TorchSim: An efficient atomistic simulation engine in PyTorch

Abhijeet S. Gangan; Adeesh Kolluru; Gerbrand Ceder; Janosh Riebesell; Jorge Colindres; Joseph Krause; Orion Cohen; Rhys Goodall; Stefano Falletta

REVIEW 3 major objections 2 minor 1 references

TorchSim: An efficient atomistic simulation engine in PyTorch

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read TorchSim's central claim is that rewriting atomistic simulation primitives as batched PyTorch operations makes machine-learned interatomic-potential simulations orders of magnitude faster on GPUs while keeping them differentiable.

desk verdict The abstract describes a plausible and useful PyTorch MD engine, but the full text in the packet is an unrelated paper, so there is no evidence to support the headline 'orders of magnitude' claim. read the letter →

arxiv 2508.06628 v1 pith:XSROWM6L submitted 2025-08-08 physics.comp-ph cond-mat.mtrl-sci

Orion Cohen , Janosh Riebesell , Rhys Goodall , Adeesh Kolluru , Stefano Falletta , Joseph Krause , Jorge Colindres , Gerbrand Ceder

show 1 more author

Abhijeet S. Gangan

This is my paper · ORCID

classification physics.comp-phcond-mat.mtrl-sci

keywords atomisticsimulationmachine-learnedinteratomicpotentialsPyTorchGPUbatchingmoleculardynamicsdifferentiablestructuralrelaxationhigh-throughputscreening

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

TorchSim is an open-source atomistic simulation engine built around PyTorch, aimed at the machine-learned interatomic potential (MLIP) era. The paper's central claim is that conventional molecular dynamics packages underuse modern GPUs because they evolve one system at a time, whereas TorchSim rewrites the core simulation primitives as batched tensor operations and evolves many systems concurrently, yielding 'orders of magnitude' speedups for popular MLIPs. It further claims support for molecular dynamics integrators, structural relaxation, machine-learned and classical potentials (Lennard-Jones, Morse, soft-sphere), automatic memory management, differentiable simulation, and interoperability with materials-informatics tools. A sympathetic reader would care because this targets the practical bottleneck of MLIP-based simulation—the surrounding code, not just the neural network—and would make ensemble and high-throughput simulations much cheaper. Caveat: the supplied full text is an unrelated pancreatic-cancer EHR manuscript, so only the abstract supports these claims in this submission.

What carries the argument

The central object is the PyTorch-native batched simulation primitive: every step of the molecular dynamics loop is written as tensor operations over a batch of independent systems, with neighbor lists and forces computed in a vectorized way and memory allocated automatically. This batching is what lets one GPU work on many systems at once instead of idling while a single system's CPU-side loops run, and the use of autodiff makes the whole trajectory differentiable.

What would settle it

Run a published machine-learned interatomic potential on a single molecular system with a standard molecular dynamics package and with TorchSim on the same GPU, comparing wall-clock time per step, energy conservation, and force error at the same batch size. If TorchSim is not orders of magnitude faster, or if its batched trajectories do not conserve energy to the same tolerance as the reference, the central acceleration claim fails.

Watch

Extended reading notes

Core claim

TorchSim's claim is that the standard architecture of atomistic simulation packages—one system per process, CPU-bound neighbor and integration loops, no GPU batching—is the main obstacle to using machine-learned potentials at scale. By expressing the whole simulation loop (neighbor lists, force evaluation, integrators, relaxers) as differentiable PyTorch operations and by packing multiple independent systems into a single GPU batch with automatic memory management, TorchSim says it can accelerate popular MLIPs by orders of magnitude. The paper also presents differentiability as a first-class feature: because the entire trajectory is built from tensor ops, gradients can flow back through the

Load-bearing premise

The load-bearing premise is that running many independent molecular systems concurrently on one GPU, with automatic memory management and automatic differentiation, is stable and accurate while delivering the claimed orders-of-magnitude speedup; no benchmark accompanies the abstract, and the supplied full text is an unrelated pancreatic-cancer EHR manuscript, so the premise is unverified in this submission.

Editorial extensions

If this is right

If the acceleration claim holds, MLIP molecular dynamics on a single GPU moves from one-system-per-run to ensemble simulation, cutting wall-clock time for statistical sampling and free-energy calculations.
Differentiable simulation makes it possible to define loss functions on whole trajectories and backpropagate into potential parameters, opening direct gradient-based training and inverse design.
Batched relaxation and MD can turn high-throughput screening workflows into a single GPU call over thousands of candidate structures.
Because classical potentials are included in the same framework, users can mix or compare classical and machine-learned models without a separate codebase.

Reading between the lines

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

If the orders-of-magnitude claim survives benchmarking, the slow step in MLIP workflows will shift from force evaluation to data generation and model accuracy, making replica-exchange and ensemble methods the default on ordinary GPU workstations.
The same batched, differentiable design could plausibly extend to multi-GPU and mixed-precision training of interatomic potentials, since the autodiff path already connects simulation outputs to parameter gradients.
Verification caveat (editorial): the full text supplied for this paper is a different manuscript—an early-detection study for pancreatic cancer using electronic health records—so none of TorchSim's implementation details, benchmarks, or code can be checked here; the claims rest on the abstract alone.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the short version: I can't review what isn't there. The abstract is about TorchSim, a PyTorch atomistic simulation engine with batched concurrent systems. The full text in the packet is a pancreatic-cancer EHR paper. So the actual software, benchmarks, and code are not in the record. Everything I say about the substance comes from the abstract.

Now, the abstract does make sense. TorchSim fits a real need: most MLIP workflows still run one simulation per process, and batched multi-system execution on GPUs could genuinely speed up screening and make differentiation across systems cheaper. The automatic memory management and integration with materials informatics tools are sensible features. If the implementation delivers what the abstract says, it's a legitimate contribution that improves on single-system loops in a way that matters for high-throughput applications.

The problem is the evidence. 'Orders of magnitude acceleration' is an empirical claim, and the record contains no benchmark numbers, no baseline (LAMMPS, ASE, JAX-MD, or whatever), no hardware description, and no accuracy checks. That's not a minor omission; it's the central claim. The engineering premise — that batched autodiff simulation is stable and faster without corrupting accuracy — is exactly the kind of thing that depends on details like GPU occupancy, memory fragmentation, and integrator behavior. Without those details, there's no way to tell whether 'orders of magnitude' is marketing or fact.

There's also no prior art discussion in the abstract, so I can't tell how much of the batching framing is new relative to JAX-MD or other differentiable MD engines. The authors may cover this in the real paper; but the abstract alone doesn't.

This is not an internal contradiction. The design could be entirely sound. But as a record for review, it's insufficient. My recommendation: get the real manuscript. If it has the promised benchmarks and code, send it out to a referee who knows the MD and MLIP landscape. If it doesn't, or if the 'orders of magnitude' claim evaporates under measurement, that's a desk reject or a request for revision. I'd neither accept nor reject on the current packet.

Referee Report

3 major / 2 minor

Summary. The submission, arXiv:2508.06628, presents an abstract for 'TorchSim: An efficient atomistic simulation engine in PyTorch.' The abstract claims that TorchSim is an open-source engine that rewrites core atomistic simulation primitives in PyTorch, achieves 'orders of magnitude acceleration' for popular machine-learned interatomic potentials (MLIPs), performs batched simulations on GPUs, supports molecular dynamics integrators, structural relaxation optimizers, classical and ML potentials, batching with automatic memory management, differentiable simulation, and integration with materials informatics tools. However, the full text supplied with the submission is not the TorchSim paper; it is an unrelated manuscript on early detection of pancreatic cancer using electronic health records (arXiv:2508.06627). Consequently, the submission contains no technical description of TorchSim, no implementation details, no benchmark results, no accuracy checks, and no code or documentation beyond the abstract.

Significance. If the abstract's claims were substantiated, TorchSim could be a practically valuable contribution to atomistic simulation in the MLIP era, offering GPU-accelerated batched simulations and differentiable dynamics that are currently missing from many standard packages. The claimed 'orders of magnitude acceleration' and the batching-with-autodiff design would be directly useful to the computational materials community. However, as submitted, none of these claims can be evaluated. The central performance assertion is an empirical claim that requires a benchmark protocol, baseline definitions, hardware details, and error bars; none are provided. The absence of the actual manuscript text makes the significance unassessable: the contribution may be real and important, but the submitted record does not demonstrate it.

major comments (3)

[Full Text] The supplied full text is not the paper advertised in the abstract. It is an entirely different manuscript on pancreatic cancer detection from electronic health records (arXiv:2508.06627). This is a load-bearing defect: the review record contains no TorchSim content beyond the abstract, so the core technical claims cannot be checked. The manuscript as submitted is not a coherent paper about TorchSim and cannot be revised into one without resubmitting the correct text.
[Abstract] The headline claim of 'orders of magnitude acceleration for popular MLIPs' is unsupported by any benchmark data, baseline definition, hardware specification, or measurement protocol. There is no comparison against standard packages (e.g., LAMMPS or ASE on CPU/GPU) nor against a single-system PyTorch loop. The abstract gives no error bars or accuracy checks, so the central numerical claim is not verifiable even from the abstract alone.
[Abstract] The abstract advertises 'batched simulations that efficiently utilize modern GPUs' and 'automatic memory management' as key features. These engineering claims are load-bearing for the performance assertion, yet no evidence is provided that batched multi-system evolution with PyTorch autodiff is numerically stable, memory-efficient, or faster than one-system-at-a-time in practice. Integrator accuracy and relaxation convergence are not discussed. This missing support is not a minor omission; it is the core of the proposed contribution.

minor comments (2)

[Abstract] The abstract states that TorchSim is 'open-source' but does not provide a repository URL or version identifier. If the correct manuscript is resubmitted, a code link should be included.
[Abstract] The abstract lists many features (integrators, relaxers, potentials, batching, autodiff, integration with tools) without any technical detail or reference to prior work. A proper manuscript would need to specify the design choices, data structures, and algorithmic contributions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation present in supplied record; full text is unrelated paper, so acceleration claim is unchecked but not circular.

full rationale

The supplied record for arXiv:2508.06628 consists only of an abstract claiming order-of-magnitude acceleration for MLIPs via PyTorch batching and concurrent multi-system GPU evolution. The 'full text' supplied is actually arXiv:2508.06627, a pancreatic-cancer EHR paper, so there is no TorchSim derivation chain, equation, fitted parameter, or self-citation to analyze. The acceleration claim is an empirical engineering assertion requiring benchmarks and baselines; its absence is an evidence gap, not a circularity. No quantity is defined in terms of a target result, no fitted input is relabeled as a prediction, and no self-citation is load-bearing. Accordingly, the paper shows no circularity in the provided record.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No fitted parameters or invented entities are visible from the abstract alone, so the ledger is nearly empty. The load-bearing assumptions are engineering premises about GPU batching and PyTorch autodiff performance, none demonstrated in the supplied text. This is consistent with an abstract-only review of a software paper whose body is unavailable.

assumptions (3)

domain assumption Batched concurrent simulation of many systems on one GPU is efficient and stable for MLIP workloads
The abstract's 'orders of magnitude acceleration' claim rests on this engineering premise; no measurement is provided in the available text.
domain assumption PyTorch-based primitives reach or exceed the performance of compiled MD codes
Implied by the claim that rewriting primitives in PyTorch yields accelerations; autodiff and memory overhead are not discussed.
standard math Standard MD integrators and relaxation optimizers remain accurate when run in batched or differentiable form
Supports the claim of support for MD integrators and optimizers; no conservation or stability tests are visible.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TorchSim: An efficient atomistic simulation engine in PyTorch." pith.science (2026). https://pith.science/paper/XSROWM6L

@misc{pith2026250806628,
  author       = {Pith},
  title        = {Pith review of: TorchSim: An efficient atomistic simulation engine in PyTorch},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XSROWM6L}},
  note         = {Machine review of arXiv:2508.06628}
}

read the original abstract

We introduce TorchSim, an open-source atomistic simulation engine tailored for the Machine Learned Interatomic Potential (MLIP) era. By rewriting core atomistic simulation primitives in PyTorch, TorchSim can achieve orders of magnitude acceleration for popular MLIPs. Unlike existing molecular dynamics packages, which simulate one system at a time, TorchSim performs batched simulations that efficiently utilize modern GPUs by evolving multiple systems concurrently. TorchSim supports molecular dynamics integrators, structural relaxation optimizers, both machine-learned and classical interatomic potentials (such as Lennard-Jones, Morse, soft-sphere), batching with automatic memory management, differentiable simulation, and integration with popular materials informatics tools.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages

[1]

�� Early Detection of Pancreatic Cancer Using Multimodal Learning on Electronic Health Records Mosbah Aouad maouad2@illinois.edu University of Illinois Urbana-Champaign Anirudh Choudhary ac67@illinois.edu University of Illinois Urbana-Champaign Awais F arooq afaroo32@uic...

work page Pith review arXiv 2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

�� Early Detection of Pancreatic Cancer Using Multimodal Learning on Electronic Health Records Mosbah Aouad maouad2@illinois.edu University of Illinois Urbana-Champaign Anirudh Choudhary ac67@illinois.edu University of Illinois Urbana-Champaign Awais F arooq afaroo32@uic...

work page Pith review arXiv 2025