{"id":"aa6295b7-1b09-488b-99ac-116c6515b765","arxiv_id":"2607.17978","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RuNNer 2.0 accelerates fourth-generation neural network potentials by implementing a quasi-linear plane-wave charge-equilibration solver, making non-local charge transfer simulations of large systems practical.","lead":"RuNNer 2.0 is a fully rewritten, open-source software suite for training and running high-dimensional neural network potentials (HDNNPs), including fourth-generation models that globally redistribute charge. Benchmarks report order-of-magnitude speedups in charge equilibration versus existing codes, and a 4G model captures non-local charge transfer that a local model cannot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"4G multi-node scaling is not demonstrated and the LAMMPS QEq design serializes the global solve on one MPI rank, so the 'all generations' quasi-linear scaling claim overreaches.","rationale":"The paper is a strong, well-documented software release with open source, public benchmark data, and credible evidence for 2G MPI scaling up to 12 million atoms and for substantial training improvements. The reader's CONDITIONAL verdict is appropriate. My stress test sharpens the most load-bearing weakness: the central 4G scaling claim is not merely missing a favorable CG iteration count; the LAMMPS MD implementation explicitly serializes the global QEq solve on the root MPI rank. This makes multi-node 4G scaling unlikely to follow from the presented architecture, independent of the O(N log^2 N) iteration-count assumption. The missing multi-node 4G benchmark is therefore a substantive gap, not a cosmetic one. The verdict should remain CONDITIONAL: either add a multi-node 4G scaling demonstration, or soften the claims of linear/quasi-linear scaling and HPC readiness for 4G-HDNNPs.","tokens_in":42476,"tokens_out":4824,"duration_ms":44908,"concrete_test":"Run 4G LAMMPS MD with the plane-wave/CG QEq solver on the 96,000-atom water system on 1, 2, 4, and 8 compute nodes, keeping 192 OpenMP threads per node and using 8 threads per MPI rank. Measure per-step wall time and root-rank peak memory. If per-step time does not decrease as nodes are added, or root-rank memory grows with total system size, the root-serialized QEq prevents multi-node scaling and the 'all generations' quasi-linear claim must be restricted to single-node/OpenMP execution.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline claim is that RuNNer 2.0 achieves linear/quasi-linear scaling across all HDNNP generations and makes large-scale 4G MD computationally tractable on HPC clusters. In the LAMMPS MD path, however, Sec. III B 2 and the MD panel of Fig. 4 specify that for 3G/4G models the full local-plus-ghost atom list is assembled on the root MPI task, where 'a single, global charge equilibration and electrostatics calculation' is performed before results are redistributed. QEq is therefore not MPI-distributed in LAMMPS MD. The 4G benchmarks in Fig. 5a and 5e are single-node OpenMP runs or systems of at most 2592 atoms; the MPI strong-scaling data in Fig. 5b are 2G-only. No multi-node 4G benchmark is presented. Thus, even if the Gubler et al. CG iteration count behaves as claimed, the implementation's MD interface places the global QEq work, the FFT grid, the charge vectors, and the associated memory all onto one rank. The conclusion's statement that RuNNer 2.0 achieves quasi-linear scaling for 'all HDNNP generations' is not supported by any 4G multi-node evidence, and the described architecture creates a concrete serial bottleneck in exactly the regime the central claim targets. This is not a missing optional benchmark: it is the main MD code path for 4G potentials.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"RuNNer 2.0 is presented as a from-scratch rewrite of the RuNNer software suite for 2G, 3G, and 4G high-dimensional neural network potentials. The central technical claim is that a plane-wave QEq solver reduces the cost of global charge equilibration from O(N^3) to O(N log^2 N), making 4G-HDNNP molecular dynamics and training quasi-linearly scalable and, for the first time, tractable for large production simulations. The paper describes a four-level software architecture, hybrid MPI/OpenMP parallelization, memory-management strategies for training, native LAMMPS and ASE interfaces, committee-based uncertainty quantification, and extensive CPU benchmarks against n2p2 and PANNA. It also includes a Au/MgO(Al) case study demonstrating that the 4G model captures non-local charge transfer that the local 2G model is categorically unable to represent.","tokens_in":42868,"tokens_out":6645,"duration_ms":57823,"significance":"If the central scaling claims are substantiated, RuNNer 2.0 would be an important enabling tool: it would allow production 4G-HDNNP simulations with global charge transfer at sizes previously restricted to local MLPs. The manuscript has notable strengths: the source code, benchmark data, and analysis scripts are publicly available under GPL; the benchmark protocols are unusually detailed (compilers, git commits, hardware, MPI/OpenMP settings); and the authors are transparent about competitor-code modifications and methodological differences, such as PANNA's angular functions and the improved n2p2 solver. These reproducibility practices are commendable and are part of why the paper has value even where some claims need tightening. However, the headline scaling statements are broader than the evidence, particularly for 4G-HDNNPs in the LAMMPS MD path.","major_comments":[{"comment":"In the LAMMPS MD path, 3G/4G electrostatics and QEq are not MPI-distributed: the full local-plus-ghost atom list is assembled on the root MPI task, where 'a single, global charge equilibration and electrostatics calculation' is performed. The 4G benchmarks supporting the scaling claim are single-node: Fig. 5a is OpenMP-only at 96,000 atoms, and Fig. 5b is 2G-only MPI strong scaling up to 12 million atoms. No multi-node 4G benchmark is presented. The conclusion that RuNNer 2.0 achieves linear/quasi-linear scaling 'across all HDNNP generations' is therefore broader than the evidence. Please either add multi-node 4G strong-scaling data or explicitly restrict the claim to single-node/OpenMP 4G operation and note the root-rank memory/time bottleneck.","section":"Sec. III B 2 / Fig. 4"},{"comment":"The O(N log^2 N) complexity claim is not proven; it rests on the empirical CG-iteration observation of Gubler et al. [62], as the text itself acknowledges. In this paper, Fig. 5e tests 4G systems only up to 2,592 atoms, and Fig. 5a (96,000 atoms) is a single-node timing decomposition without a scaling fit or CG iteration counts. Because this is the paper's headline claim, please report CG iteration counts and total QEq time versus N over at least the 10^3-10^5 atom range, and state explicitly in the abstract/conclusions that the asymptotic complexity is empirical for realistic systems rather than a worst-case bound.","section":"Sec. III F / Abstract"},{"comment":"The claim that 4G-HDNNPs can be trained 'with the same efficiency as their local counterparts' is not established for large single structures. The paper itself notes that storing dq/dr scales O(N^2) and that the Coulomb matrix inverse is cached, both quadratic in atoms per structure. The training benchmark in Fig. 7 uses 96-atom water structures, for which the quadratic memory cost is negligible. Please either benchmark 4G training on a structure large enough to expose the quadratic memory/derivative cost, or limit the claim to per-epoch timings for small/medium structures.","section":"Sec. III C 1 / Fig. 7"}],"minor_comments":[{"comment":"The text states that tanh^3(1) is approximately 0.50; the correct value is approximately 0.442. Please correct.","section":"SI Sec. 1 A 1 b"},{"comment":"The smooth spectral projection is written with different notation and normalization in the main text and the SI. The main-text equation omits the 1/n_lambda normalization and the grid-spacing factor used in the SI. Please align the two definitions.","section":"Eq. (10) vs SI Eq. (13)"},{"comment":"The x-axis tick labels in Fig. 5(e) are illegible in the version I received (e.g., '1 2 4 8 163264'). Please ensure the final figure renders distinct numeric tick values for each subpanel.","section":"Fig. 5(e)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is well taken: the LAMMPS MD path explicitly serializes QEq on the root MPI task, so the broad 'all generations' quasi-linear scaling statement overreaches. This is fixable either by adding multi-node 4G scaling data or by carefully qualifying the claims. I see no circularity problem; the scaling analysis relies on prior, independently published work, but the manuscript should present its own large-system 4G evidence before claiming a first-time tractability result. The paper's transparency about code modifications and competitor caveats is a genuine strength and should be preserved in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RuNNer 2.0 is a substantial, credible software release. What should you know: the production implementation of the plane-wave QEq solver, the memory caching that makes 4G force training nearly as cheap as 2G, and the smooth overlap-matrix descriptor are real contributions, and the benchmarks are unusually well documented. But the abstract's claim of 'quasi-linear scaling across all HDNNP generations' is not supported by the evidence: the 4G LAMMPS MD path serializes the global QEq on one MPI rank, and no multi-node 4G scaling is shown.\n\nThe genuinely new material is the Gubler et al. quasi-linear particle-mesh QEq solver brought into a production HDNNP suite, plus the memory-management strategies that cut the 4G training overhead by a large factor, and a smooth spectral projection for overlap-matrix descriptors that fixes the eigenvalue-crossing problem. The code is open source, datasets and analysis scripts are provided, and the comparison benchmarks against n2p2 and PANNA come with honest caveats: they document patches applied to competitor codes, PANNA's different angular functions, compiler and hardware details, and the n2p2 solver improvements contributed for the test. This is the right way to publish a software paper.\n\nThe main soft spot is the central complexity claim. O(N log^2 N) for QEq rests on an empirical statement from Gubler et al. about CG iteration counts; the text itself says the iteration count 'often increases' with size. And in the LAMMPS interface, Sec. III B 2 describes collecting all local and ghost atoms on the root MPI task and performing a single global QEq there. So the global solve is MPI-serial by design. The 4G benchmarks are single-node OpenMP or systems up to 2592 atoms; the MPI strong-scaling plot is 2G only. That means the conclusion 'linear or quasi-linear scaling across all HDNNP generations' is not backed by data for the regime the paper is about. This is a load-bearing gap, though not a reason to reject the paper.\n\nThe Au/MgO application demonstrates the physics: 4G captures the remote dopant-to-Au charge transfer, 2G cannot. The missing DFT reference points for the thickest doped slab are disclosed in the caption and SI, so that's a minor caveat. Training times show a real five-to-25x improvement over RuNNer 1.3 under matched settings.\n\nVerdict: a serious referee should look at this, and revision should either add 4G multi-node scaling or soften the claims to match the evidence. I would cite it, and I'd bring it to a reading group to discuss how much of the quasi-linear promise is implementation and how much is assumption.","headline":"Substantial open-source HDNNP release that makes 4G training and single-node MD practical, but the quasi-linear O(N log^2 N) scaling claim overreaches because 4G multi-node scaling is absent and the LAMMPS MD path serializes QEq on one MPI rank.","tokens_in":43388,"tokens_out":4509,"would_cite":true,"duration_ms":38957,"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 plane-wave trick cuts charge equilibration from O(N^3) to O(N log^2 N), making fourth-generation neural network potentials with global charge transfer practical at scale.","keywords":["neural network potentials","charge equilibration","fourth-generation HDNNP","plane-wave method","molecular dynamics","machine learning force fields","non-local charge transfer","software performance"],"falsifier":"Take a 4G potential and run the plane-wave/CG QEq on a series of systems from 10^3 to 10^6 atoms with a deliberately delocalized charge distribution, such as a charged metallic slab or a solvated ion at low concentration. Count the number of CG iterations per QEq solve: if iteration count grows linearly with N, or if wall time per step grows faster than N log^2 N, the paper's central complexity claim is refuted. A second check: run the same 96,000-atom 4G molecular dynamics on 1, 2, 4, and 8 compute nodes and measure whether the global QEq wall time decreases with additional MPI ranks.","tokens_in":42415,"feed_emoji":"⚡","tokens_out":5524,"duration_ms":49715,"temperature":0.7,"pith_summary":"The paper presents a full rewrite of a widely used code for training and running high-dimensional neural network potentials (HDNNPs). Its central claim is that the global charge equilibration step of fourth-generation (4G) HDNNPs can be accelerated from cubic to quasi-linear scaling—O(N^3) becomes O(N log^2 N). This matters because 4G models are the ones that let atomic charges respond to distant changes, the mechanism behind non-local charge transfer in doped oxides, charged defects, and metal–oxide interfaces, but that step previously limited them to small systems. The acceleration comes from a plane-wave/FFT approach that evaluates the Coulomb operator's action on a charge vector without ever forming the dense N×N matrix. Benchmarks on water systems up to 96,000 atoms show 4G molecular dynamics reaching about 0.03 ns/day on one compute node, outperforming existing 4G implementations by up to two orders of magnitude, and training runs that are 5–25 times faster than the predecessor code.","feed_headline":"Charge equilibration falls from cubic to quasi-linear scaling","feed_subtitle":"A rewrite makes fourth-generation neural network potentials with global charge transfer practical for large simulations.","key_machinery":"The load-bearing mechanism is a particle-mesh charge equilibration scheme: atomic Gaussian charge densities are projected onto a real-space grid, the Poisson equation is solved in Fourier space with the 4π/G^2 kernel, and the resulting electrostatic potential is integrated against each atom's localized Gaussian to obtain every component of A·q. Charge equilibration (QEq) is the global linear solve that determines environment-dependent atomic charges in 4G HDNNPs; this matrix–vector product is what the conjugate-gradient solver needs at each iteration. By evaluating the product in O(N log N) time and never materializing the dense O(N^2) Coulomb matrix, the scheme removes the cubic bottleneck","core_discovery":"The central discovery is that charge equilibration in 4G HDNNPs can be made quasi-linear by recognizing that the matrix–vector product (A·q)_i equals the gradient of the electrostatic energy with respect to the charge on atom i, which is the integral of that atom's Gaussian charge density times the electrostatic potential. For periodic systems, the potential is obtained by projecting the charge density onto a grid, applying an FFT, multiplying by the reciprocal-space Coulomb kernel 4π/G^2, and back-transforming. This avoids forming or storing the dense Coulomb matrix A. With conjugate-gradient iterations that grow only slowly with system size in practice, the total cost becomes O(N log^2 N),","pith_inferences":["The stated O(N log^2 N) complexity depends on the empirical observation that conjugate-gradient iterations grow slowly with system size; the most direct test of the headline claim is to count iterations for strongly delocalized charge distributions, such as metallic slabs or charged interfaces, where screening length scales may drive faster iteration growth.","Because the global QEq step is performed on a single MPI rank in the present molecular-dynamics interface, the paper's multi-node scaling evidence is effectively for 2G models; a fair test of 4G scalability across nodes would require a distributed QEq solver or a profile of the root-rank bottleneck.","If the quasi-linear QEq holds up in production, it removes a major barrier to using 4G models in active-learning and high-throughput screening loops, where thousands of short simulations are needed; the committee sharing already makes uncertainty estimates cheap enough to drive such loops.","The memory strategy for 4G force training stores the dense global-charge Jacobian and Coulomb matrices, which is O(N^2) per structure; for very large single structures, on-the-fly recomputation may be the only viable route, so the reported training speedups may not transfer unchanged to ultra-large single cells."],"forward_implications":["Fourth-generation HDNNP molecular dynamics becomes practical for systems of tens of thousands of atoms: the paper reports about 0.03 ns/day for a 96,000-atom water system on a single node, with FFT-based QEq scaling well up to 192 threads.","4G training no longer carries a large long-range penalty: with full RAM caching, per-epoch 4G force training is reported as only marginally slower than 2G training, and 5–25 times faster than the previous code version.","Committee-based uncertainty quantification is nearly free at inference because neighbor lists and features are shared across committee members; the marginal cost of adding a member is roughly one extra neural-network forward pass per atom.","The Au/MgO benchmark shows 4G models reproduce doping-induced non-local charge transfer across multiple slab thicknesses, while a strictly local 2G model remains blind to a deep subsurface dopant, implying 4G is necessary for such systems.","Quasi-linear scaling across all HDNNP generations means one codebase can cover local and non-local electrostatics without a sharp performance cliff at the 4G level."],"fun_headline_variants":["RuNNer 2.0 speeds charge equilibration to quasi-linear","Charge equilibration drops from O(N³) to O(N log²N)","RuNNer 2.0 makes 4G neural network potentials practical","Quasi-linear charge transfer with RuNNer 2.0","RuNNer 2.0: global charge transfer without the cubic cost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The quasi-linear scaling claim rests on the observation that conjugate-gradient iterations in the plane-wave QEq solver grow only slowly with system size; if some materials require iteration counts that grow much faster, the O(N log^2 N) advantage—and with it the thesis that 4G models are now scalable—weakens.","fun_headline_variants_meta":{"raw":{"variants":["RuNNer 2.0 speeds charge equilibration to quasi-linear","Charge equilibration drops from O(N³) to O(N log²N)","RuNNer 2.0 makes 4G neural network potentials practical","Quasi-linear charge transfer with RuNNer 2.0","RuNNer 2.0: global charge transfer without the cubic cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1168,"prompt_tokens":786,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":284}},"tokens_in":530,"tokens_out":382,"duration_ms":3667,"temperature":1.0,"reasoning_tokens":284,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:27:30.805366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a 4G potential and run the plane-wave/CG QEq on a series of systems from 10^3 to 10^6 atoms with a deliberately delocalized charge distribution, such as a charged metallic slab or a solvated ion at low concentration. Count the number of CG iterations per QEq solve: if iteration count grows linearly with N, or if wall time per step grows faster than N log^2 N, the paper's central complexity claim is refuted. A second check: run the same 96,000-atom 4G molecular dynamics on 1, 2, 4, and 8 compute nodes and measure whether the global QEq wall time decreases with additional MPI ranks.","supporting_citations":[],"review_version":1}