012001 (2023)

2023
[51]

& Tanaka, I

Togo, A., Chaput, L., Tadano, T. & Tanaka, I. Implementation strategies in phonopy and phono3py. J. Phys.: Condens. Matter 35, 353001 (2023)

2023
[52]

Thompson, A. P. et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Comput. Phys. Commun. 271, 108171 (2022)

2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Four Generations of High -Dimensional Neural Network Potentials

Behler, J. Four Generations of High -Dimensional Neural Network Potentials. Chem. Rev. 121, 10037– 10072 (2021)

work page 2021

[2] [2]

Perspective: Machine learning potentials for atomistic simulations

Behler, J. Perspective: Machine learning potentials for atomistic simulations. J. Chem. Phys. 145, 170901 (2016)

work page 2016

[3] [3]

Machine-learning interatomic potentials for materials science

Mishin, Y . Machine-learning interatomic potentials for materials science. Acta Mate r. 214, 116980 (2021)

work page 2021

[4] [4]

L., Caro, M

Deringer, V . L., Caro, M. A. & Csányi, G. Machine Learning Interatomic Potentials as Emerging Tools for Materials Science. Adv. Mater. 31, 1902765 (2019)

work page 2019

[5] [5]

Kocer, E., Ko, T. W. & Behler, J. Neural Network Potentials: A Concise Overview of Methods. Annu. Rev. Phys. Chem. 73, 163–186 (2022)

work page 2022

[6] [6]

P., Batatia, I., Arany, E

Kovács, D. P., Batatia, I., Arany, E. S. & Csányi, G. Evaluation of the MACE force field architecture: From medicinal chemistry to materials science. J. Chem. Phys. 159, 044118 (2023)

work page 2023

[7] [7]

Musaelian, A. et al. Learning local equivariant representations for large-scale atomistic dynamics. Nat. Commun. 14, 579 (2023)

work page 2023

[8] [8]

Deng, B. et al. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nat. Mach. Intell. 5, 1031–1041 (2023)

work page 2023