Are Machine Learning Interatomic Potentials Truly Practical? A Benchmark of 23 Mainstream Models

Hanwen Kang; Miao Liu; Sheng Meng; Tenglong Lu

REVIEW 2 major objections 6 minor 4 references

Big AI atom models barely beat small ones

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 03:43 UTC pith:JS5HPNEC

load-bearing objection The throughput-vs-accuracy Pareto framing for 23 MLIPs is a genuine service to the field, but the central '3–5 meV/atom' claim is unsupported by the paper's own data. the 2 major comments →

arxiv 2607.07647 v1 pith:JS5HPNEC submitted 2026-07-08 cond-mat.mtrl-sci cs.CEphysics.comp-ph

Are Machine Learning Interatomic Potentials Truly Practical? A Benchmark of 23 Mainstream Models

Hanwen Kang , Tenglong Lu , Sheng Meng , Miao Liu This is my paper

classification cond-mat.mtrl-sci cs.CEphysics.comp-ph

keywords machine learning interatomic potentialsmolecular dynamicsPareto frontieraccuracy-efficiency trade-offbenchmarkgraph neural networksmaterials simulationDFT

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.

This paper benchmarks 23 mainstream machine-learning interatomic potentials (MLIPs) — neural-network models that approximate quantum-mechanical energy calculations for molecular dynamics — across three dimensions: prediction accuracy, simulation throughput, and maximum simulatable system size. The central finding is a sharp and systematic accuracy-efficiency trade-off. The largest state-of-the-art models (10–730 million parameters) deliver only 3–5 meV/atom more accuracy than lightweight models (under 5 million parameters), an improvement smaller than room-temperature thermal noise, while losing one to three orders of magnitude in throughput. In the worst cases, the heaviest MLIPs run less than twice as fast as the density functional theory (DFT) calculations they were meant to replace, effectively surrendering the entire reason MLIPs exist. Lightweight models, by contrast, occupy the Pareto frontier: they run hundreds to thousands of times faster than DFT, scale to hundreds of thousands of atoms on modest hardware, and their accuracy is competitive for nearly all routine molecular dynamics work. The paper argues that the field's dominant benchmark — Matbench Discovery — optimizes a single dimension (static accuracy) and has therefore incentivized model bloat with negligible practical payoff.

Core claim

The paper's central claim is that the accuracy gap between the heaviest and lightest MLIPs (3–5 meV/atom) is smaller than thermal noise at room temperature, while the throughput gap between them spans one to three orders of magnitude. This means the Pareto frontier of accuracy versus efficiency is dominated by lightweight and medium-scale models, not by the largest state-of-the-art architectures. The heaviest models, in the worst case, simulate at less than 2× DFT speed and are limited to 500–1,000-atom systems — comparable to what DFT itself can handle — making them impractical for the large-scale molecular dynamics that MLIPs were invented to enable. The paper also identifies a secondary,

What carries the argument

The central object is the Pareto frontier of accuracy (measured by thermal-conductivity error κ_SRME from the Matbench Discovery protocol) versus inference throughput (measured in MD steps per second, normalized to CPU-based DFT). The paper constructs this frontier across 23 models spanning 0.5–730 million parameters, on fixed hardware (NVIDIA DGX Spark, 80 GB memory cap), using a 192-atom LiCoO₂ test system. A second key mechanism is the material-density effect: because graph-neural-network MLIPs use a fixed 6–7 Å cutoff radius, denser materials generate more graph edges per atom, causing the same model to simulate two orders of magnitude fewer atoms in a dense oxide than in a sparse solid.

Load-bearing premise

The throughput and scalability rankings are established from a single 192-atom LiCoO₂ system on one hardware configuration. The paper itself acknowledges that quantitative results depend on system size, material density, and hardware, so the Pareto frontier shape — and the practical recommendations drawn from it — may not hold for larger systems, different materials, or different GPU architectures.

What would settle it

If a future large MLIP (10M+ parameters) were shown to achieve both top-tier accuracy and throughput comparable to lightweight models — for instance through architecture innovations that decouple parameter count from inference cost — the central trade-off claim would be weakened.

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

If this is right

If the 3–5 meV/atom accuracy gap is truly below thermal noise for most finite-temperature simulations, then the dominant strategy in MLIP development — scaling model parameters to climb accuracy leaderboards — is misaligned with practical utility and should be supplemented with efficiency and scalability metrics.
Lightweight MLIPs (1–5M parameters) can run on desktop-class hardware ($4,000–5,000) and scale to ~200,000 atoms, which could democratize large-scale molecular dynamics for research groups without HPC access.
The material-density effect means published MLIP scalability numbers are meaningless without specifying the target material, suggesting benchmarking standards need to report performance across a density-diverse material set.
If throughput rankings shift substantially for larger systems (where neighbor-list construction costs scale differently than at 192 atoms), the Pareto frontier shape could change, and the paper's recommendations would need revalidation at the system sizes practitioners actually use.

Where Pith is reading between the lines

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

The paper benchmarks throughput on a single 192-atom system of one material (LiCoO₂). Because DFT scales as O(N³) while MLIPs scale closer to O(N), the DFT-normalized speedup numbers will grow with system size — meaning the gap between heavyweight MLIPs and DFT could look worse or better at different system sizes than the 192-atom benchmark suggests.
The finding that large models gain only 3–5 meV/atom over lightweight ones raises the question of whether the accuracy ceiling is being approached for the current generation of training datasets. If so, further parameter scaling may yield even smaller returns, strengthening the case for redirecting effort toward efficiency and architecture optimization rather than model enlargement.
If lightweight models were fine-tuned or phonon-optimized (as the paper hints with the Nequix MP PFT model), they might close part of the accuracy gap while retaining their throughput advantage, potentially pushing the Pareto frontier further toward the accuracy corner and making the large-model tier even harder to justify.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

The throughput-vs-accuracy Pareto framing for 23 MLIPs is a genuine service to the field, but the central '3–5 meV/atom' claim is unsupported by the paper's own data.

read the letter

The headline you need: the Pareto framing across 23 models on unified hardware is a real contribution, and the qualitative finding — that large MLIPs trade orders-of-magnitude throughput for marginal accuracy gains — is correct and important. But the paper's signature quantitative claim ('3–5 meV/atom more accuracy') is never substantiated by any data shown, and that gap matters for the argument's load-bearing role. The stress-test note is right on this point. The paper's actual accuracy coordinate is κ_SRME (a thermal-conductivity relative error from Matbench Discovery), and the values reported are ≈0.12–0.24 for large models versus ≈0.31–0.58 for lightweight baselines — a 2–5× difference. No figure or table shows energy MAE for the 23 models. The 3–5 meV/atom figure appears in the abstract and §3.1 as an assertion, never traced to a source or derived from the κ_SRME data. Whether a 2–5× κ_SRME difference is 'marginal' is a substantive question the paper doesn't engage with; it just declares it so. This is the main soft spot, and it's not minor because the entire 'lightweight models are the practical choice' conclusion depends on the accuracy gap being negligible. What the paper does well: the unified ASE pipeline across 23 models on identical hardware, the memory-scaling analysis (Figure 3), and the material-density finding (Figure 4) are all genuinely useful. The observation that the same model handles 10⁵ atoms of LGPS but only ~8,000 atoms of LiCoO₂ is the kind of practical insight that working researchers actually need. The cross-validation with TorchSim (1.11–1.4× speedup) is a reasonable robustness check. The DFT-normalized throughput hierarchy (Figure 2) is informative, though the 192-atom LiCoO₂ single-material, single-size design limits generalizability — and the paper itself acknowledges this in §4.4. The self-citations ([41–43], [19] GPTFF) are from the same group, but GPTFF is included as one of 23 models on the same footing as the others, which is fine. No code or data release is mentioned, which weakens reproducibility. Who gets value: anyone selecting an MLIP for production MD work, and anyone developing MLIPs who needs to understand the efficiency landscape. The qualitative message lands even without the precise meV/atom claim. Recommendation: this deserves a serious referee. The core trade-off is real, the benchmarking effort is substantial, and the Pareto framing is a genuine contribution to evaluation culture. But the referee must push hard on the '3–5 meV/atom' claim — either show the energy MAE data or stop using that framing and justify why the κ_SRME differences are practically negligible. Without that fix, the central quantitative argument is unsupported.

Referee Report

2 major / 6 minor

Summary. This manuscript benchmarks 23 mainstream machine-learning interatomic potentials (MLIPs) across three dimensions—predictive accuracy (via the Matbench Discovery κ_SRME metric), MD simulation throughput, and GPU memory scalability—on a standardized, low-cost hardware platform (NVIDIA DGX Spark, 80 GB memory cap). The central finding is a sharp accuracy–efficiency trade-off: large SOTA models (10–730M parameters) achieve only marginally better accuracy than lightweight models (<5M parameters) while losing orders of magnitude in throughput and memory scalability. The authors argue that lightweight MLIPs sit on the Pareto frontier and are the practical choice for routine MD work. The study is timely, the multi-dimensional evaluation framework is well-motivated, and the practical model-selection guidance is useful for the community.

Significance. The manuscript addresses an important gap in MLIP evaluation: most existing benchmarks focus solely on static accuracy, ignoring inference efficiency and hardware scalability. By systematically benchmarking 23 models on standardized hardware with a unified pipeline, the authors provide a valuable service to working researchers. The identification of material-density-dependent scalability (§3.4) is a particularly insightful and underappreciated contribution. The practical recommendations in §4.3, if supported by the data, would be genuinely useful to the community. The use of an external metric (κ_SRME from Matbench Discovery) for accuracy and independent throughput measurements provides a reasonable foundation. However, the central quantitative claim ('3–5 meV/atom more accuracy') is not supported by the data presented in the paper, which is a significant gap that must be addressed.

major comments (2)

§3.1 and Abstract: The headline claim that large MLIPs deliver 'only 3–5 meV/atom more accuracy' than lightweight models is not supported by any data shown in the paper. The paper's actual accuracy metric is κ_SRME (a thermal-conductivity relative error), with values of ≈0.12–0.24 for large models versus ≈0.31–0.58 for lightweight baselines—a 2–5× difference. No figure or table presents energy MAE values for the 23 models, and the 3–5 meV/atom figure is never derived from the κ_SRME data or traced to a specific source. This is load-bearing because the 'marginal accuracy gain' narrative underpins the entire Pareto argument. The authors must either (a) present energy MAE data for the 23 models that supports the 3–5 meV/atom claim, or (b) justify why a 2–5× κ_SRME difference qualifies as 'marginal' for practical MD applications, or (c) reframe the claim in terms of κ_SRME directly.
§3.2: The throughput benchmark uses a single 192-atom LiCoO₂ cell, which is small relative to the systems MLIPs are designed to simulate. The paper itself acknowledges in §4.4 that 'quantitative numbers are dependent on system size, material density, and hardware configuration,' and §3.4 demonstrates that material density changes scalability by up to two orders of magnitude. Since the Pareto frontier argument depends on throughput rankings, and since per-atom overhead for large models may amortize differently at 1,000+ atoms, the authors should either (a) supplement the 192-atom benchmark with at least one larger system size (e.g., 1,000–10,000 atoms) to test whether rankings are stable, or (b) explicitly caveat the Pareto frontier claim as specific to the 192-atom LiCoO₂ workload and discuss how it might shift for larger or sparser systems.

minor comments (6)

§2: The κ_SRME equation is difficult to parse in the typeset form. Please ensure proper formatting of the summation indices, volume factor, and the symmetric relative mean error definition for the final version.
§3.1: The phrase 'Below the zero-point vibrational energy of a typical bond' appears as a sentence fragment after the κ_SRME values. Clarify or remove.
§3.3: The text mentions 'eSEN 30M' and 'PET-OAM XL' in the figure caption but 'PET-730M' in the body. Please reconcile the model names.
§4.3: The recommendation lists 'M3GNet, MACE-small' for the lightweight tier, but §3.1 highlights 'MatterSim v1 1M and Nequix MP PFT' as Pareto-optimal lightweight models. Please align the recommendations with the benchmark results.
References [41–43] and [19] (GPTFF) are from the same group as the authors. GPTFF appears as one of the benchmarked models. This is not problematic per se, but a brief disclosure of the relationship would be appropriate for transparency.
§3.2: The DFT baseline uses a Gamma-only 1×1×1 k-point mesh for a 192-atom cell. While this is a reasonable choice for a large supercell, it would be helpful to note whether this makes the DFT baseline faster or slower than a typical production calculation with denser k-point sampling.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for a careful and constructive report. The referee correctly identifies that our headline accuracy claim ('3–5 meV/atom') is not supported by the κ_SRME data actually presented in the manuscript, and that the throughput benchmark rests on a single 192-atom system. We address both major comments below. On Comment 1, we will reframe the claim in terms of κ_SRME and add supporting energy MAE data from Matbench Discovery where available. On Comment 2, we will add a larger-system throughput benchmark and explicitly caveat the Pareto frontier claim. Both revisions are substantive.

read point-by-point responses

Referee: §3.1 and Abstract: The headline claim that large MLIPs deliver 'only 3–5 meV/atom more accuracy' than lightweight models is not supported by any data shown in the paper. The paper's actual accuracy metric is κ_SRME, with values of ≈0.12–0.24 for large models versus ≈0.31–0.58 for lightweight baselines—a 2–5× difference. No figure or table presents energy MAE values for the 23 models, and the 3–5 meV/atom figure is never derived from κ_SRME or traced to a specific source. The authors must either (a) present energy MAE data supporting the claim, (b) justify why a 2–5× κ_SRME difference qualifies as 'marginal,' or (c) reframe the claim in terms of κ_SRME directly.

Authors: The referee is correct. The '3–5 meV/atom' figure appears in the abstract and §3.1 but is not derived from any data shown in the manuscript. Our primary accuracy metric is κ_SRME (the Matbench Discovery thermal-conductivity relative error), and the values we report—≈0.12–0.24 for large models versus ≈0.31–0.58 for lightweight baselines—represent a 2–5× relative difference, which we should not have described as 'marginal' without explicit justification. The '3–5 meV/atom' figure was drawn from published energy MAE values on the Matbench Discovery leaderboard for several of the models in our benchmark, but we failed to present those MAE values in the manuscript itself or to cite their source. This is a genuine gap in the paper. We will revise as follows: (1) The abstract and §3.1 will be rewritten to state the accuracy comparison directly in terms of κ_SRME, with the 2–5× relative difference reported transparently. (2) We will add a supplementary table listing energy MAE values from the Matbench Discovery leaderboard for the models in our benchmark, clearly attributed to that external source, so that readers can see both the κ_SRME and energy-MAE pictures. (3) We will add an explicit discussion of why a 2–5× κ_SRME difference may or may not be 'marginal' in practical MD contexts—specifically, that κ_SRME probes higher-order derivative quality of the learned PES, and that its practical impact on observable MD quantities (phase stability, defect migration barriers, phonon spectra) is not yet well calibrated to meV/atom-scale energy errors. We will not claim that the accuracy difference is 'marginal' without supporting this with either the κ_SRME data or the energy MAE data. We agree this is a major revision. revision: yes
Referee: §3.2: The throughput benchmark uses a single 192-atom LiCoO₂ cell, which is small relative to the systems MLIPs are designed to simulate. The paper itself acknowledges in §4.4 that quantitative numbers are dependent on system size, material density, and hardware configuration, and §3.4 demonstrates that material density changes scalability by up to two orders of magnitude. Since the Pareto frontier argument depends on throughput rankings, and since per-atom overhead for large models may amortize differently at 1,000+ atoms, the authors should either (a) supplement the 192-atom benchmark with at least one larger system size (e.g., 1,000–10,000 atoms) to test whether rankings are stable, or (b) explicitly caveat the Pareto frontier claim as specific to the 192-atom LiCoO₂ workload and discuss how it might shift for larger or sparser systems.

Authors: The referee raises a valid concern. The 192-atom LiCoO₂ cell was chosen for the throughput benchmark to provide a uniform, reproducible comparison point across all 23 models, including those that cannot fit larger systems in 80 GB of memory. However, we agree that a single system size is insufficient to establish that throughput rankings are stable, and the referee is right that per-atom overhead for large models may amortize differently at larger scales. We will address this in two ways. First, we will add a supplementary throughput benchmark at a larger system size (approximately 2,000–5,000 atoms of LiCoO₂, depending on which models can fit) covering the subset of models that can run at that scale. This will directly test whether throughput rankings change. We expect that large models with high fixed overhead will, if anything, become relatively more competitive on a per-atom basis at larger sizes, but we will report the data honestly regardless of the outcome. Second, we will explicitly caveat the Pareto frontier claim in §3.1 and §4.3 as specific to the 192-atom LiCoO₂ workload, and add a discussion of how the Pareto frontier may shift for larger systems, sparser materials (as shown in §3.4), and different hardware configurations. We note that the memory-scalability results in §3.4 already show that the same model can handle systems differing by two orders of magnitude depending on material density, which reinforces the need for this caveat. We agree this is a major revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; benchmark study with independent measurements, minor non-load-bearing self-citations

full rationale

This is a benchmark study, not a derivation chain, and the central claims are constructed from independently measured data: accuracy is quantified using the external Matbench Discovery κ_SRME metric (Eq. in §2, ref [36]), and throughput is measured on standardized hardware via a unified ASE pipeline. The Pareto frontier (Figure 1) is built from these two independently obtained coordinates, not from any self-referential definition or fit-then-predict loop. The paper does contain four self-citations from the same group: [19] GPTFF (Xie, Lu, Meng, Liu) is one of the 23 benchmarked models, and [41–43] (Kang/Lu/Meng/Liu and collaborators) are cited in §3.1 as supporting references for the claim that lightweight models suffice for practical MD work ('they offer more than enough accuracy at a cost a desktop machine can bear [41] [42] [43]'). However, these citations are not load-bearing for the central Pareto argument: the frontier is established by the measured data points themselves, not by the cited works. GPTFF is not favored in the paper's conclusions (§4.3 recommends M3GNet and MACE-small, not GPTFF), and its use in §3.4 as an illustrative example of material-density effects is a legitimate technical point that would hold for any GNN-based MLIP. The skeptic's concern that the '3–5 meV/atom' headline claim is not directly derived from the κ_SRME data is a correctness/unsupported-claim issue, not a circularity issue—the claim is not defined in terms of its own outputs. No step in the paper reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new entities are postulated. The paper is a benchmark study measuring existing models on existing hardware. The free parameters are experimental design choices (memory cap, system size) rather than fitted model parameters. The axioms are domain assumptions about benchmarking methodology, not new theoretical postulates.

free parameters (3)

80 GB memory cap = 80 GB
Chosen to mimic ordinary lab hardware; affects all scalability numbers. Not fitted to data but is an arbitrary constraint that shapes results.
192-atom system size = 192 atoms (LiCoO₂)
Fixed system for throughput benchmarking; chosen as a common test case but affects throughput rankings.
6–7 Å cutoff (inherited from MLIP architectures) = 6–7 Å
Not a parameter of this paper but of the benchmarked models; drives the density-dependent scalability finding.

axioms (4)

domain assumption ASE-based pipeline is sufficient for reliable large-scale benchmarking (no framework bias)
§2: Cross-validated on TorchSim with 1.11–1.4× speedup. The assumption that this small difference does not reorder model rankings is unstated.
domain assumption κ_SRME from Matbench Discovery is a valid primary accuracy coordinate for MLIPs
§2: The paper adopts this metric from an external benchmark rather than independently computing it.
standard math DFT cost scales as O(N³) while MLIPs scale close to linearly
§3.2: Used to justify DFT-normalized speedup interpretation. Standard result from computational physics.
domain assumption Default official configurations and public pre-trained weights represent each model's intended performance
§2: Models tested with defaults. If some models have under-tuned public releases, rankings may not reflect their best possible performance.

pith-pipeline@v1.1.0-glm · 11800 in / 2894 out tokens · 448209 ms · 2026-07-09T03:43:12.608620+00:00 · methodology

0 comments

read the original abstract

Most MLIP benchmarks reward static accuracy while ignoring inference efficiency and hardware scalability -- driving model bloat with unclear real-world value. We benchmark 23 mainstream open-source MLIPs on a low-cost NVIDIA DGX Spark (128 GB native memory, capped at 80 GB to mimic ordinary lab hardware), using a fixed 192-atom system under a unified ASE-based pipeline. We evaluate three dimensions: predictive accuracy, MD simulation throughput, and atomic scalability. Our results expose a sharp accuracy-efficiency trade-off: large SOTA models deliver only 3-5 meV/atom more accuracy than lightweight ones, but lose orders of magnitude in throughput -- in the worst case, becoming only marginally faster than DFT itself. Lightweight MLIPs, by contrast, sit on the Pareto frontier and run on modest hardware. The lesson is that single-dimensional benchmarks mislead the field, and that future MLIP development should value efficiency and scalability alongside accuracy.

Figures

Figures reproduced from arXiv: 2607.07647 by Hanwen Kang, Miao Liu, Sheng Meng, Tenglong Lu.

**Figure 1.** Figure 1: Accuracy–efficiency Pareto trade-off across the 23 benchmarked MLIPs. Inference throughput (x-axis, normalized MD steps per second) is plotted against prediction accuracy (y-axis, normalized thermal-conductivity MAE). Accuracy and computational efficiency constitute a tradeoff. The headline result is shown in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: MD Simulation throughput across multiple mainstream MLIPs, normalized to CPU-based DFT (1×). Classical empirical potentials lead: the Buckingham potential achieves ~11,600× DFT speedup, but at the cost of chemical fidelity. Lightweight MLIPs follow, delivering hundreds-tothousands-fold speedups — the speed tier at which the original MLIP promise is realized. Mediumscale MLIPs lose roughly half an order o… view at source ↗

**Figure 3.** Figure 3: (a) GPU memory consumption versus simulated atom count, with linear fits extrapolated to the unified 80 GB memory cap.(b) Relationship between maximum atom count and model parameters. Per-atom memory footprints differ by orders of magnitude across architectures. Lightweight MLIPs scale to roughly 200,000 atoms; medium-tier models (e.g., M3GNet) handle 10⁴– 10⁵ atoms; heavyweight models — DPA4.0 Pro, Equifo… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages · 3 internal anchors

[1]

Accuracy–efficiency Pareto trade-off across the 23 benchmarked MLIPs

Results 3.1 Accuracy–Efficiency Pareto Trade-off Figure 1. Accuracy–efficiency Pareto trade-off across the 23 benchmarked MLIPs. Inference throughput (x-axis, normalized MD steps per second) is plotted against prediction accuracy (y-axis, normalized thermal-conductivity MAE). Accuracy and computational efficiency constitute a trade-off. The headline resul...

work page
[2]

EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers

Conclusions We benchmarked 23 mainstream MLIPs across accuracy, throughput, and scalability, on standardized hardware with a conservative memory budget. The headline is unflattering: largest does not mean best. Bloated architectures deliver a few meV/atom of extra accuracy while surrendering orders of magnitude in throughput and the ability to simulate re...

work page internal anchor Pith review Pith/arXiv arXiv 1964
[3]

B. Deng, P. Zhong, K. Jun, J. Riebesell, K. Han, C. J. Bartel, and G. Ceder, CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling, Nat Mach Intell 5, 1031 (2023). [14] H. Yang et al., MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures, arXiv:2405.04967. [15] Y. Lysogorskiy, A...

work page internal anchor Pith review Pith/arXiv arXiv 2023
[4]

Orb-v3: atomistic simulation at scale

B. Rhodes, S. Vandenhaute, V. Šimkus, J. Gin, J. Godwin, T. Duignan, and M. Neumann, Orb-v3: Atomistic Simulation at Scale, arXiv:2504.06231. [31] T. Koker, A. Gangan, M. Kotak, J. Marian, and T. Smidt, PFT: Phonon Fine-Tuning for Machine Learned Interatomic Potentials, arXiv:2601.07742. [32] T. Koker, M. Kotak, and T. Smidt, Training a Foundation Model f...

work page internal anchor Pith review Pith/arXiv arXiv 2023

[1] [1]

Accuracy–efficiency Pareto trade-off across the 23 benchmarked MLIPs

Results 3.1 Accuracy–Efficiency Pareto Trade-off Figure 1. Accuracy–efficiency Pareto trade-off across the 23 benchmarked MLIPs. Inference throughput (x-axis, normalized MD steps per second) is plotted against prediction accuracy (y-axis, normalized thermal-conductivity MAE). Accuracy and computational efficiency constitute a trade-off. The headline resul...

work page

[2] [2]

EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers

Conclusions We benchmarked 23 mainstream MLIPs across accuracy, throughput, and scalability, on standardized hardware with a conservative memory budget. The headline is unflattering: largest does not mean best. Bloated architectures deliver a few meV/atom of extra accuracy while surrendering orders of magnitude in throughput and the ability to simulate re...

work page internal anchor Pith review Pith/arXiv arXiv 1964

[3] [3]

B. Deng, P. Zhong, K. Jun, J. Riebesell, K. Han, C. J. Bartel, and G. Ceder, CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling, Nat Mach Intell 5, 1031 (2023). [14] H. Yang et al., MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures, arXiv:2405.04967. [15] Y. Lysogorskiy, A...

work page internal anchor Pith review Pith/arXiv arXiv 2023

[4] [4]

Orb-v3: atomistic simulation at scale

B. Rhodes, S. Vandenhaute, V. Šimkus, J. Gin, J. Godwin, T. Duignan, and M. Neumann, Orb-v3: Atomistic Simulation at Scale, arXiv:2504.06231. [31] T. Koker, A. Gangan, M. Kotak, J. Marian, and T. Smidt, PFT: Phonon Fine-Tuning for Machine Learned Interatomic Potentials, arXiv:2601.07742. [32] T. Koker, M. Kotak, and T. Smidt, Training a Foundation Model f...

work page internal anchor Pith review Pith/arXiv arXiv 2023