{"id":"d1ea09e8-8c40-4047-b46c-e5f930f8e652","arxiv_id":"2505.06711","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A layer-by-layer ghost-atom message synchronization scheme makes message passing neural network potentials scale linearly with message-passing depth, enabling molecular dynamics with over 100 million atoms.","lead":"This paper shows how to run fast molecular simulations with a class of machine learning potentials that previously were hard to parallelize. The new method keeps communication local in each message-passing layer and reaches simulations with more than 100 million atoms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"O(T) claim is sound for one-hop MPNNs, but the 'general framework' assertion is broader than the single REANN demonstration; global or nonlocal MPNNs break the ghost-feature-exchange assumption.","rationale":"The paper's central algorithmic claim is internally consistent, and the REANN-MPI demonstration is credible: the source code is available, the RDF validation in Figure S1 checks correctness, and the strong and weak scaling plots support the implementation. My read does not overturn the reader's conditional verdict. The most load-bearing unresolved issue is the breadth of the 'general' claim: the O(T) analysis is proven only for message functions that are one-hop-local under a fixed cutoff, and the paper supplies only one architecture as evidence. This is addressable by porting the exchange scheme to a second MPNN and measuring both scaling and numerical agreement. The literature-based speed comparisons are a secondary weakness but do not bear on the core scaling argument as directly.","tokens_in":16948,"tokens_out":15646,"duration_ms":176707,"concrete_test":"Port the ghost-feature MPI exchange to a second, non-REANN one-hop MPNN (e.g., SchNet or NequIP) in LAMMPS, and run the same 1M-atom water benchmark with T = 1 through 4 on 2, 8, and 32 MPI ranks. Check (i) wall-time per MD step grows approximately linearly in T, and (ii) energies and forces agree with a single-process reference to numerical precision. If (ii) fails or (i) is superlinear, the 'general framework' claim must be restricted to one-hop models; additionally, repeat with a global-attention layer to map the boundary of applicability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core argument in Section 2.2 is correct for the REANN architecture: since each MP layer aggregates only one-hop neighbor features within the original r_c, distributing the computation and exchanging only ghost-atom feature tensors per layer gives per-layer cost proportional to the number of edges and communication proportional to the halo, hence total cost O(T). The weak spot is the jump from this REANN-specific implementation to the 'general parallelization framework' asserted in the abstract, Section 2.2, and conclusion. The O(T) communication pattern depends on every update being a local function of one-hop neighbor features from the previous layer; it is not applicable to MPNN potentials with global attention, global normalization, long-range electrostatics, or charge equilibration, because those introduce couplings that are not representable as ghost-feature halo exchanges. The paper's own scope clause ('with messages aggregated over neighboring atoms') excludes such models, but the title and abstract do not carry this caveat. No second MPNN architecture is implemented or benchmarked, so the generality claim currently rests on argument rather than evidence. Secondary, the speed comparisons with Allegro, SevenNet, and MACE use literature timings from different hardware and software versions, so the reported speed advantage is not cleanly attributable to the parallelization scheme.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a parallelization algorithm for message passing neural network (MPNN) potentials used in molecular dynamics. The central idea is that in each message-passing layer, only ghost-atom feature tensors need to be exchanged between processes, so that the cost and communication overhead scale linearly with the number of layers T rather than with the effective cutoff volume. The authors implement this with the REANN model (REANN-MPI) and report strong and weak scaling results on CPU and GPU benchmarks, including a simulation of roughly 100 million Ag atoms. They also compare performance against Allegro, SevenNet, MACE-MP-0, and Deep Potential, and argue that the linear-in-T scaling overturns earlier estimates of O(T^3)-O(T^4) cost for MPNNs.","tokens_in":17147,"tokens_out":9401,"duration_ms":98388,"significance":"If the algorithmic claims hold, the contribution is valuable: it demonstrates that standard one-hop MPNN potentials can be parallelized with halo-only communication, making large-scale molecular dynamics with MPNNs as tractable as with strictly local potentials. The conceptual clarification that the growth of the effective receptive field does not require explicit enumeration of a T*r_c neighbor list is correct and important for the community. The reverse-mode differentiation on a partitioned graph with communicated vector-Jacobian products is a sound foundation, and the authors provide their implementation as open-source code with benchmark data, which is a strength. However, the paper over-claims generality beyond the tested REANN architecture, and the absolute speed comparisons with other models are not controlled, which tempers the significance of the reported speed advantages.","major_comments":[{"comment":"The abstract and conclusion claim that this is a 'general parallelization framework' that can 'empower various MPNN models,' but the algorithm relies on the assumption that each MP layer aggregates only one-hop neighbor features within the original cutoff. Models with global attention (e.g., SO3KRATES), long-range electrostatics, or charge equilibration (e.g., PhysNet, CENT) violate this assumption because their updates depend on atoms outside the ghost halo. The paper's own scope clause 'messages aggregated over neighboring atoms' covers only a subset of MPNN potentials. Please either implement the scheme on a second non-REANN MPNN architecture (e.g., SchNet or NequIP) to support the generality claim, or explicitly qualify the title, abstract, and conclusion to state that the framework applies to local one-hop message-passing architectures.","section":"Abstract, Section 2.2"},{"comment":"The claimed speed advantages in Figure 7 rest on uncontrolled comparisons: the Allegro and MACE-MP-0 timings are taken from Refs. 47 and 54, which used different hardware, library versions, and implementation details. Thus the 'two orders of magnitude faster' statement is not a valid measure of the parallelization efficiency and could be dominated by differences in model architecture (REANN is a relatively small invariant model). To support the claim that the parallel algorithm makes MPNNs as fast as strictly local models, the authors should measure all models on the same hardware with the same software stack, or present the comparison with explicit caveats and error margins.","section":"Section 2.3, Figure 7"},{"comment":"The text states that 'Figure 5 offers ambiguous numerical evidence that a properly parallelized MPNN model scales linearly with T.' This appears to be a typo for 'unambiguous,' but as written it directly contradicts the subsequent sentence and undermines the central scaling claim. Please correct this and ensure the figure caption and text consistently describe the evidence as supporting linear scaling.","section":"Section 2.3, Figure 5"}],"minor_comments":[{"comment":"The exact number of atoms in the claimed 'over 100 million' simulation is not specified; 32 nodes with 'roughly 3 million atoms per node' gives about 96 million. Please provide the exact atom count (e.g., 100,663,296) and the precise system size in the text or figure.","section":"Section 2.3"},{"comment":"The notation for the backward pass is terse and some terms (e.g., the dimensions of ∂ρ_i^t/∂c_j^{t-1} and the exact communication patterns for a generic MPNN) are not defined. Please clarify the tensor shapes and state explicitly which vector-Jacobian products are exchanged between which processes in each layer.","section":"Section 2.2, Eqs. (3)-(5)"},{"comment":"The comparison with Deep Potential (Ref. 41) uses different core counts (768 vs 3360 cores). Reporting per-core efficiency or a normalized metric would make the comparison more interpretable than raw times per step per atom.","section":"Section 2.3, Figure 6 and text"},{"comment":"The distinction between 'false impression' and 'actual growth' is clear conceptually, but the caption could be more explicit: in (b), the number of contributing paths grows, but the computational graph only needs the original neighbor list. Consider adding one sentence to state this directly.","section":"Figure 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The core parallelization algorithm is correct for local one-hop MPNNs and the scaling measurements for REANN are convincing. The main concern is that the paper's title and abstract promise a general framework without providing evidence beyond one architecture, and the speed comparisons with Allegro/MACE use literature timings that are not directly comparable. Both issues are fixable by either adding a second implementation or tempering the claims. I recommend major revision rather than rejection because the central idea is sound and the contribution would be useful after proper scoping."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — read this if you care about whether MPNN potentials can be pushed to 100M atoms. The core algorithmic idea is real: instead of treating each MP layer as expanding the effective cutoff and recomputing neighbor lists, exchange only the ghost-atom feature tensors that owners computed in the previous layer. That turns the per-layer communication into a halo exchange and makes total cost linear in T. Reverse-mode forces via communicated vector-Jacobian products are the right way to handle gradients without shipping whole subgraphs. I checked the computational graph reasoning around Fig. 4 and Eqs. 3–5; the pieces fit. This is a legitimate correction to the T^3/T^4 scaling story in the Allegro and SevenNet literature.\n\nThe implementation is also a plus: REANN-MPI is in C++/LibTorch, MPI, LAMMPS interface, code available. Strong scaling to 1000 CPU cores at ~80%, weak scaling flat to 32 nodes, and 100M-atom MD are useful concrete benchmarks. The RDF check against REANN-JIT confirms energies and forces are unchanged, which addresses the worry that communication breaks correctness. Credit where due: this is the first scheme I know that makes the per-layer ghost-feature exchange explicit for MPNN potentials.\n\nSoft spots, in proportion. First, the general-framework claim is broader than the evidence. Section 2.2 says any MPNN with messages aggregated over neighboring atoms can adopt the scheme; that is true for one-hop local message passing of the REANN/SchNet/NequIP/MACE type, but not for global attention, global normalization, long-range electrostatics, or charge equilibration. Those models have couplings that no finite halo exchange can represent. The paper's own “messages aggregated over neighboring atoms” clause is doing quiet work that the title, abstract, and conclusion do not carry. A second demonstration architecture, or a clear statement of the local-message-passing restriction in the abstract, would fix this.\n\nSecond, the speed comparisons against Allegro, SevenNet, and MACE are from literature timings on different software versions and hardware. They are qualitative, not apples-to-apples. That does not undermine the scaling curves, but it means I would not cite the “two orders of magnitude faster than Allegro” number without checking the original setups.\n\nMinor: Figure 5 says “ambiguous numerical evidence” — obviously meant “unambiguous”; there is also an “Allgero” typo in the intro. Both trivial.\n\nWho gets value: anyone implementing MPNN potentials in LAMMPS or similar, especially at GPU cluster scale. The paper deserves a serious referee; the O(T) argument and the REANN-MPI implementation are concrete enough to be verified. I would send it out, and I would ask the authors to scope the generality claim and to report one timing set where the comparison models run on the same hardware.","headline":"A genuinely useful parallelization scheme for one-hop MPNN potentials, with an O(T) argument that holds for REANN-class models; the 'any MPNN' framing runs ahead of the evidence.","tokens_in":17667,"tokens_out":2752,"would_cite":true,"duration_ms":27929,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Message-passing neural network potentials scale linearly with layers, not cubically, enabling 100-million-atom molecular dynamics.","keywords":["message passing neural network potential","molecular dynamics","parallelization","linear scaling","recursively embedded atom neural network","spatial decomposition","ghost atoms","machine learning interatomic potential"],"falsifier":"Take an MPNN whose layer update contains a nonlocal operation—such as global softmax attention, an all-atom normalization constant, or an electrostatic term summed over the entire box—and run it under this spatial-decomposition scheme. If the result is either wrong energies at subdomain boundaries or communication volume that grows with system size rather than with surface area times $T$, the linear-scaling claim does not generalize to that model. The test is concrete: instrument per-layer communication volume at increasing box sizes.","tokens_in":16723,"feed_emoji":"⚛️","tokens_out":6224,"duration_ms":63048,"temperature":0.7,"pith_summary":"The paper sets out to overturn a standard objection to message-passing neural network (MPNN) potentials: that their parallel cost grows steeply with the number of message-passing layers, effectively as $T^3$ or $T^4$. It argues that, because each layer updates an atom from its immediate one-hop neighbors within the original cutoff, a spatially decomposed simulation only needs to exchange ghost-atom feature tensors once per layer, so total cost grows linearly with $T$. This claim matters because MPNNs are among the most accurate machine-learned potentials but have been considered impractical for very large molecular dynamics. Using its recursively embedded atom neural network, the accompanying implementation demonstrates near-linear strong and weak scaling and reports simulation of more than 100 million atoms at speeds comparable to strictly local models. The paper presents this as a general framework that other MPNN architectures can adopt.","feed_headline":"MPNN potentials now scale to 100 million atoms","feed_subtitle":"A ghost-message scheme makes message-passing molecular dynamics match strictly local model speed.","key_machinery":"The central object is the ghost-atom message exchange across subdomain boundaries in each message-passing layer. A process stores its local atoms plus ghost atoms (atoms owned by other processes that sit inside the cutoff), and each layer it receives the already-computed feature tensors of those ghosts from their owning processes rather than recomputing them. This keeps each computational subgraph intact without redundant work; gradients are handled symmetrically by exchanging only the derivative components that involve neighbor processes, using vector–Jacobian products rather than full Jacobians.","core_discovery":"On the paper's own terms, the discovery is that the previously claimed $\\mathcal{O}(T^3)$—or cumulatively $\\mathcal{O}(T^4)$—growth of MPNN cost with layer number is an artifact of treating message passing as if each layer rebuilt a growing neighbor list. In practice, at each layer an atom's new feature is built from features of the original neighbors inside $r_c$, so one neighbor list suffices for every layer. In parallel spatial decomposition, each process computes features for its local atoms, sends the ghost-atom subset of those features to neighboring processes, and proceeds; no process recomputes a ghost atom's features from an incomplete neighborhood. Backward force evaluation is decomposed through the same chain rule, with vector–Jacobian products limiting inter-process gradient traffic. Integrated with the REANN architecture, this yields linear scaling with $T$, strong scaling to 1,000 CPU cores and 8 GPUs, and a 100-million-atom silver simulation whose per-atom speed matches strictly local models.","pith_inferences":["A direct test of the generality claim would implement this ghost-message scheme in a second, independent MPNN architecture (for example, an equivariant model) and check whether linear-in-$T$ scaling persists at the same per-atom cost; the current evidence is on one architecture.","The speed advantage reported against strictly local Allegro likely reflects both the parallel scheme and REANN's cheaper invariant features; separating those effects would require the same algorithm wrapped around Allegro's architecture.","Models with genuine long-range physics, such as electrostatics or global attention, will need an additional communication layer on top of ghost messages; the scheme as stated applies to one-hop message passing only.","Direct GPU-to-GPU communication, listed by the authors as future work, would likely push the strong-scaling efficiency above the ~67–88% range observed on 2–8 GPUs."],"forward_implications":["An MPNN with many message-passing layers is no longer inherently harder to parallelize; each added layer raises cost only linearly, so depth can be used for accuracy without sacrificing scale.","Invariant and equivariant MPNNs that aggregate messages from neighboring atoms—SchNet-like, NequIP-like, and MACE-like architectures—can adopt the same communication pattern without structural changes.","MPNN-based molecular dynamics can reach system sizes previously reserved for strictly local potentials, as shown by the over-100-million-atom silver simulation.","By using vector–Jacobian products rather than full Jacobians, per-process gradient tensors shrink by a factor related to the feature-vector size, cutting both compute and communication in the backward pass.","The parallelization is compatible with local models as well: when $T=1$, the new implementation performs comparably to existing just-in-time parallelized REANN, showing negligible added overhead."],"supporting_citations":[{"why":"Defines the recursively embedded atom neural network used in all benchmarks and demonstrations.","marker":"[17]"},{"why":"Supplies the Allegro strictly local model, the parallel-scaling argument being overturned, and comparison timing data for silver.","marker":"[47]"},{"why":"Gives the SevenNet spatial decomposition algorithm the paper compares against in the high-entropy alloy benchmark.","marker":"[48]"},{"why":"Provides the 100-million-atom Deep Potential simulation baseline and per-atom timings used for the efficiency comparison.","marker":"[41]"},{"why":"Provides the embedded atom neural network local descriptors from which REANN's feature construction is derived.","marker":"[49]"},{"why":"Supplies the MACE-MP-0 foundation model timing data used in the multi-GPU comparison.","marker":"[54]"}],"fun_headline_variants":["MPNN MD at 100M atoms: now as fast as local models","Parallel MPNN breaks 100M-atom barrier at local speed","Linear-scaling MPNN: 100M atoms, local-model speed","Ghost-message MPNN: 100M atom MD, linear with layers","MPNN potentials reach 100M atoms at local-model pace"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme assumes that at every layer, an atom's updated feature depends only on features of its original one-hop neighbors inside the cutoff; if an MPNN uses global normalization, long-range attention, or any nonlocal term beyond that neighborhood, ghost-atom messages alone will not be enough.","fun_headline_variants_meta":{"raw":{"variants":["MPNN MD at 100M atoms: now as fast as local models","Parallel MPNN breaks 100M-atom barrier at local speed","Linear-scaling MPNN: 100M atoms, local-model speed","Ghost-message MPNN: 100M atom MD, linear with layers","MPNN potentials reach 100M atoms at local-model pace"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2839,"prompt_tokens":916,"completion_tokens":1923,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":1828}},"tokens_in":532,"tokens_out":1923,"duration_ms":12668,"temperature":1.0,"reasoning_tokens":1828,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:35:38.055936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an MPNN whose layer update contains a nonlocal operation—such as global softmax attention, an all-atom normalization constant, or an electrostatic term summed over the entire box—and run it under this spatial-decomposition scheme. If the result is either wrong energies at subdomain boundaries or communication volume that grows with system size rather than with surface area times $T$, the linear-scaling claim does not generalize to that model. The test is concrete: instrument per-layer communication volume at increasing box sizes.","supporting_citations":[],"review_version":1}