{"id":"25ed022e-1a81-41f8-821d-287b19aaa802","arxiv_id":"2509.00966","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A multiscale light-matter dynamics software on Aurora claims 152x and 3,780x speedups and 1.87 EFLOP/s, but the performance comparisons and peak-efficiency figure are questionable.","lead":"This paper reports a new exascale software for simulating how light, electrons, and atoms interact in materials, from quantum to device scales. It claims large speedups on the Aurora supercomputer, including the first simulation of light-induced switching of topological structures in a ferroelectric.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3,780× XS-NNQMD speedup is an artifact of dividing time-to-solution by neural-network weights; without that normalization, the claimed improvement collapses from ~3,780× to ~2.4×.","rationale":"The reader's REJECT verdict is well supported, and the weakest assumption they identify—the T2S metric—is the most load-bearing soft spot in the paper's central claims. The 3,780× speedup is a headline number that rests entirely on dividing by the number of neural-network weights. Since the previous SOTA used a 440-weight model and this work uses a 690,000-weight model, the denominator changes by a factor of 1,568, which alone accounts for the bulk of the claimed improvement. When the weight factor is removed, the actual per-atom wall-clock improvement is only about 2.4×. This directly undermines the '3,780× faster' statement in the abstract and the Gordon Bell justification, making the claim misleading as a time-to-solution comparison. I focused on this rather than the additional FLOP/s inconsistencies (e.g., the 100.2% of FP64 peak figure) because the T2S metric is the stated basis for the strongest speedup claims. There is no need to change the reader's verdict: the paper's central quantitative claims are not supported by the evidence as presented. A concrete re-analysis of the raw timings would settle the issue definitively.","tokens_in":16727,"tokens_out":9545,"duration_ms":112080,"concrete_test":"Recompute the XS-NNQMD comparison using only wall-clock time per atom per MD step, i.e., omit the weight denominator, using the Table II timings: SOTA = 3142.66 s / 1.007e12 atoms, present = 1590.31 s / 1.2288e12 atoms. If the ratio is ~2.4 rather than ~3,780, the 3,780× claim is an artifact of weight normalization. Additionally, if feasible, run the SOTA 440-weight model on the same Aurora hardware for the same PbTiO3 system and compare raw end-to-end wall-clock times; if the speedup is nowhere near 3,780×, the headline claim is not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline '3,780×' speedup for XS-NNQMD (Sec. IV, VII.C.2) is based on a time-to-solution (T2S) metric defined as wall-clock time per MD step divided by (number of atoms × number of neural-network weights). This normalization factors out exactly the quantity that differs most between the SOTA and the present work: the prior model [11] has 440 weights, while this work uses 690,000 weights. Using the paper's own numbers (Table II), the SOTA T2S is 3142.66 s / (1.007e12 atoms × 440) = 7.09e-12 s/(atom·weight), while the present work is 1590.31 s / (1.2288e12 atoms × 690,000) = 1.88e-15 s/(atom·weight), giving a ratio of ~3,780. However, if one compares the actual wall-clock time per atom per step—the real time-to-solution—the SOTA is 3142.66 / 1.007e12 = 3.12e-9 s/(atom·step), and this work is 1590.31 / 1.2288e12 = 1.29e-9 s/(atom·step), a speedup of only ~2.4×. The weight denominator therefore inflates the speedup by a factor of ~1,568 (the weight ratio itself). This is not a minor normalization choice: the central claim of a '3,780× improvement' in the abstract and Gordon Bell justification is unsupported by the reported timings as stated. The metric rewards using vastly more parameters rather than measuring end-to-end simulation speed, so the headline result does not hold as a genuine time-to-solution comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MLMD, a multiscale light-matter dynamics software integrating the DC-MESH (divide-and-conquer Maxwell–Ehrenfest–surface hopping) and XS-NNQMD (excited-state neural-network quantum molecular dynamics) modules. On the Aurora supercomputer, the authors report 1.873 EFLOP/s for a 15.36-million-electron PbTiO3 system, a 152-fold time-to-solution (T2S) improvement over prior Maxwell–Ehrenfest simulations, a 3,780-fold T2S improvement for a 1.23-trillion-atom XS-NNQMD simulation, and near-perfect weak scaling for both modules. The paper claims this is the first end-to-end exascale multiscale light-matter dynamics simulation and demonstrates light-induced switching of topological superlattices.","tokens_in":17257,"tokens_out":7609,"duration_ms":92657,"significance":"If the headline numbers were sound, this would be a significant Gordon Bell–level achievement: integrating Ehrenfest and surface-hopping dynamics with neural-network potentials at exascale, exploiting mixed precision and heterogeneous hardware, and reporting near-perfect weak scaling. The algorithmic contributions—DCR spatial/physical decomposition, shadow dynamics, GEMMification of nonlocal corrections, and parameterized mixed precision—are interesting and the per-kernel optimization data are valuable. However, the central performance claims are undermined by two load-bearing problems: the XS-NNQMD 3,780× speedup is an artifact of a weight-normalized T2S metric, and the 1.873 EFLOP/s figure is reported as 100.2% of FP64 peak while the stated machine FP64 peak is approximately 2 EFLOP/s. The scientific claim about topological superlattice switching is also not substantiated with results. As written, the paper's headline claims do not survive scrutiny.","major_comments":[{"comment":"The 3,780× XS-NNQMD speedup is manufactured by the T2S definition. The paper defines T2S as wall-clock time per MD step divided by (number of atoms × number of neural-network weights). With the authors' own numbers, SOTA [11] gives 3142.66 s/(1.007e12 atoms × 440 weights)=7.09e-12 s/(atom·weight), while this work gives 1590.31 s/(1.2288e12 atoms × 690,000 weights)=1.88e-15 s/(atom·weight), a ratio of ~3,780. If one instead compares the actual wall-clock time per atom per step—the only quantity that reflects time-to-solution—the SOTA is 3.12e-9 s/(atom·step) and this work is 1.29e-9 s/(atom·step), a speedup of only ~2.4×. Dividing by the number of weights rewards models with more parameters and has no direct connection to accuracy or time-to-solution. Since the abstract and Gordon Bell justification rest on the 3,780× number, this is a load-bearing flaw.","section":"Sec. IV, Sec. VII.C.2, Table II"},{"comment":"The peak-performance claim is internally inconsistent. The paper states in Sec. VI.B that the entire Aurora machine has a theoretical FP64 peak of approximately 2 EFLOP/s, yet Sec. VII.B reports 1.873 EFLOP/s on 10,000 nodes as \"100.2% of FP64 peak.\" 1.873/2 = 93.65%, not 100.2%. If the denominator is instead the FP64 peak of the allocated 10,000 nodes, that should be stated explicitly, along with how that peak is computed. Additionally, Table IV lists a single-tile FP64 peak of 23 TFLOP/s, while Sec. VI.B says a blade has 187 TFLOP/s peak; with 6 GPUs × 2 tiles per blade, these numbers do not reconcile. The FLOP accounting (Intel SDE, unitrace, scaling from a single tile to 10,000 nodes) needs clarification before the 1.87 EFLOP/s claim can be assessed.","section":"Sec. VII.B, Sec. VI.B, Table I"},{"comment":"The 152× DC-MESH T2S improvement is also not an apples-to-apples comparison. The paper compares 1.11e-7 s/electron for a 15,360,000-electron PbTiO3 MESH simulation on 10,000 Aurora nodes with 1.69e-5 s/electron for a 71,040-electron silica Maxwell–Ehrenfest subset simulation on 27,648 Fugaku nodes. These runs differ in problem size by more than two orders of magnitude, in physical system, and in the physics included (MESH superset vs ME subset). The large speedup is at least partly a consequence of scale and machine generation. A convincing claim would require either same-machine, same-size controlled comparisons or a clear argument that the per-electron T2S is insensitive to these differences.","section":"Sec. VII.C.1, Table I"},{"comment":"The mixed-precision accuracy claim is asserted rather than demonstrated. The paper says BF16 with FP32 accumulation is sufficient for the nonlocal correction and refers to [34], but no error analysis or validation is provided for the production 15.36-million-electron run that achieves 1.873 EFLOP/s. Since the FLOP/s number and the scientific validity of the dynamics both depend on this precision choice, the paper should include a direct convergence check (e.g., comparing BF16/FP32 results against FP64 for observables such as electronic occupations or excited-state populations).","section":"Sec. V.B.7, Sec. VI.C, Sec. VII.B"}],"minor_comments":[{"comment":"Several superscripts are garbled, e.g., \"1.88×10!\"# [sec/(atom•weight•step)]\". Please fix all exponent and unit formatting.","section":"Abstract and Performance Attributes"},{"comment":"The hardware description is ambiguous: \"PVC tiles per blade have a peak floating-point performance of 187 TFLOP/s\" is unclear whether 187 TFLOP/s is per tile, per GPU, or per blade. The later statement that machine peak is ~2 EFLOP/s implies per blade, but Table IV uses 23 TFLOP/s per tile. Please clarify.","section":"Sec. VI.B"},{"comment":"The paper claims to have \"enabled the first study of light-induced switching of topological superlattices,\" but no switching dynamics, order-parameter time series, or comparison with experiment is presented. As a performance paper this may be outside scope, but the scientific claim should be toned down or supported.","section":"Sec. VIII and Fig. 3"},{"comment":"Typo: \"Insitute\" should be \"Institute\" in the Stanford PULSE affiliation.","section":"Author affiliations"},{"comment":"The SOTA XS-NNQMD comparison uses a 440-weight model from 2022, while this work uses a 690,000-weight Allegro-FM fine-tuned model. Even apart from the T2S normalization, the difference in model architecture and the claimed \"accuracy-speed trade-off\" are not quantified; please state the accuracy difference, e.g., force/energy errors on the same test set.","section":"Sec. IV and Table II"}],"recommendation":"reject","confidential_remarks":"The 3,780× XS-NNQMD speedup is the central Gordon Bell justification and it is an artifact of normalizing T2S by the number of neural-network weights. Using the paper's own timings, the honest speedup is about 2.4×. The 1.873 EFLOP/s figure also conflicts with the stated machine peak. These are not local presentation issues; they change what the paper claims. The authors could potentially revise the manuscript around the 2.4× speedup and a correctly scoped FLOP/s claim, but that would be a substantially different paper. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a serious submission from a serious group, and the scientific integration is genuinely new. They combine DC-MESH (Ehrenfest + surface hopping + Maxwell) with XS-NNQMD using Allegro-FM, and they actually ran a 15.4-million-electron QD and a 1.23-trillion-atom NNQMD on 10,000 Aurora nodes. That is a real engineering achievement. The algorithmic narrative—divide-and-conquer-recombine, shadow dynamics, GEMMification of nonlocal corrections, mixed precision—is plausible and largely builds on their own prior work.\n\nBut the performance claims don't survive scrutiny. The 3,780x XS-NNQMD speedup is the biggest problem. The paper defines T2S as wall-clock per step divided by (atoms × NN weights). The SOTA model has 440 weights; theirs has 690,000. That normalization alone inflates the ratio by a factor of ~1,568. Using the more natural metric of wall-clock per atom per step, the raw speedup is about 2.4x. The paper says the weight normalization accounts for accuracy, but it never demonstrates any accuracy improvement. This is not a nitpick: the abstract and Gordon Bell justification lead with 3,780x. The second issue is the peak-performance claim. They state the machine FP64 peak is ~2 EFLOP/s, then report 1.873 EFLOP/s as 100.2% of FP64 peak. That only works if you use the peak of the 10,000 nodes actually used (1.87 EFLOP/s), but then the per-tile peak of 23 TFLOP/s in Table IV contradicts it. And a real application hitting 100% of theoretical peak is not credible; that number suggests FLOP counting that matches theoretical ops rather than executed ops. The 152x DC-MESH speedup is less egregious but still a cross-machine, cross-code comparison with different problem sizes and no error bars. No code or data are provided.\n\nThe work is worth engaging with. The integration is a first, and the ideas are sound enough to warrant a serious referee. But the draft as written would need major revision: drop the weight-normalized metric, re-analyze the peak number, and add some validation of accuracy or at least error estimates. I would not cite the current version, and I'd encourage the authors to resubmit with honest metrics.","headline":"The integration is real but the headline numbers are not: the 3,780x speedup is a metric artifact and the 100.2% of peak is internally inconsistent.","tokens_in":17712,"tokens_out":6137,"would_cite":false,"duration_ms":71241,"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":"A single multiscale simulation code sustains 1.87 EFLOP/s and cuts quantum-dynamics time-to-solution 152x versus the prior record.","keywords":["light-matter dynamics","quantum molecular dynamics","divide-and-conquer algorithms","exascale computing","neural-network potentials","mixed-precision arithmetic","topological materials","ferroelectric superlattice"],"falsifier":"Instrument the 15,360,000-electron run with an independent FLOP counter that separates FP64, FP32, and BF16 operations and divide by wall-clock time; if the result is well below 1.87 EFLOP/s, the peak-fraction claim fails. Also rerun the previous state-of-the-art benchmark—71,040 electrons, comparable accuracy, same time-to-solution definition—with this code; if the time-to-solution is not near the claimed 1.11e-7 seconds per electron per step, the 152x comparison is configuration-dependent.","tokens_in":16685,"feed_emoji":"⚡","tokens_out":8556,"duration_ms":103298,"temperature":0.7,"pith_summary":"This paper claims that the full multiscale chain of light-matter dynamics—Maxwell's equations for light, Schrödinger equations for electrons, Newton's equations for atoms, and neural-network surrogate dynamics—can be simulated end to end on an exascale machine by decomposing the problem physically as well as spatially and mapping each piece to the hardware that suits it. On a 10,000-node exascale system it reports sustaining 1.873 EFLOP/s (100.2% of FP64 peak) for a 15,360,000-electron quantum-dynamics run, along with 152x and 3,780x reductions in time-to-solution for quantum dynamics and excited-state neural-network molecular dynamics versus prior state-of-the-art results. These numbers, if correct, make this the first end-to-end exascale multiscale light-matter simulation and enable the first first-principles study of light-induced switching in a ferroelectric skyrmion superlattice. The broader claim is that hardware heterogeneity and low-precision arithmetic can be used as resources rather than obstacles for scientific computing.","feed_headline":"Exascale code hits 1.87 EFLOP/s for quantum light-matter dynamics","feed_subtitle":"A divide-and-conquer scheme cuts time-to-solution 152x and simulates 15.4 million electrons end to end.","key_machinery":"The load-bearing mechanism is the divide-conquer-recombine split into physical subproblems with small dynamic ranges: local field dynamics—light-driven Schrödinger and Maxwell propagation on regular grids—runs on GPU, while quantum excitation molecular dynamics—nonadiabatic electron-atom coupling and surface hopping—runs on CPU. Shadow dynamics and metamodel-space algebra let the two exchange only tiny occupation-number deltas, amortized over roughly a hundred quantum-dynamics steps, so large wave functions stay GPU-resident. Nonlocal corrections are 'GEMMified' into dense matrix multiplications that tolerate BF16 precision; excited-state neural-network molecular dynamics blends a ground-sta","core_discovery":"The central claim is that a divide-conquer-recombine algorithmic paradigm, combined with a metamodel-space algebra that minimizes inter-unit communication, turns the exascale heterogeneity challenge into a speed advantage. The paper decomposes the coupled light-electron-atom problem into local field dynamics run on GPU tiles and quantum excitation molecular dynamics run on CPUs, exchanging only small occupation-number changes instead of large wave-function arrays. Nonlocal corrections are reframed as dense matrix multiplications and run in FP32/BF16, with FP64 kept for the chemically sensitive part. On 10,000 nodes (60,000 GPUs), this yields 1.873 EFLOP/s for a 15,360,000-electron PbTiO3 sys","pith_inferences":["The divide-conquer-recombine pattern suggests a recipe for other multiphysics codes: decompose by dynamic range, pair each subproblem with its best precision, and exchange only low-dimensional summaries at coarse time steps; this could be tested on couplings such as fluid-plasma or chemistry-atmosphere systems.","The shadow-dynamics exchange of occupation numbers rather than wave functions implies that multi-fidelity AI/physics handshaking can work when the AI module receives only a scalar summary of the quantum state, which may generalize to other neural-network-accelerated simulations.","The reported system sizes open a concrete comparison target: experimental pump-probe measurements of ferroelectric skyrmion switching should now be simulable at comparable length and time scales, so threshold fluences and switching times can be checked directly.","Because the code uses OpenMP target offload and standard BLAS calls, porting it to a different vendor's GPUs is a direct test of whether the heterogeneity exploitation is portable or specific to the reported machine."],"forward_implications":["Quantum light-matter dynamics of multimillion-electron systems becomes practical on exascale machines, not just ground-state electronic structure.","The divide-conquer-recombine/metamodel-space algebra pattern provides a template for mapping other multiscale and multiphysics problems onto heterogeneous, low-precision hardware.","Excited-state molecular dynamics of trillion-atom systems with first-principles-accuracy neural networks becomes feasible, enabling long-time simulations that were previously out of reach.","Simulation-guided study of light-induced topological switching can proceed at device-relevant scales, potentially informing experiments on ferroelectric topotronic devices.","Mixed-precision arithmetic is sufficient for the local part of quantum propagation, while full FP64 is reserved for chemically sensitive parts—a precision hierarchy for future exascale applications."],"supporting_citations":[{"why":"Provides the prior state-of-the-art time-to-solution for Maxwell-Ehrenfest simulation that the 152x improvement is measured against.","marker":"[23]"},{"why":"Provides the only prior excited-state neural-network quantum molecular dynamics time-to-solution, the 440-weight baseline for the 3,780x comparison.","marker":"[11]"},{"why":"Supplies the equivariant Allegro model whose 1,000x speedup over quantum simulation underlies the XS-NNQMD performance.","marker":"[26]"},{"why":"Adds sharpness-aware training that suppresses unphysical force outliers and lengthens time-to-failure in long simulations.","marker":"[27]"},{"why":"Introduces the Allegro foundation model and total-energy alignment that give the ground-state model transferability across elements.","marker":"[28]"},{"why":"Earlier GPU acceleration of this code; supplies the kernel optimizations and GEMMification used here.","marker":"[12]"},{"why":"Shows hybrid FP32/BF16 precision is accurate enough for nonlocal correction, justifying the mixed-precision settings.","marker":"[34]"},{"why":"Defines the divide-conquer-recombine and globally-scalable-locally-fast solver framework used for spatial decomposition.","marker":"[30]"},{"why":"Applies multiscale divide-and-conquer to the Maxwell-Ehrenfest subset, the closest prior work this paper extends to full surface-hopping and neural-network coupling.","marker":"[25]"}],"fun_headline_variants":["Divide-conquer-recombine hits 1.87 EFLOP/s on 60k GPUs","Simulating 15.4 million electrons hits 1.87 EFLOP/s","Quantum materials code 152x faster, hits exascale","1.23 trillion atoms simulated 3780x faster"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The headline speedups rest on the assumption that time-to-solution normalized by electron count or atom-weight is a fair way to compare runs on different machines with very different models; if the reference runs are not equivalently configured, the speedup factors lose their meaning.","fun_headline_variants_meta":{"raw":{"variants":["Divide-conquer-recombine hits 1.87 EFLOP/s on 60k GPUs","Simulating 15.4 million electrons hits 1.87 EFLOP/s","Quantum materials code 152x faster, hits exascale","1.23 trillion atoms simulated 3780x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001319,"raw_usage":{"total_tokens":5230,"prompt_tokens":785,"completion_tokens":4445,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":4363}},"tokens_in":529,"tokens_out":4445,"duration_ms":36852,"temperature":1.0,"reasoning_tokens":4363,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:00:46.172924+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument the 15,360,000-electron run with an independent FLOP counter that separates FP64, FP32, and BF16 operations and divide by wall-clock time; if the result is well below 1.87 EFLOP/s, the peak-fraction claim fails. Also rerun the previous state-of-the-art benchmark—71,040 electrons, comparable accuracy, same time-to-solution definition—with this code; if the time-to-solution is not near the claimed 1.11e-7 seconds per electron per step, the 152x comparison is configuration-dependent.","supporting_citations":[{"cited_title":"$ This work Aurora 1.88×10!","cited_arxiv_id":null,"evidence_quote":"Provides the only prior excited-state neural-network quantum molecular dynamics time-to-solution, the 440-weight baseline for the 3,780x comparison."},{"cited_title":"V.A.6-7)","cited_arxiv_id":null,"evidence_quote":"Introduces the Allegro foundation model and total-energy alignment that give the ground-state model transferability across elements."}],"review_version":1}