Pith. sign in

REVIEW 4 major objections 8 minor 4 references

Developing a Neural Network Machine Learning Interatomic Potential for Molecular Dynamics Simulations of La-Si-P Systems

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a neural-network machine-learning interatomic potential can be trained to accurately describe crystalline and liquid La-Si-P structures and to predict melting temperatures and LaP nucleation in this ternary system.

desk verdict A genuinely useful first MLIP for La–Si–P, but the melting-temperature validation is less independent than claimed because transition configurations were added to the training set. read the letter →

arxiv 2506.08339 v1 pith:RLOOSTS4 submitted 2025-06-10 cond-mat.mtrl-sci cond-mat.dis-nnphysics.chem-ph

Ling Tang , Weiyi Xia , Gayatri Viswanathan , Ernesto Soto , Kirill Kovnir , Cai-Zhuang Wang This is my paper
classification cond-mat.mtrl-scicond-mat.dis-nnphysics.chem-ph
keywords machinelearninginteratomicpotentialneuralnetworkmoleculardynamicsLa-Si-Pternarysystemmeltingtemperaturenucleationandgrowthiterativetrainingphasestability
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

The paper claims that a neural-network machine-learning interatomic potential can be built for the ternary La-Si-P system with enough accuracy and transferability for molecular dynamics simulations of phase stability and transformations. Training data come from distorted crystal structures, liquid snapshots, and crystal-liquid transition configurations, and the model is refined iteratively by up-weighting poorly learned phases and adding new melting data. The resulting potential reproduces energy-volume curves for known elemental, binary, and ternary phases and liquid pair-correlation functions when checked against density-functional theory. It predicts the melting temperatures of two ternary compounds within about 5 to 10 percent of experiment, and molecular dynamics runs show LaP nucleating and growing from La-Si-P liquids as observed in synthesis. If correct, this makes molecular dynamics a practical tool for guiding the synthesis of new ternary La-Si-P compounds.

What carries the argument

The central machinery is a local-environment neural network potential whose total energy is the sum of per-atom energies, so any structure can be evaluated quickly once the network is trained. The argument rides on the iterative training loop: the model is first fit to a broad set of distorted crystals and ab initio liquid snapshots, then re-trained with adaptively increased weights for phases it fits poorly, and finally supplemented with snapshots collected at the solid-liquid transition so the potential learns barriers and molten states rather than only equilibrium structures. That last step is what rescues the melting-temperature predictions, as shown by the La2SiP4 case where adding transition configurations raises the predicted melting point substantially.

What would settle it

Compare the training energies and forces against spin-polarized DFT calculations with an explicit treatment of lanthanum's 4f electrons for a representative subset of La-containing crystals and liquids; if the reference data shift enough to reorder competing phases or move predicted melting temperatures by more than a few percent, the transferability claim inherits a systematic error. A second check is to measure the melting temperature of LaP at controlled phosphorus pressure, since the potential predicts roughly 2690 K while no direct measurement exists above 1373 K.

Watch

Extended reading notes

Core claim

The central claim is that an accurate and transferable artificial-neural-network machine-learning interatomic potential can be developed for the La-Si-P system. The potential represents the total energy as a sum of atomic energies, each obtained by passing a local environment descriptor, namely neighbor positions within a 7.0 Å cutoff, through a filter network and then through four hidden layers of 120 nodes; forces follow as derivatives of the total energy. The training set contains about 71,800 distorted crystal structures and 130,000 liquid snapshots, plus configurations added later from solid-liquid transition simulations. On the final model, the energy RMS error is about 12 meV per atom and the force error about 0.23 eV per angstrom. The model reproduces the energy-volume curves of all known La-Si-P crystalline phases and the pair-correlation functions of La-Si-P liquids at 2500 K against ab initio results. It yields melting temperatures of the Aea2 polymorph of LaSiP3 and of La2SiP4 that are 5.2 percent and 9.8 percent below the measured values, and it captures LaP nucleation and growth from undercooled liquids at 1400 K.

Load-bearing premise

The load-bearing premise is that the non-spin-polarized DFT calculations used to label the training data are accurate enough for lanthanum-containing phases; if those reference energies and forces are biased, the neural network faithfully learns the bias.

Editorial extensions

If this is right

  • Molecular dynamics studies of the La-Si-P system can now be run at sizes and time scales that ab initio MD cannot reach, allowing phase competition and transformation kinetics to be studied directly.
  • The predicted melting temperatures of the ternary compounds are systematically low by 5 to 10 percent, so the potential is reliable for trends and for guiding synthesis conditions rather than for exact thermodynamic values.
  • The simulations reproduce LaP and Si-substituted LaP crystallization from undercooled La-Si-P liquids, consistent with the experimental observation that LaP nucleates readily during synthesis.
  • The iterative training recipe, adaptive re-weighting of existing data plus addition of crystal-liquid transition configurations, is what carries the transferability, and the same recipe can be applied to other ternary systems.
  • For binary and elemental phases, larger melting-temperature errors of up to about 20 percent indicate that those transitions need dedicated training data before the potential is used quantitatively there.

Reading between the lines

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

  • This suggests the same iterative data-generation strategy could be transferred to other rare-earth pnictide ternaries, where synthesis is often limited by unknown liquid-phase kinetics rather than by thermodynamic stability alone.
  • Because the reference data come from non-spin-polarized DFT for a lanthanum-containing system, the potential may carry a systematic bias from the treatment of 4f electrons; testing against spin-polarized reference calculations would reveal the size of that bias.
  • Given the systematic low bias in melting temperatures, relative stability rankings and crystallization trends from this potential are more trustworthy than absolute melting points, and a small set of experimental anchor melting temperatures could calibrate the model for quantitative use.
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.

Referee Report

4 major / 8 minor

Summary. The manuscript reports the development of a DeePMD-based neural network machine learning interatomic potential for the La-Si-P ternary system. The authors construct a large training set of DFT energies and forces for distorted crystals and liquid snapshots across many compositions, use iterative retraining with adjusted data weights and added transition-state structures, and validate the potential against DFT energy-volume (E-V) curves and liquid pair-correlation functions. They then predict melting temperatures of ternary LaSiP3 and La2SiP4 via solid/liquid coexistence molecular dynamics (MD), compare with experimental values (including a new differential scanning calorimetry measurement), and simulate LaP nucleation from La-Si-P liquids. The paper claims that the potential is accurate and transferable, with a 5-10% underestimation of ternary melting temperatures and qualitative agreement with observed crystallization behavior.

Significance. If the developed potential is genuinely accurate and transferable, it would provide a valuable computational tool for guiding synthesis of ternary La-Si-P phases and for studying phase competition and growth kinetics in this materials family. The paper has notable strengths: a systematic iterative training workflow, broad compositional coverage of the training set, a new DSC measurement of the La2SiP4 melting temperature, and successful reproduction of DFT E-V curves and liquid structures. However, the load-bearing evidence for transferability--the melting-temperature predictions--is significantly weakened by the paper's own description of the training procedure, which added solid-liquid transition configurations to the training set. The agreement for ternary compounds is therefore partly a fitting result, and the large errors reported for LaSi and elemental La (about 20% and 19%, respectively) further temper the transferability claim. The choice of non-spin-polarized PBE for lanthanum-containing phases also raises a correctness risk that is not addressed in the manuscript.

major comments (4)
  1. [Section 3, page 11; Section 2, page 6 (Fig. 2d); Section 4, page 15] The statement in Section 3 that 'the melting temperatures are not explicitly included in training' is internally inconsistent with the procedure described elsewhere. Section 2 states that in the 4th iteration, 'additional structures collected during the crystal to liquid transition' were added to the training data set, and Fig. 2(d) shows that this addition substantially increased the predicted melting temperature of La2SiP4. Section 4 repeats that adding structures near the melting temperature 'can improve the performance of the ANN-ML model significantly, especially for melting temperature prediction.' The predicted Tm values for La2SiP4 (and possibly LaSiP3, given the general statement) are therefore not independent predictions but tests of how well the model interpolates on transition-state configurations whose DFT energies and forces were explicitly provided during training. The authors should remove or substantially qualify the 'remarkable' claim and instead discuss the agreement as a validation that the potential describes the solid-liquid coexistence region, not as evidence that Tm was predicted from data that omitted melting-relevant configurations.
  2. [Section 3, Fig. 8] The potential underestimates the melting temperatures of LaSi and elemental La by about 20% and 19%, respectively, while the ternary compounds are within 5-10% of experiment. The paper attributes the systematic underestimation to the use of GGA-DFT reference data, but this explanation does not account for the large variation across phases. Given the demonstrated sensitivity of Tm to the addition of melting-path training data in Fig. 2(d), the difference between the ternary and elemental/binary results likely reflects a deficiency in transition-state training data for the latter phases rather than a purely systematic DFT offset. The claim of an 'accurate and transferable' potential should be tempered, or the authors should add melting-transition training data for LaSi and La to test whether the errors can be reduced, rather than attributing the full discrepancy to the DFT functional.
  3. [Section 2, page 5 (DFT settings)] The exchange-correlation functional is non-spin-polarized PBE, and no justification or test of this choice is provided for lanthanum, a rare-earth element with 4f electrons. If the DFT reference is inaccurate for La-containing phases due to unpaired 4f spins or strong correlations, the machine learning potential will inherit these errors, undermining the claimed transferability across the La-Si-P phase space. The authors should either verify that spin-polarized (or DFT+U) calculations give negligible energy differences for representative elemental, binary, and ternary La-containing structures, or explicitly discuss this limitation as a correctness risk in the computed melting temperatures and nucleation behavior.
  4. [Section 3, page 8 (E-V curves)] The E-V curves are presented as evidence of transferability, but Section 2 indicates that the training set already includes the same stable elemental, binary, and ternary phases taken from the Materials Project, as well as the previously predicted ternary compounds. The E-V agreement is therefore partly a check on the training set, not an independent test. To substantiate the transferability claim, the authors should identify which phases in Fig. 4 were not included in the training data, report their E-V errors separately, and make clear whether the iterative training procedure was adjusted based on these same E-V results.
minor comments (8)
  1. [Section 2, page 4] The description of training data generation states that '21 to 41 non-distorted structures' and '2100 to 4100 distorted structures' are generated, but it is unclear whether these numbers are per parent phase or in total; please clarify and verify that the stated total of 71,800 distorted crystal structures is consistent with the number of parent compounds.
  2. [Fig. 2(d) caption] The caption says 'adding relevant training data from MD simulation of liquid to solid transitions,' while the main text describes 'structures collected during the crystal to liquid transition'; please unify the terminology.
  3. [Section 3, page 10] The phase notation 'Aea2 LaSiP3' is used without explanation; please define the space group and, if needed, the structure type, so that readers unfamiliar with the polymorph can follow the discussion.
  4. [Section 3, page 11 (DSC of La2SiP4)] The DSC sample is described as a mixture of LaP and La2SiP4, and the endothermic peak at ~1330 K is only 'likely' assigned to the melting of La2SiP4; please provide additional evidence for this assignment, such as a comparison with pure La2SiP4 or a discussion of the LaP-La2SiP4 phase diagram and possible eutectic effects.
  5. [Section 3, page 10 (coexistence simulations)] The coexistence MD description does not specify the length of the NVE simulation over which the temperature (kinetic energy) is averaged; please provide the simulation time and any equilibration criteria used to determine when coexistence is reached.
  6. [Section 4 and reference [57]] The nucleation and growth results are stated to be in agreement with experimental observations from reference [57], which is listed as 'to be published'; this comparison cannot currently be verified, so please provide the supporting data or cite a published source.
  7. [Abstract and page 2] The name 'Behler and Perrinello' is a typo; it should read 'Behler and Parrinello.'
  8. [Fig. 3] The RMS errors reported in the caption are computed using 1320 randomly selected snapshots from the training data set; please clarify that these are training errors, not held-out test errors, and consider reporting errors on a separate validation set for a more meaningful accuracy assessment.

Circularity Check

2 steps flagged · score 6.0 of 10

Melting-temperature 'prediction' is not an independent test: the model was explicitly fine-tuned on crystal-to-liquid transition snapshots, while Section 3 calls the same agreement 'remarkable' because melting temperatures were 'not explicitly included.'

  1. fitted input called prediction [Section 4 (Summary), Fig. 2(d); contrasted with Section 3, p. 11]
    "In the present study, we focus on crystal growth and melting during the materials synthesis. Therefore, we also add structures near the melting temperature to fine-tune the ANN-ML model. We use the pretrained ANN-ML model to generate snapshots close to the crystal melting. Then, we calculate the energy and forces of these snapshots using DFT. We found that adding these new training data can improve the performance of the ANN-ML model significantly, especially for melting temperature prediction."

    The paper's central evidence of transferability is the 5-10% agreement with experimental melting temperatures for LaSiP3 and La2SiP4. But Section 4 explicitly states that structures near melting were added to the training set for these systems, and Fig. 2(d) shows that this addition substantially changed the predicted Tm for La2SiP4. The final model therefore interpolates DFT energies and forces from the very solid-liquid transition configurations probed in the coexistence simulations, so the melting-temperature agreement is partly a consequence of the training choices rather than an independent prediction.

  2. fitted input called prediction [Section 2 (training data) and Section 3 (E-V transferability assessment)]
    "For each of these parent crystalline structures, 21 to 41 non-distorted structures are generated by uniformly expanding or contracting the volume of the initial structure, then 2100 to 4100 distorted structures are generated by randomly distorting atomic positions from their lattice position with a displacement amplitude of 0.2 Å. In this way, a total of 71,800 'distorted' crystal structures are generated for the ANN-ML training."

    Section 3 evaluates transferability by comparing ANN-ML E-V curves with DFT for 'all known stable elemental, binary, and ternary phases' and some ML-predicted ternary phases. However, Section 2 states that the parent crystalline structures used to generate the training data are exactly the stable elemental, binary, ternary, and ML-predicted phases from the Materials Project and recent ML studies. Thus the E-V agreement shown in Fig. 4 is an in-sample reproduction of the training data, not an out-of-sample prediction. The same applies to the liquid g(r) comparisons, since liquid snapshots at the same compositions were part of the training set. This does not invalidate the potential, but it means these comparisons cannot support the transferability claim.

full rationale

The paper is a standard ANN-ML potential development: a neural network is fitted to DFT energies and forces for a deliberately constructed set of crystals and liquids, and the reported RMS errors, E-V curves, and pair-correlation functions are largely in-sample accuracy checks of that fit. The strongest claimed independent output is the prediction of melting temperatures by solid-liquid coexistence MD. That claim is partially circular because the paper's own Section 4 and Fig. 2(d) show that solid-to-liquid transition snapshots for the ternary phases were added to the training set specifically to improve melting-temperature prediction, and that this addition substantially changed the predicted Tm. Consequently, the 5-10% agreement with experiment is not a fully independent validation, and the Section 3 statement that melting temperatures were 'not explicitly included' is contradicted by the paper's own methods. No load-bearing self-citation, uniqueness-importation, or ansatz-smuggling issue is present; reference [57] is a future same-group work and is not needed for the main derivation. The circularity is partial, not total: the coexistence dynamics and nucleation observations still require the potential to integrate the trained data, but the key quantitative claim is weakened by the inclusion of melting-path configurations in training.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central contribution is a fitted interatomic potential. The free parameters are the hand-chosen hyperparameters and the large set of neural network weights (not enumerated) that are fitted to the DFT data. The key axioms are the accuracy of the DFT reference and the representational assumptions of the DeePMD framework. No new physical entities are introduced.

free parameters (5)
  • cutoff radius = 7.0 A
    Chosen by hand to balance cost and accuracy; all local environment descriptors are sampled within this radius (Section 2).
  • network architecture = 4 hidden layers, 120 nodes per layer
    Selected for the ANN-ML model; no systematic hyperparameter optimization is reported (Section 2).
  • activation function = tanh
    Chosen as the activation function for the neural network model (Section 2).
  • distortion amplitude = 0.2 A
    Used to generate distorted crystalline structures for training data; chosen ad hoc (Section 2).
  • liquid simulation temperature = 2500 K
    Used to generate liquid snapshots and for pair-correlation function comparisons; a protocol choice, not a fitted value (Sections 2 and 3).
assumptions (5)
  • domain assumption PBE-GGA DFT energies and forces are accurate enough to serve as reference data for the interatomic potential.
    The model is trained entirely on DFT data. Any error in the PBE functional, particularly for La 4f electrons, propagates into the potential. The paper uses non-spin-polarized PBE for all systems (Section 2).
  • domain assumption The DeePMD local environment descriptor within a 7.0 A cutoff can represent the potential energy surface of this ternary system.
    The model assumes that all relevant atomic interactions are captured within the cutoff radius and by the descriptor architecture (Section 2).
  • domain assumption The total potential energy can be decomposed as a sum of atomic energy contributions, E = sum Ei.
    This is the standard DeePMD assumption, stated in Section 2, and is not independently verified for this system.
  • domain assumption The solid/liquid coexistence simulation method yields the correct melting temperature for the given model potential.
    The method is taken as a reliable estimator of Tm (Section 3), following standard practice. Errors in the method itself are not assessed here.
  • domain assumption The iterative training process, including adding melting-path configurations and adjusting usage probabilities, leads to a transferable model without overfitting.
    The paper assumes that these refinements improve the model globally. The large Tm errors for LaSi and La (about 20 percent) show this assumption is only partially met for those phases (Section 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Developing a Neural Network Machine Learning Interatomic Potential for Molecular Dynamics Simulations of La-Si-P Systems." pith.science (2026). https://pith.science/paper/RLOOSTS4

@misc{pith2026250608339,
  author       = {Pith},
  title        = {Pith review of: Developing a Neural Network Machine Learning Interatomic Potential for Molecular Dynamics Simulations of La-Si-P Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RLOOSTS4}},
  note         = {Machine review of arXiv:2506.08339}
}
read the original abstract

While molecular dynamics (MD) is a very useful computational method for atomistic simulations, modeling the interatomic interactions for reliable MD simulations of real materials has been a long-standing challenge. In 2007, Behler and Perrinello first proposed and demonstrated an artificial neural network machine learning (ANN-ML) scheme, opening a new paradigm for developing accurate and efficient interatomic potentials for reliable MD simulation studies of the thermodynamics and kinetics of materials. In this paper, we show that an accurate and transferable ANN-ML interatomic potential can be developed for MD simulations of La-Si-P system. The crucial role of training data in the ML potential development is discussed. The developed ANN-ML potential accurately describes not only the energy versus volume curves for all the known elemental, binary, and ternary crystalline structures in La-Si-P system, but also the structures of La-Si-P liquids with various compositions. Using the developed ANN-ML potential, the melting temperatures of several crystalline phases in La-Si-P system are predicted by the coexistence of solid-liquid phases from MD simulations. While the ANN-ML model systematically underestimates the melting temperatures of these phases, the overall trend agrees with experiment. The developed ANN-ML potential is also applied to study the nucleation and growth of LaP as a function of different relative concentrations of Si and P in the La-Si-P liquid, and the obtained results are consistent with experimental observations.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages

  1. [1]

    universal

    Introduction Molecular dynamics (MD) simulation, for which Prof. Parrinello has introduced many groundbreaking methodologies over the last 45 years, is a powerful computational tool in studying the thermodynamics and kinetics of real materials in atomistic detail and has been widely used in today’s materials research [1,2]. However, a major bottleneck for...

    work page 1995

  2. [9]

    Z. Fan, Z. Zeng, C. Zhang, Y. Wang, K. Song, H. Dong, Y. Chen, T. Ala-Nissila, Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport, Phys. Rev. B 104, 104309(2021). [10] Y. Zhang, J. Xia, B. Jiang, Physically Motivated Recursively Embedded Atom Neural Networks: Incorpor...

    work page 2021

  3. [34]

    Peter Kaiser and Wolfgang Jeitschko, The Rare Earth Silicon Phosphides LnSi2P6(Ln= La, Ce, Pr, and Nd), Journal of Solid State Chemistry 124, 346-352(1996). [35] Y. S. Sun, J. D. Chen, S. D. Yang, B. X. Li, G. L. Chai, C. S. Lin, M. Luo, N. Ye, LaSiP3 and LaSi2P6: Two Excellent Rare-Earth Pnictides with Strong SHG Responses as Mid- and Far-Infrared Nonlin...

    arXiv 1996

  4. [47]

    Kresse, J

    G. Kresse, J. Furthmüller, Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set. Comput. Mater. Sci. 6, 15-50(1996). [48] G. Kresse, J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 54, 11169-11186(1996). [49] J. P. Perdew,...

    work page 1996

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.