Pith. sign in

REVIEW 3 major objections 9 minor 115 references

One platform, 60+ AI potentials, zero dependency hell

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 05:13 UTC pith:TOWLWPUZ

load-bearing objection MLIP Studio is a genuinely useful software platform paper — 60+ MLIPs unified in one free web interface with benchmarking. The headline 33× DFT speedup is real but cherry-picked from a worst-case starting geometry. the 3 major comments →

arxiv 2607.07606 v1 pith:TOWLWPUZ submitted 2026-07-08 cond-mat.mtrl-sci

MLIP Studio: An Open Platform for Interactive Benchmarking and Atomistic Simulations Using Machine Learning Interatomic Potentials

Manas Sharma , Sudeep Punnathanam , Ananth Govind Rajan This is my paper
classification cond-mat.mtrl-sci
keywords machine learning interatomic potentialsuniversal MLIPsatomistic simulationbenchmarkingDFT accelerationgraph neural networkscomputational materials sciencegeometry optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces MLIP Studio, a free web-based platform that consolidates over 60 universal machine learning interatomic potentials (MLIPs) from six major model families into a single interactive interface. The core problem it addresses is practical: these foundation models, which approximate quantum-mechanical DFT calculations at a fraction of the cost, are scattered across mutually incompatible software environments, making systematic comparison and adoption difficult. MLIP Studio resolves this by providing a unified environment where users can upload a structure, switch between any of the 60+ models instantly, and run end-to-end simulation workflows — energy/force/stress prediction, geometry optimization, vibrational analysis, equation-of-state fitting, spin-state determination, and batch benchmarking against DFT reference data — without writing code or managing dependencies. The platform also supports custom model upload, wall-time benchmarking across CPU and GPU hardware, and automated generation of parity plots and element-resolved error tables. Through a case study on CrCl₃ adsorption on sapphire, the authors demonstrate that the best MLIP for a given task is property-dependent: the model with the lowest energy error differs from the model with the lowest force error, making side-by-side comparison essential rather than optional.

Core claim

The central contribution is the platform itself and the demonstration that consolidating 60+ MLIPs into a dependency-resolved, code-free environment is both technically achievable and practically transformative for how researchers select and deploy these models. A secondary empirical claim is that MLIP-based geometry pre-optimization can reduce subsequent DFT optimization effort, with the authors reporting speedups ranging from approximately 5× to 33× (measured in SCF steps) across three test systems. The largest speedup corresponds to a 64-water-molecule periodic box with severe initial atomic overlaps from Packmol, where MLIP pre-optimization brought the starting maximum force from ~1.3 eV

What carries the argument

Universal machine learning interatomic potentials (MLIPs): graph neural network models pretrained on millions of DFT calculations that predict energies, forces, and stresses for arbitrary atomic configurations without system-specific retraining. The platform unifies six architectural families — MACE, FairChem/UMA, ORB, MatterSim, SevenNet, and PET — which differ in how they enforce rotational symmetry (explicit equivariant tensors vs. data-augmented approximate invariance), whether they predict forces conservatively (via energy gradients) or directly, and how they trade accuracy against inference speed.

Load-bearing premise

The headline ~33× DFT speedup comes from a single worst-case starting geometry — a Packmol-generated water box with severe atomic overlaps and initial forces of ~1.3 eV/Å — where any reasonable pre-relaxation method would yield large speedups. The paper does not establish whether comparable speedups hold for typical DFT workflows starting from moderately reasonable structures.

What would settle it

If MLIP pre-optimization were tested on a broad set of starting structures with moderate (not pathological) initial forces and the median DFT speedup fell below ~2×, the practical-acceleration claim would lose its force. Similarly, if the platform's dependency-resolved environment were shown to produce silently incorrect results for specific model combinations (validated against known DFT reference values), the core value proposition of unified, trustworthy multi-model comparison would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If MLIP pre-optimization reliably reduces DFT effort by even 5–10× on moderately reasonable starting structures (not just worst-case overlapped geometries), it changes the economics of DFT workflows: researchers could routinely pre-screen with MLIPs and reserve DFT for final refinement, effectively expanding the system sizes accessible to first-principles methods.
  • The finding that the best model depends on the target observable (energy vs. forces vs. lattice parameters) implies that single-model evaluations are insufficient for trustworthy MLIP deployment; multi-model comparison, as the platform enables, should become standard practice.
  • Cross-model consistency on the 1,000-configuration water-box potential energy surface — where eight independently trained MLIPs agreed on the same highest- and lowest-energy structures — suggests that these foundation models are converging toward shared physical representations, which would strengthen confidence in their use for configuration screening even without DFT validation.
  • The custom model upload feature, combined with automated parity plotting, creates a rapid feedback loop for model developers: finetuned models can be benchmarked against DFT reference data in minutes without writing code, potentially accelerating the iterative refinement cycle.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The ~33× DFT speedup is measured on a single extreme case (Packmol-generated water box with ~1.3 eV/Å initial forces). The other two test systems show 5× and 26× speedups, and the 26× case compares against an unoptimized starting geometry rather than the more standard UFF pre-relaxation baseline (where the speedup drops to ~23×). The generalizability of these speedups to typical DFT workflows with
  • The platform's value as a benchmarking tool would be significantly enhanced if it supported standardized benchmark datasets (analogous to MatBench Discovery) with pre-loaded reference data, enabling apples-to-apples comparisons across the 60+ models on identical chemical spaces rather than relying on user-supplied trajectories.
  • The plugin architecture's ability to accommodate new models with minimal effort is a design claim that has not yet been stress-tested against the rapid pace of MLIP development; if new architectures introduce fundamentally different inference interfaces (e.g., non-graph-based or multi-modal models), the integration cost may be higher than the modular design anticipates.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 9 minor

Summary. This manuscript introduces MLIP Studio, a free, web-based platform that provides unified access to over 60 universal machine learning interatomic potentials (MLIPs) from six model families (MACE, FAIRChem/UMA, ORB, MatterSim, SevenNet, PET) within a single interactive interface. The platform supports end-to-end atomistic simulation workflows including energy/force/stress prediction, geometry optimization, vibrational analysis, equation-of-state fitting, spin-state determination, batch benchmarking against DFT reference data, and custom model deployment. The authors demonstrate the platform's capabilities through a pre-optimization study showing reduced DFT effort, CPU/GPU performance benchmarks across models, and a comprehensive case study on CrCl₃ adsorption on sapphire that illustrates task-specific MLIP selection. The source code is publicly available on GitHub, and a hosted instance is provided.

Significance. MLIP Studio addresses a genuine practical gap in the computational chemistry and materials science community: the fragmentation of MLIP frameworks across incompatible software environments makes systematic, fair benchmarking difficult. By providing a unified, free, and interactive platform supporting 60+ models, the tool lowers the barrier to MLIP adoption for both research and education. The platform ships with reproducible code (GitHub repository), provides falsifiable benchmarking data (CPU/GPU timing, parity metrics against DFT reference trajectories), and includes a custom multi-stage optimizer (Fast-MSO). The CrCl₃/sapphire case study demonstrates a practical end-to-end workflow for task-specific model selection. The computational performance benchmarks (Section 4.9) provide useful indicative guidance for practitioners choosing between models on different hardware. The platform's plugin-based architecture and support for custom MACE model uploads add practical value for model developers.

major comments (3)
  1. §4.3.1, Table 2, and Abstract: The headline ~33× DFT speedup claim is drawn from the most extreme of three test systems — a Packmol-generated (H₂O)₆₄ box with substantial atomic overlaps (initial max force ~1.3 eV/Å). The paper itself acknowledges this dramatic improvement 'arises because the randomly packed Packmol geometries contain substantial atomic overlaps and high initial forces.' The other two test systems show 5× and 26× speedups. No test system uses a moderately reasonable starting geometry (e.g., a crystal from a database or a molecule built with standard bond lengths), which is the regime most DFT practitioners operate in. Additionally, the 33× figure is measured in SCF steps (5,939 → 180), while in ionic steps the speedup is ~27× (649 → 24). The abstract elevates the SCF-step metric from the single extreme case without qualification. The authors should either (a) add at lea[
  2. §4.3.1, Table 2: No UFF pre-optimization comparison is provided for the (H₂O)₆₄ water box, unlike the caffeine and ibuprofen cases. Since the paper frames MLIP pre-optimization as an alternative to the 'conventional' UFF pre-relaxation strategy, the absence of a UFF baseline for the system that produces the headline speedup figure is a gap. Adding this comparison (or explaining why it was omitted) would strengthen the claim that MLIP pre-optimization offers a meaningful advantage over standard practice.
  3. §4.9, Fig. 10: The CPU and GPU benchmarks are performed on a single (H₂O)₆₄ configuration set with 1,000 frames. While the authors note that 'benchmarking the performance of computational codes is inherently sensitive to specific configurations,' the system-size dependence of MLIP inference is well-known and can change relative rankings between models. The paper would benefit from at least one additional system size (e.g., a smaller molecule or a larger periodic system) to demonstrate whether the recommended model rankings (PET-MAD XS, ORBv3 OMAT Direct, MACE OMAT-0 Small for GPU) are robust to system size. This is relevant because the recommendations in §4.9.2 are presented as general guidance.
minor comments (9)
  1. §2.1, Table 1: The text in §1 states 'six major model families' but §2.1 lists five families in the table header (MACE, FAIRChem, ORB, MatterSim, SevenNet, PET) — this is actually six, but the §2 introduction text says 'five major model families.' Please reconcile.
  2. §4.3: The Fast-MSO optimizer is described with default thresholds f_FIRE = 0.8 and f_MDMin = 0.25 eV/Å, but no validation or comparison against single-optimizer performance is provided. A brief demonstration (even one system) showing that Fast-MSO outperforms its constituent optimizers used individually would justify its inclusion.
  3. §4.6: The spin-state determination is validated on five molecules, all of which are relatively small. The CrCl₃ case is discussed in the case study (§5) but with different models (MACE POLAR 1 M, UMA OMOL s1p1, MACE OMOL XL 1024) than the UMA OMOL s1.1 used in §4.6. Clarifying which model is recommended for spin-state determination would help users.
  4. §5, Table 5: The force MAE for UMA OMAT s1p1 is reported as 51.8 eV/Å, which appears to be a typo — it should likely be 51.8 meV/Å, consistent with the other entries in the table and the RMSE of 95.5 meV/Å.
  5. §4.7.2: The text states that 1,000 water-box configurations were evaluated, but Figure 8 shows only the first 100. While the reason is given ('to be able to see a clear distinction'), providing the full 1,000-frame comparison in the Supporting Information would be useful.
  6. §4.10: The MLIP Studio QM9 Gap model is described as a 'four-layer message-passing graph neural network' with a test MAE of 0.062 eV, but no details on training/validation split, architecture specifics, or comparison with the cited models (SchNet, PPGN, etc.) beyond the MAE number are provided. A brief description or reference to where these details can be found would be appropriate.
  7. §1: The phrase 'source-available' is used to describe the license, but the ASL (Academic Software License) is not widely known. Clarifying whether this is OSI-approved or what restrictions it imposes (e.g., commercial use) would help potential users.
  8. §4.3.1: The DFT convergence criteria differ between the molecular systems (10⁻³ a.u./Bohr for forces, TURBOMOLE) and the water box (0.05 eV/Å, VASP). While this is understandable due to different codes, a brief note acknowledging that the speedup factors are not directly comparable across systems due to differing convergence thresholds would improve clarity.
  9. References: Several arXiv preprints are cited without published versions where they may exist (e.g., ref 15, ref 17, ref 18). The authors should check for published versions where available.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a constructive and balanced assessment. All three major comments are well-taken and have prompted concrete revisions to the manuscript, as detailed below.

read point-by-point responses

  1. Referee: §4.3.1, Table 2, and Abstract: The headline ~33× DFT speedup claim is drawn from the most extreme of three test systems — a Packmol-generated (H₂O)₆₄ box with substantial atomic overlaps (initial max force ~1.3 eV/Å). The paper itself acknowledges this dramatic improvement 'arises because the randomly packed Packmol geometries contain substantial atomic overlaps and high initial forces.' The other two test systems show 5× and 26× speedups. No test system uses a moderately reasonable starting geometry (e.g., a crystal from a database or a molecule built with standard bond lengths), which is the regime most DFT practitioners operate in. Additionally, the 33× figure is measured in SCF steps (5,939 → 180), while in ionic steps the speedup is ~27× (649 → 24). The abstract elevates the SCF-step metric from the single extreme case without qualification. The authors should either (a) add at lea[

    Authors: The referee is correct on all counts. The abstract currently elevates the single most extreme metric (33× in SCF steps) from the single most extreme test system without adequate qualification. This is misleading and we will revise it. Specifically: (1) The abstract will be rewritten to report the range of speedups observed across all three test systems (approximately 5× to 27× in ionic steps, with the SCF-step metric reaching ~33× for the water box specifically due to the high initial forces in the Packmol-generated geometry). (2) We will add a fourth test system with a moderately reasonable starting geometry — a crystal structure imported from the Materials Project database — to demonstrate the speedup in the regime most DFT practitioners actually operate in. We expect the speedup to be more modest in this case and will report it honestly. (3) In Section 4.3.1 and Table 2, we will clarify that the 33× figure refers specifically to SCF steps for the (H₂O)₆₄ system, while the ionic-step speedup for the same system is ~27×, and will present both metrics consistently for all systems. (4) We will add explicit discussion noting that the magnitude of speedup depends strongly on the quality of the initial geometry, with poorly packed structures (high initial forces) benefiting the most. We agree that the abstract should not cherry-pick the most favorable number. revision: yes

  2. Referee: §4.3.1, Table 2: No UFF pre-optimization comparison is provided for the (H₂O)₆₄ water box, unlike the caffeine and ibuprofen cases. Since the paper frames MLIP pre-optimization as an alternative to the 'conventional' UFF pre-relaxation strategy, the absence of a UFF baseline for the system that produces the headline speedup figure is a gap. Adding this comparison (or explaining why it was omitted) would strengthen the claim that MLIP pre-optimization offers a meaningful advantage over standard practice.

    Authors: This is a fair point. The UFF baseline was omitted for the (H₂O)₆₄ system because UFF, as implemented in ASE, does not handle periodic systems with the same robustness as it handles molecular systems, and we encountered convergence difficulties when attempting UFF pre-relaxation on the water box. However, the referee is correct that this omission is inadequately explained, particularly given that the water box is the system producing the headline speedup figure. We will address this in the revised manuscript by either (a) performing the UFF pre-optimization using a different implementation (e.g., via LAMMPS or another code that handles periodic UFF more reliably) and adding the result to Table 2, or (b) if we cannot obtain a converged UFF result in a reasonable timeframe, explicitly stating in the text that UFF pre-optimization was attempted but did not converge for this periodic system, and explaining why this itself is informative — it illustrates a practical limitation of classical force fields for disordered periodic systems where MLIPs succeed. Either way, the gap in the current presentation will be closed. revision: yes

  3. Referee: §4.9, Fig. 10: The CPU and GPU benchmarks are performed on a single (H₂O)₆₄ configuration set with 1,000 frames. While the authors note that 'benchmarking the performance of computational codes is inherently sensitive to specific configurations,' the system-size dependence of MLIP inference is well-known and can change relative rankings between models. The paper would benefit from at least one additional system size (e.g., a smaller molecule or a larger periodic system) to demonstrate whether the recommended model rankings (PET-MAD XS, ORBv3 OMAT Direct, MACE OMAT-0 Small for GPU) are robust to system size. This is relevant because the recommendations in §4.9.2 are presented as general guidance.

    Authors: The referee raises a valid concern. MLIP inference cost does not scale uniformly across architectures — some models have higher fixed overhead but better scaling with system size, while others show the opposite pattern. Relative rankings can indeed shift with system size. We will address this by adding benchmarks on at least one additional system size. Specifically, we plan to benchmark a smaller system (e.g., a single water molecule or a small molecule such as caffeine, ~20–30 atoms) to complement the 192-atom (H₂O)₆₄ box. This will allow us to demonstrate whether the recommended model rankings are robust or shift with system size. We will also add a brief discussion of the expected scaling behavior: models with larger fixed overhead (e.g., UMA) are expected to be relatively more disadvantaged for small systems, while models with efficient message passing (e.g., PET-MAD XS) may maintain their advantage across sizes. We agree that the current recommendations in §4.9.2 should be qualified as applying specifically to the system size benchmarked, with the additional data providing evidence of how generalizable the rankings are. revision: yes

Circularity Check

0 steps flagged

No circularity found — software platform paper with empirical demonstrations

full rationale

This is a software platform paper describing MLIP Studio, a web application for benchmarking and running atomistic simulations with machine learning interatomic potentials. The paper makes no first-principles derivation claims that could be circular. The headline ~33× DFT speedup (Section 4.3.1, Table 2) is an empirical measurement from a specific test system (Packmol-generated (H₂O)₆₄ box), not a derived result that reduces to its inputs. The CrCl₃/sapphire case study (Section 5) uses DFT reference data from ref 95, where three of the four current authors are co-authors. However, this self-citation is not load-bearing for the paper's central claim (that MLIP Studio is a useful unified platform). The reference data serves as an external benchmark for demonstrating the platform's capabilities — any other DFT reference dataset would serve the same purpose. The platform's utility does not depend on any theoretical premise established in ref 95. No uniqueness theorems are invoked, no parameters are fitted and then presented as predictions, and no ansatz is smuggled through citation. The derivation chain is straightforward: external MLIP models (from MACE, FAIRChem, ORB, MatterSim, SevenNet, PET) are applied to structures, and results are compared against DFT references. The paper is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

MLIP Studio is a well-executed software platform paper that addresses a real gap in the MLIP ecosystem. The unified access to 60+ models and the benchmarking capabilities are genuinely useful. However, the headline ~33× DFT speedup is drawn from an extreme case (randomly packed water molecules with atomic overlaps) and may not generalize to typical workflows. The case study relies on the authors' own prior DFT data. The paper is primarily a software description and demonstration rather than a scientific advance, which limits its significance. The platform's long-term value depends on continued maintenance as the MLIP field evolves rapidly. Code and data are publicly available, supporting reproducibility.

free parameters (3)
  • f_FIRE threshold = 0.8 eV/Å
    Force threshold for transitioning from FIRE to MDMin in the Fast-MSO optimizer. Chosen heuristically as a default value (Section 4.3).
  • f_MDMin threshold = 0.25 eV/Å
    Force threshold for transitioning from MDMin to LBFGS in the Fast-MSO optimizer. Chosen heuristically as a default value (Section 4.3).
  • QM9 Gap model architecture (4-layer MPNN) = N/A
    Model architecture choice for the in-house HOMO-LUMO gap predictor. Test MAE of 0.062 eV reported on QM9 test set (Section 4.10).
axioms (3)
  • domain assumption MLIPs trained on diverse DFT datasets achieve transferable accuracy across chemical systems without system-specific retraining.
    Foundational assumption underlying the entire platform. Invoked in Section 2 and throughout. This is the standard assumption of the universal MLIP field, supported by the cited literature.
  • ad hoc to paper Library versions that differ from each MLIP package's default requirements still produce correct and validated results.
    Section 1 states that compatibility testing identified library versions that 'while not always matching those specified in each package's default requirements file, nonetheless produce correct and validated results.' No systematic validation protocol or numerical tolerance is provided for this claim.
  • domain assumption MLIP pre-optimized structures are sufficiently close to the DFT minimum to accelerate convergence without introducing artifacts.
    Assumed in Section 4.3.1. The paper does not discuss cases where MLIP pre-optimization might converge to a different local minimum than DFT would reach.
invented entities (2)
  • Fast-MSO (Fast Multi-Stage Optimizer) no independent evidence
    purpose: Chains FIRE→MDMin→LBFGS optimizers with force-threshold-based transitions for geometry optimization.
    A custom optimizer introduced in this paper. No comparison against single-optimizer performance is provided to justify the multi-stage design. The transition thresholds are heuristic defaults.
  • MLIP Studio QM9 Gap model independent evidence
    purpose: Predicts HOMO-LUMO gaps of organic molecules using a 4-layer message-passing GNN trained on QM9.
    Test MAE of 0.062 eV on the QM9 test set is reported, which is a standard falsifiable benchmark. However, no external validation on non-QM9 molecules is provided.

pith-pipeline@v1.1.0-glm · 27080 in / 3877 out tokens · 325583 ms · 2026-07-09T05:13:23.172330+00:00 · methodology

0 comments
read the original abstract

Universal machine learning interatomic potentials (MLIPs) are foundation AI models transforming atomistic simulations, but their practical use remains hindered by fragmented software ecosystems, dependency conflicts, and the lack of accessible benchmarking tools. These models approach first-principles density functional theory (DFT) accuracy at a fraction of the computational cost. We introduce MLIP Studio (available at https://mlipstudio.iisc.ac.in), an open and free platform that brings more than 60 universal MLIPs into a unified interactive interface for molecules and materials. The platform enables end-to-end MLIP-driven workflows, including property prediction, geometry optimization, vibrational and equation-of-state analysis, spin-state determination, custom model deployment, and high-throughput benchmarking against reference data. Automated parity plots and sortable error tables facilitate rapid identification of element-wise outliers and problematic data points. We demonstrate that MLIP-based pre-optimization can reduce subsequent DFT optimization effort by ~33$\times$. Additionally, the application enables benchmarking of computational performance. Through a comprehensive case study involving the 2D magnetic material CrCl$_3$ on a sapphire substrate, we show how cross-model comparisons of various properties and potential-energy landscapes can guide task-specific MLIP selection. Overall, MLIP Studio lowers the barrier to the reliable use of foundation models in end-to-end research workflows, benchmarking, and education in computational chemistry and materials science.

Figures

Figures reproduced from arXiv: 2607.07606 by Ananth Govind Rajan, Manas Sharma, Sudeep Punnathanam.

Figure 1
Figure 1. Figure 1: Annotated screenshot of the MLIP Studio home page. (a) Input options panel showing [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Screenshots of MLIP Studio output for representative calculations. (a) Energy, force, [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Geometry optimization output for ibuprofen using the UMA OMOL s1.1 model and [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Structures used for the MLIP pre-optimization study. (a) Caffeine molecule. (b) Ibupro [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Vibrational mode analysis results for an H [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Equation of state calculation for silicon using the MACE OMAT-0 Medium model. (a) [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Spin-state determination using the UMA OMOL s1.1 model. Energy as a function of the [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Potential energy profiles for the first 100 of 1,000 randomly packed [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Benchmarking of the MACE OMAT-0 Medium model against DFT reference data for [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Computational performance benchmarks for universal MLIPs, measured as wall-clock [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Prediction of dipole moment, atom-resolved partial charges, band gap, and DOS. [PITH_FULL_IMAGE:figures/full_fig_p032_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Rotational potential energy curves for a tilted CrCl [PITH_FULL_IMAGE:figures/full_fig_p036_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

115 extracted references · 115 canonical work pages · 1 internal anchor

  1. [1]

    and Sham, L

    Kohn, W. and Sham, L. J. , title =. Phys. Rev. , volume =

    work page

  2. [2]

    and Furthmüller, J

    Kresse, G. and Furthmüller, J. , title =. Comput. Mater. Sci. , volume =

    work page

  3. [3]

    Staroverov, V. N. and Scuseria, G. E. and Tao, J. and Perdew, J. P. , title =. Phys. Rev. B , volume =

    work page

  4. [4]

    and Marsman, M

    Paier, J. and Marsman, M. and Hummer, K. and Kresse, G. and Gerber, I. C. and Ángyán, J. G. , title =. J. Chem. Phys. , volume =

    work page

  5. [5]

    and Tran, F

    Haas, P. and Tran, F. and Blaha, P. , title =. Phys. Rev. B , volume =

    work page

  6. [6]

    Nørskov, J. K. and Abild-Pedersen, F. and Studt, F. and Bligaard, T. , title =. Proc. Natl. Acad. Sci. U. S. A. , volume =

    work page

  7. [7]

    , title =

    Burke, K. , title =. J. Chem. Phys. , volume =

    work page

  8. [8]

    and Grabowski, B

    Freysoldt, C. and Grabowski, B. and Hickel, T. and Neugebauer, J. and Kresse, G. and Janotti, A. and Van de Walle, C. G. , title =. Rev. Mod. Phys. , volume =

    work page

  9. [9]

    P.; Burke, K.; Ernzerhof, M

    Perdew, John P. and Burke, Kieron and Ernzerhof, Matthias , journal =. 1996 , publisher =. doi:10.1103/PhysRevLett.77.3865 , url =

    work page doi:10.1103/physrevlett.77.3865 1996

  10. [10]

    Perdew, Kieron Burke, and Matthias Ernzerhof

    Perdew, John P. and Burke, Kieron and Ernzerhof, Matthias , journal =. 1997 , publisher =. doi:10.1103/PhysRevLett.78.1396 , url =

    work page doi:10.1103/physrevlett.78.1396 1997

  11. [11]

    2019 , publisher=

    Sharma, Manas and Mishra, Debabrata , journal=. 2019 , publisher=

    work page 2019

  12. [12]

    2025 , publisher=

    Jacobs, Ryan and Morgan, Dane and Attarian, Siamak and Meng, Jun and Shen, Chen and Wu, Zhenghao and Xie, Clare Yijia and Yang, Julia H and Artrith, Nongnuch and Blaiszik, Ben and others , journal=. 2025 , publisher=

    work page 2025

  13. [13]

    Chiang, Yuan and Kreiman, Tobias and Weaver, Elizabeth and Amin, Ishan and Kuner, Matthew and Zhang, Christine and Kaplan, Aaron and Chrzan, Daryl and Blau, Samuel M and Krishnapriyan, Aditi S and others , booktitle=

    work page

  14. [14]

    2026 , publisher=

    Yuan, Eric C-Y and Liu, Yunsheng and Chen, Junmin and Zhong, Peichen and Raja, Sanjeev and Kreiman, Tobias and Vargas, Santiago and Xu, Wenbin and Head-Gordon, Martin and Yang, Chao and others , journal=. 2026 , publisher=

    work page 2026

  15. [15]

    Journal of the American Chemical Society , volume=

    Kov. Journal of the American Chemical Society , volume=. 2025 , publisher=

    work page 2025

  16. [16]

    Barroso-Luque, Luis and Shuaibi, Muhammed and Fu, Xiang and Wood, Brandon M and Dzamba, Misko and Gao, Meng and Rizvi, Ammar and Zitnick, C Lawrence and Ulissi, Zachary W , journal=

    work page

  17. [17]

    Orb-v3: atomistic simulation at scale

    Rhodes, Benjamin and Vandenhaute, Sander and. arXiv preprint arXiv:2504.06231 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Neumann, Mark and Gin, James and Rhodes, Benjamin and Bennett, Steven and Li, Zhiyi and Choubisa, Hitarth and Hussey, Arthur and Godwin, Jonathan , journal=

    work page

  19. [19]

    Wood, Brandon M and Dzamba, Misko and Fu, Xiang and Gao, Meng and Shuaibi, Muhammed and Barroso-Luque, Luis and Abdelmaqsoud, Kareem and Gharakhanyan, Vahe and Kitchin, John R and Levine, Daniel S and others , journal=

    work page

  20. [20]

    2020 , publisher=

    Jain, Anubhav and Montoya, Joseph and Dwaraknath, Shyam and Zimmermann, Nils ER and Dagdelen, John and Horton, Matthew and Huck, Patrick and Winston, Donny and Cholia, Shreyas and Ong, Shyue Ping and others , booktitle=. 2020 , publisher=

    work page 2020

  21. [21]

    2013 , publisher=

    Jain, Anubhav and Ong, Shyue Ping and Hautier, Geoffroy and Chen, Wei and Richards, William Davidson and Dacek, Stephen and Cholia, Shreyas and Gunter, Dan and Skinner, David and Ceder, Gerbrand and others , journal=. 2013 , publisher=

    work page 2013

  22. [22]

    2016 , publisher=

    Mathew, Kiran and Singh, Arunima K and Gabriel, Joshua J and Choudhary, Kamal and Sinnott, Susan B and Davydov, Albert V and Tavazza, Francesca and Hennig, Richard G , journal=. 2016 , publisher=

    work page 2016

  23. [23]

    2025 , publisher=

    Horton, Matthew K and Huck, Patrick and Yang, Ruo Xi and Munro, Jason M and Dwaraknath, Shyam and Ganose, Alex M and Kingsbury, Ryan S and Wen, Mingjian and Shen, Jimmy X and Mathis, Tyler S and others , journal=. 2025 , publisher=

    work page 2025

  24. [24]

    Journal of Physics: Condensed Matter , volume=

    Hjorth Larsen, Ask and J. Journal of Physics: Condensed Matter , volume=. 2017 , publisher=

    work page 2017

  25. [25]

    Yang, Han and Hu, Chenxi and Zhou, Yichi and Liu, Xixian and Shi, Yu and Li, Jielan and Li, Guanzhi and Chen, Zekun and Chen, Shuizhou and Zeni, Claudio and others , journal=

    work page

  26. [26]

    Fu, Xiang and Wood, Brandon M and Barroso-Luque, Luis and Levine, Daniel S and Gao, Meng and Dzamba, Misko and Zitnick, C Lawrence , journal=

    work page

  27. [27]

    arXiv preprint arXiv:2511.20487 , year=

    Wehrhan, Leon and Walewski, Lucien and Bluntzer, Marie and Chomet, Heloise and Tilly, Jules and Brunken, Christoph and Acosta-Guti. arXiv preprint arXiv:2511.20487 , year=

    work page arXiv

  28. [28]

    Riebesell, Janosh and Goodall, Rhys EA and Benner, Philipp and Chiang, Yuan and Deng, Bowen and Lee, Alpha A and Jain, Anubhav and Persson, Kristin A , journal=

    work page

  29. [29]

    arXiv preprint arXiv:2505.08762 , year=

    Levine, Daniel S and Shuaibi, Muhammed and Spotte-Smith, Evan Walter Clark and Taylor, Michael G and Hasyim, Muhammad R and Michel, Kyle and Batatia, Ilyes and Cs. arXiv preprint arXiv:2505.08762 , year=

    work page arXiv

  30. [30]

    Introduction to transformers for NLP: With the hugging face library and models to solve problems , pages=

    Hugging face , author=. Introduction to transformers for NLP: With the hugging face library and models to solve problems , pages=. 2022 , publisher=

    work page 2022

  31. [31]

    Journal of Chemical Education , year =

    Seshadri, Keshavan and Liu, Peng and Koes, David Ryan , title =. Journal of Chemical Education , year =

    work page

  32. [32]

    Software Development , volume=

    Khorasani, Mohammad and Abdou, Mohamed and Fern. Software Development , volume=. 2022 , publisher=

    work page 2022

  33. [33]

    Advances in neural information processing systems , volume=

    Batatia, Ilyes and Kovacs, David P and Simm, Gregor and Ortner, Christoph and Cs. Advances in neural information processing systems , volume=

    work page

  34. [34]

    2022 , publisher=

    Batzner, Simon and Musaelian, Albert and Sun, Lixin and Geiger, Mario and Mailoa, Jonathan P and Kornbluth, Mordechai and Molinari, Nicola and Smidt, Tess E and Kozinsky, Boris , journal=. 2022 , publisher=

    work page 2022

  35. [35]

    2024 , publisher=

    Park, Yutack and Kim, Jaesun and Hwang, Seungwoo and Han, Seungwu , journal=. 2024 , publisher=

    work page 2024

  36. [36]

    2024 , publisher=

    Wang, Guanjie and Wang, Changrui and Zhang, Xuanguang and Li, Zefeng and Zhou, Jian and Sun, Zhimei , journal=. 2024 , publisher=

    work page 2024

  37. [37]

    2024 , publisher=

    Wan, Kaiwei and He, Jianxin and Shi, Xinghua , journal=. 2024 , publisher=

    work page 2024

  38. [38]

    2023 , publisher=

    Anstine, Dylan M and Isayev, Olexandr , journal=. 2023 , publisher=

    work page 2023

  39. [39]

    2022 , publisher=

    Lehtola, Susi and Karttunen, Antti J , journal=. 2022 , publisher=

    work page 2022

  40. [40]

    Journal of computational chemistry , volume=

    Hafner, J. Journal of computational chemistry , volume=. 2008 , publisher=

    work page 2008

  41. [41]

    Properties of Complex Inorganic Solids , pages=

    Hafner, J. Properties of Complex Inorganic Solids , pages=. 1997 , publisher=

    work page 1997

  42. [42]

    Journal of computational chemistry , volume=

    Mart. Journal of computational chemistry , volume=. 2009 , publisher=

    work page 2009

  43. [43]

    1970 , publisher=

    Broyden, Charles George , journal=. 1970 , publisher=

    work page 1970

  44. [44]

    1970 , publisher=

    Fletcher, Roger , journal=. 1970 , publisher=

    work page 1970

  45. [45]

    Goldfarb, Donald , journal=

    work page

  46. [46]

    Shanno, David F , journal=

    work page

  47. [47]

    Nocedal, Jorge , journal=

    work page

  48. [48]

    Nucleic acids research , volume=

    PubChem 2023 update , author=. Nucleic acids research , volume=. 2023 , publisher=

    work page 2023

  49. [49]

    Nucleic acids research , volume=

    PubChem substance and compound databases , author=. Nucleic acids research , volume=. 2016 , publisher=

    work page 2016

  50. [50]

    Nucleic acids research , volume=

    PubChem 2019 update: improved access to chemical data , author=. Nucleic acids research , volume=. 2019 , publisher=

    work page 2019

  51. [51]

    Nucleic acids research , volume=

    PubChem 2025 update , author=. Nucleic acids research , volume=. 2025 , publisher=

    work page 2025

  52. [52]

    2012 , publisher=

    Hanwell, Marcus D and Curtis, Donald E and Lonie, David C and Vandermeersch, Tim and Zurek, Eva and Hutchison, Geoffrey R , journal=. 2012 , publisher=

    work page 2012

  53. [53]

    2021 , publisher=

    Meunier, Marc and Robertson, Struan , journal=. 2021 , publisher=

    work page 2021

  54. [54]

    2019 , publisher=

    Bannwarth, Christoph and Ehlert, Sebastian and Grimme, Stefan , journal=. 2019 , publisher=

    work page 2019

  55. [55]

    Journal of the American chemical society , volume=

    Rapp. Journal of the American chemical society , volume=. 1992 , publisher=

    work page 1992

  56. [56]

    2016 , publisher=

    Mardirossian, Narbe and Head-Gordon, Martin , journal=. 2016 , publisher=

    work page 2016

  57. [57]

    2D Materials , volume=

    Gjerding, Morten Niklas and Taghizadeh, Alireza and Rasmussen, Asbj. 2D Materials , volume=. 2021 , publisher=

    work page 2021

  58. [58]

    2020 , publisher=

    Talirz, Leopold and Kumbhar, Snehal and Passaro, Elsa and Yakutovich, Aliaksandr V and Granata, Valeria and Gargiulo, Fernando and Borelli, Marco and Uhrin, Martin and Huber, Sebastiaan P and Zoupanos, Spyros and others , journal=. 2020 , publisher=

    work page 2020

  59. [59]

    2025 , note =

    MLIP Playground – Run, test and compare universal machine learning interatomic potentials , howpublished =. 2025 , note =

    work page 2025

  60. [60]

    Kaplan, Aaron D and Liu, Runze and Qi, Ji and Ko, Tsz Wai and Deng, Bowen and Riebesell, Janosh and Ceder, Gerbrand and Persson, Kristin A and Ong, Shyue Ping , journal=

    work page

  61. [61]

    2026 , note =

    Rowan: ML-Powered Molecular Design and Simulation , howpublished =. 2026 , note =

    work page 2026

  62. [62]

    2026 , note =

    QpiAI , howpublished =. 2026 , note =

    work page 2026

  63. [63]

    2026 , note =

    Quantistry , howpublished =. 2026 , note =

    work page 2026

  64. [64]

    2026 , note =

    Matlantis , howpublished =. 2026 , note =

    work page 2026

  65. [65]

    2011 , publisher=

    Grimme, Stefan and Ehrlich, Stephan and Goerigk, Lars , journal=. 2011 , publisher=

    work page 2011

  66. [66]

    Nocedal, Jorge and Wright, Stephen J , year=

    work page

  67. [67]

    Physical Review B , volume=

    Garijo del R. Physical Review B , volume=. 2019 , publisher=

    work page 2019

  68. [68]

    Physical review letters , volume=

    Bitzek, Erik and Koskinen, Pekka and G. Physical review letters , volume=. 2006 , publisher=

    work page 2006

  69. [69]

    2025 , publisher=

    Sharma, Manas and Franzke, Yannick J and Holzer, Christof and Pauly, Fabian and Sierka, Marek , journal=. 2025 , publisher=

    work page 2025

  70. [70]

    2005 , publisher=

    Weigend, Florian and Ahlrichs, Reinhart , journal=. 2005 , publisher=

    work page 2005

  71. [71]

    Geiger, Mario and Smidt, Tess , journal=

    work page

  72. [72]

    Murnaghan, Francis Dominic , journal=

    work page

  73. [73]

    1947 , publisher=

    Birch, Francis , journal=. 1947 , publisher=

    work page 1947

  74. [74]

    Vinet, PJJR and Ferrante, J and Smith, JR and Rose, JH , journal=

    work page

  75. [75]

    The Journal of Physical Chemistry A , volume=

    Csonka, G. The Journal of Physical Chemistry A , volume=. 2005 , publisher=

    work page 2005

  76. [76]

    Duval, Alexandre and Mathis, Simon V and Joshi, Chaitanya K and Schmidt, Victor and Miret, Santiago and Malliaros, Fragkiskos D and Cohen, Taco and Lio, Pietro and Bengio, Yoshua and Bronstein, Michael , journal=

    work page

  77. [77]

    2019 , publisher=

    Drautz, Ralf , journal=. 2019 , publisher=

    work page 2019

  78. [78]

    and Sharma, Piyush and Chiglintsev, Emil O

    Kumar, Vivek and Jangid, Abhishek and Sharma, Manas and Verma, Manvi and Pasyanthi, Jampala and Kumar, Keerthana S. and Sharma, Piyush and Chiglintsev, Emil O. and Panin, Mikhail I. and Punnathanam, Sudeep N. and Chernov, Alexander I. and Govind Rajan, Ananth and Singh, Akshay , title =. Advanced Materials , volume =. doi:https://doi.org/10.1002/adma.2025...

    work page doi:10.1002/adma.202514405

  79. [79]

    2021 , eprint=

    PFP: Universal Neural Network Potential for Material Discovery , author=. 2021 , eprint=

    work page 2021

  80. [80]

    2016--2018 , howpublished =

    work page 2016

Showing first 80 references.