{"id":"d370be24-31d1-4242-b4a6-29929661c011","arxiv_id":"2610.23371","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":7,"one_line_summary":"A receiver-chunked three-pass schedule enables exact foundation-model supervision of million-atom MD, catching rare local failures and repairing them via on-the-fly cluster retraining.","lead":"This paper shows that an exact equivariant foundation model (MACE-MPA-0) can check every atom of million-atom molecular dynamics simulations, finding sparse local failures that averaged metrics and random audits miss. A reorganized three-pass inference schedule reduces memory per atom 340-fold, making whole-system teacher evaluations feasible on one GPU at 1.3% of the run time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"No significant objection identified: the paper's central claims are well-supported, limitations are honestly stated, and the repair demonstration is robust within its scope.","rationale":"The paper delivers a clean methodological advance with careful numerical verification and honest reporting of limitations. The three-pass schedule is exact, the memory reduction is real, and the detection/repair experiments are well-designed with multiple seeds at 12k atoms and a single demonstration at 1M atoms. The reader's verdict captures the strengths accurately. I find no additional load-bearing concern that would warrant changing the verdict. The million-atom seed replication is the obvious remaining check, but the authors already note this as a limitation.","tokens_in":47794,"tokens_out":2549,"duration_ms":30231,"concrete_test":"Re-run the million-atom SiO2 repair experiment (Section 3.3) with two additional random seeds of the initial velocities and random-number generator, while keeping the student initialization and training data fixed; report the spread of the final student force RMSE and flagged fraction on the independent teacher frames. This would quantify the reproducibility of the repair at scale, addressing the acknowledged limitation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that whole-system supervision with an exact foundation model detects and repairs local failures of distilled potentials is supported by extensive verification: the three-pass schedule is algebraically exact and verified numerically, cluster labels are tested for fidelity, and retraining is shown to reduce failure tails and pressure biases across multiple seeds and scales. The most significant limitation—that the teacher's own errors bound achievable student accuracy and that the million-atom repair run lacks seed replication—is explicitly acknowledged in the manuscript (Sections 3.3 and 4). No hidden assumption or internal inconsistency undermines the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper introduces a method for whole-system foundation-model supervision of distilled machine-learned interatomic potentials in million-atom molecular dynamics simulations. The key technical contribution is a receiver-chunked three-pass evaluation schedule for two-layer MACE models that discards per-edge intermediates, reducing the marginal memory per atom 340-fold and enabling exact evaluation of up to 2.86×10^7 atoms on a single H200 GPU. The authors use the exact MACE-MPA-0 teacher to check every atom of production trajectories driven by faster distilled students. They demonstrate that system-averaged error metrics and random audits miss sparse local failures (under-coordinated tails and isolated unphysical contacts) that whole-system checks reveal. On-the-fly retraining on teacher-labelled 13-Å clusters removes three quarters of the failure tail and fixes kilobar-level pressure biases in molten SiO2 and compressed water. The repair converges the student to the teacher's accuracy, not to DFT, and the approach incurs a supervision overhead of 1.3% of wall time at one check per 100 ps for 10^6 atoms.","tokens_in":47966,"tokens_out":1618,"duration_ms":17477,"significance":"If the results hold, this work provides a practical and principled framework for validating and repairing distilled potentials in production MD. The three-pass schedule is algebraically exact and numerically verified, and the memory reduction is dramatic, making exact foundation-model supervision feasible on a single GPU. The controlled experiments clearly separate what random audits miss, what cluster retraining fixes, and where the teacher's own DFT errors limit the repair. The paper is explicit about its limitations (single seed at 10^6 atoms, teacher-bounded accuracy, no stress in clusters, short-range students). This is a significant advance for large-scale MD with equivariant models.","major_comments":[],"minor_comments":[{"comment":"The million-atom repair run (1.03×10^6 atoms) uses only one seed. The authors acknowledge this in Section 4 as a limitation. While the three seeds at 12,288 atoms provide statistical replication for the smaller system, adding a second seed for the million-atom case (or stating a resource constraint more clearly) would strengthen confidence in the scalability of the repair.","section":"Section 3.3, Figure 4 and Table S14","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The analysis of transient vs. persistent tails is valuable, but the suggestion that 'criteria based on the persistence of flagged atoms over consecutive checks or on their absolute number would separate the two cases' is not implemented. Adding a simple demonstration (e.g., applying a persistence criterion to the water data) would make the recommendation more concrete.","section":"Section 3.4, Figure 5a-c","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation G[C ∪ S] += ∂e2/∂H^(1) could be clarified: is the gradient accumulated on both the receiver chunk and its senders? The text says 'back-propagates the first term of eq 2 into the pair forces and ∂e_2/∂H^(1) into G' and that G has size N×512. A short comment in the algorithm caption would help readers unfamiliar with adjoint accumulation.","section":"Section 2.1, Algorithm 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The cluster radius study (Table S7) shows that 13 Å clusters are sufficient for label fidelity. However, the text mentions that 15 Å clusters reduce the label deviation further but give the same retrained student accuracy. This result is important and could be moved to the main text, perhaps as a sentence in Section 3.1 or 3.3, to support the claim that the 13 Å choice is justified.","section":"Supplementary Note 4, Table S7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The throughput for SevenNet-0 in LAMMPS is given as 8.5×10^4 atom-steps s^-1 with a 23 GB peak. It would be helpful to state the number of atoms for this measurement (presumably 983,040 as for the other entries) and the chunk size used, for reproducibility.","section":"Section 3.6, Table 2","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This paper delivers what it promises: a way to run an exact equivariant foundation model on million-atom frames with one GPU, by reorganising two-layer MACE inference into three receiver-chunked passes that keep no per-edge intermediates resident.  The schedule is proved partition-invariant in the SI, verified numerically to machine precision, and the memory reduction (340-fold, to 4.57 kB per atom) is cleanly measured.  That alone is a useful piece of work.\n\nWhat is actually new: the combination of whole-system supervision of a distilled student with on-the-fly cluster retraining.  The empirical finding that a 1 % random audit misses an isolated six-atom O–O contact 94 % of the time is striking, and the controlled repair experiments (three seeds at 12 k atoms, one run at 1 M) show the method works.  The paper also honestly measures where it does not work: the teacher’s own error on under-coordinated Si bounds the student, repair converges to the teacher not to DFT, and the 1 M-atom repair demonstration lacks seed replication.  Code, data and models are public.\n\nSoft spots in proportion: (i) The million-atom repair run has no seed replication, which the authors acknowledge.  Fine.  (ii) The check-interval analysis shows that transient failures in water can be missed by sparse checks, but the paper does not propose a criterion to separate persistent from transient tails.  That is a development question, not a flaw.  (iii) The students’ short-range cutoff costs 1–2 % of the water density, and the pressure-bias in ambient water after cluster retraining is a side effect of missing stress labels.  Both are stated.  (iv) The throughput comparison with NEP89 is honest about what it compares (GPUMD on the same GPU), but the ratio will change as implementations improve.  Not a weakness of the paper.  Overall the central claims hold up, and the limitations are explicitly scoped.\n\nWho benefits: anyone building or using distilled potentials who wants to audit whole trajectories, and anyone developing memory-efficient inference for graph-based potentials.  This paper is not a breakthrough in accuracy—the repair ceiling is the teacher’s DFT error—but it is a clear methodological advance with rigorous validation.\n\nRecommendation: send to a serious referee.  The paper is well-argued, reproducible, and its claims are supported by the evidence presented.  Minor revisions could tighten the discussion of transient vs persistent failures, but the work is ready for peer review as is.","headline":"Solid engineering: the three-pass schedule is algebraically exact and verified, the failure-detection experiments are well-controlled, and the limitations are stated honestly.  Worth a serious referee.","tokens_in":48522,"tokens_out":1379,"duration_ms":15836},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"This paper shows that an exact equivariant foundation model can supervise every atom of million-atom molecular dynamics simulations, finding and repairing local failures of distilled potentials that system averages and small random…","keywords":["machine-learned interatomic potentials","foundation models","distillation","whole-system supervision","three-pass evaluation","molecular dynamics","MACE","on-the-fly retraining"],"falsifier":"Run a whole-system check on a 1,000,000-atom molten SiO₂ frame at 3000 K with a distilled student whose per-atom force deviation tail has been suppressed by the proposed cluster retraining, then compare the student's forces against DFT on the flagged atoms. If the student's error on under-coordinated Si remains below the teacher's error (≈158 meV/Å) and the teacher itself agrees with DFT to within the same margin, the central claim of feasible repair up to the teacher's accuracy is supported; if the student error exceeds the teacher's or the teacher disagrees with DFT by more than stated, the approach is limited to teacher accuracy, not DFT accuracy.","tokens_in":49226,"feed_emoji":"🔬","tokens_out":3137,"duration_ms":34824,"temperature":0.7,"pith_summary":"The paper establishes that whole-system supervision with an exact foundation model is feasible for million-atom molecular dynamics at just 1.3% overhead. A receiver-chunked three-pass evaluation schedule reduces the marginal memory of MACE-MPA-0 by 340-fold, fitting up to 28.6 million atoms on a single H200 GPU. This makes it possible to check every atom of a production simulation. In molten SiO₂, such checks uncover a persistent tail of under-coordinated atoms and an isolated unphysical O–O contact that a 1% random audit misses 94% of the time. On-the-fly cluster retraining on teacher-labelled 13-Å clusters removes three quarters of the failure tail and fixes kilobar-level pressure biases, though the repair converges to the teacher, not to DFT.","feed_headline":"Exact foundation model supervises million-atom MD at 1.3% overhead","feed_subtitle":"Receiver-chunked three-pass schedule cuts memory 340-fold; on-the-fly cluster retraining removes three-quarters of hidden failure tail","key_machinery":"Receiver-chunked three-pass schedule. The total energy of a two-layer MACE model is computed in three sweeps over disjoint chunks of receiver atoms: Pass A evaluates layer-1 features for each chunk without autograd; Pass B evaluates the layer-2 energy and back-propagates it, accumulating the upstream gradient G and the first term of the force derivative; Pass C recomputes layer-1 with autograd and back-propagates a surrogate that includes G, yielding the second term. Per-edge intermediates exist only inside the active chunk, reducing the resident state from 1.55 MB per atom to 4.57 kB per atom (340-fold). On-the-fly cluster retraining: when a segment fails, up to four 13-Å clusters are carved around the atoms with the largest per-atom force deviation, labelled by the teacher in vacuum, and the student is fine-tuned for 40–60 epochs with the force loss restricted to the inner 6 Å to avoid surface label error.","core_discovery":"The paper demonstrates that whole-system checks with an exact equivariant foundation model (MACE-MPA-0) detect sparse, clustered local failures in million-atom MD simulations that system-averaged error metrics and small random audits systematically miss. In molten SiO₂ at 3000 K, a distilled MACE student shows a force RMSE of 66–74 meV/Å—below the 90 meV/Å threshold—yet 1,140 atoms (1.1×10⁻³ of the system) exceed a per-atom force deviation of 0.5 eV/Å, with under-coordinated Si and O enriched 8- and 17-fold. An isolated O–O contact involving six atoms is detected only by checking every atom; a 1% random audit has a 6% probability of containing it. Fine-tuning the student on 13-Å clusters carved around the worst-performing atoms and labelled by the teacher in vacuum (force loss restricted to the inner 6 Å) removes three quarters of the flagged tail, reduces under-coordinated Si error from 143 to 105 meV/Å, and eliminates kilobar-level pressure biases in SiO₂ and compressed water. The repair converges the student toward the teacher, not toward DFT; where the teacher itself errs (e.g. under-coordinated Si at 3000 K, 158 meV/Å against DFT), the student inherits that error.","pith_inferences":["The three-pass schedule removes the memory bottleneck of equivariant message-passing potentials without altering the model or requiring multiple GPUs, so any existing MACE (or NequIP-type) foundation model can immediately be used for whole-system supervision at scale.","The finding that random placement of clusters works as well as worst-atom placement for diffuse tails suggests a two-tier strategy: a cheap global check triggers retraining on a modest number of random clusters, and worst-atom targeting is reserved for rare, isolated events.","Because the repair converges to the teacher and not to DFT, the practical limit of this supervision pipeline is the accuracy of the foundation model itself. Systems or conditions where the teacher is known to be inaccurate (e.g. under-coordinated Si at high T) would require a DFT fine-tuning stage on the teacher, not just the student.","The transient failure tails observed in ambient water (present in only 44 of 189 frames) imply that check-interval design must distinguish between persistent and transient failures—a simple persistence criterion over consecutive checks could avoid unnecessary retraining on noise."],"forward_implications":["Whole-system supervision can be applied as a routine quality-control step in large-scale MD driven by any distilled potential, with a cost that is negligible at production check intervals (1.3% at one check per 100 ps).","The three-pass schedule generalises layer by layer to deeper NequIP-type networks (e.g. SevenNet-0), enabling exact evaluation of multi-layer equivariant models on millions of atoms on a single GPU.","On-the-fly cluster retraining corrects diffuse coordination-defect tails and isolated catastrophic failures without requiring new DFT data, because the teacher labels are the foundation model; the approach is bounded by the teacher’s own accuracy against DFT.","Random placement of clusters suffices for distributed tails, whereas isolated extreme failures (e.g. the O–O contact) require clusters centred on the worst atoms; the checks should keep running during retraining.","The approach identifies that system-averaged RMSE is an unreliable metric for detecting sparse failures in large systems, and that small random audits have vanishing probability of capturing isolated extreme events."],"supporting_citations":[{"why":"MACE architecture and its equivariant message-passing formulation; the paper uses MACE-MPA-0 as the teacher model.","marker":"3, 10"},{"why":"MACE-MPA-0 foundation model trained on MPtrj and Alexandria; provides the teacher labels for distillation and supervision.","marker":"15"},{"why":"NEP89 universal neuroevolution potential; used as the primary speed and accuracy baseline.","marker":"9"},{"why":"NEP architecture and GPUMD implementation; provides the descriptor-based reference for performance comparison.","marker":"5, 6"},{"why":"LAMMPS molecular dynamics engine with Kokkos ML-IAP interface; the three-pass evaluator is coupled through this interface.","marker":"27"},{"why":"Symmetrix-XL provides a concurrent compiled inference engine for MACE; the paper compares performance and exactness with it.","marker":"34"},{"why":"SevenNet-0 is a NequIP-type foundation model; the paper extends the three-pass schedule to its five-layer architecture.","marker":"31"},{"why":"cuEquivariance library provides CUDA-accelerated equivariant tensor product kernels used in both the stock and the three-pass evaluation.","marker":"25"}],"fun_headline_variants":["Atom-by-atom supervision of million-atom MD at 1.3% overhead","Exact AI model catches hidden MD failures at 1.3% cost","340× memory cut enables whole-system check of million-atom MD","On-the-fly retraining removes 75% of missed MD errors"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The MACE-MPA-0 foundation model provides sufficiently accurate force and energy labels for all chemical environments encountered in the simulations, including the under-coordinated defect sites that the distilled student fails on—an assumption the paper shows is violated for under-coordinated Si at 3000 K, where the teacher's own error against DFT is 158 meV/Å.","fun_headline_variants_meta":{"raw":{"variants":["Atom-by-atom supervision of million-atom MD at 1.3% overhead","Exact AI model catches hidden MD failures at 1.3% cost","340× memory cut enables whole-system check of million-atom MD","On-the-fly retraining removes 75% of missed MD errors"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":7.3e-05,"raw_usage":{"cost":7.331688e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":7.331688e-05,"upstream_inference_prompt_cost":1.71612e-05,"upstream_inference_completions_cost":5.615568e-05},"total_tokens":2992,"prompt_tokens":1135,"completion_tokens":1857,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":1774}},"tokens_in":1135,"tokens_out":1857,"cost_source":"openrouter_usage","duration_ms":17523,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":1774,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-19T14:48:48.204287+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run a whole-system check on a 1,000,000-atom molten SiO₂ frame at 3000 K with a distilled student whose per-atom force deviation tail has been suppressed by the proposed cluster retraining, then compare the student's forces against DFT on the flagged atoms. If the student's error on under-coordinated Si remains below the teacher's error (≈158 meV/Å) and the teacher itself agrees with DFT to within the same margin, the central claim of feasible repair up to the teacher's accuracy is supported; if the student error exceeds the teacher's or the teacher disagrees with DFT by more than stated, the approach is limited to teacher accuracy, not DFT accuracy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MACE-MPA-0 foundation model trained on MPtrj and Alexandria; provides the teacher labels for distillation and supervision."},{"cited_title":"NEP89 : universal neuroevolution potential for inorganic and organic materials across 89 elements","cited_arxiv_id":null,"evidence_quote":"NEP89 universal neuroevolution potential; used as the primary speed and accuracy baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LAMMPS molecular dynamics engine with Kokkos ML-IAP interface; the three-pass evaluator is coupled through this interface."},{"cited_title":"Train for Accuracy, Execute at Scale: Architecture-Preserving Inference for Equivariant Atomistic Foundation Models","cited_arxiv_id":"2610.01036","evidence_quote":"Symmetrix-XL provides a concurrent compiled inference engine for MACE; the paper compares performance and exactness with it."},{"cited_title":"Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations","cited_arxiv_id":null,"evidence_quote":"SevenNet-0 is a NequIP-type foundation model; the paper extends the three-pass schedule to its five-layer architecture."},{"cited_title":"GitHub repository, 2024; https://github.com/NVIDIA/cuEquivariance, First release v0.1.0, November 2024","cited_arxiv_id":null,"evidence_quote":"cuEquivariance library provides CUDA-accelerated equivariant tensor product kernels used in both the stock and the three-pass evaluation."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}