{"id":"3bd0b11a-0229-46af-a2b7-09aad4cc86c4","arxiv_id":"2608.03280","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Polynomial angular descriptors in aenet and n2p2 potentials can be evaluated exactly from one-neighbor Cartesian moments, removing the quadratic pair loop and accelerating trained models without retraining.","lead":"AccelNet rewrites the angular part of two common machine-learning interatomic potentials so each atom's descriptor comes from one-neighbor sums instead of neighbor-pair loops, giving identical energies and forces. The result is an exact, no-retraining speedup that grows with local atom density and is most useful for dense environments or long simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"n2p2 backward-compatibility breadth rests on one TiO2 model at LAMMPS output precision; a compatibility suite across supported n2p2 conventions is needed to settle.","rationale":"The algebraic derivation is sound: I rechecked the key manipulations—B_p as the ordered product of moments, the diagonal subtraction D_i, T_n(1)=1 for the Chebyshev case, the 1/(1+δ) normalization for β=δ, and D=0 for β≠δ in the n2p2 case. None of these steps introduces an error, and the paper's self-assessment of an exact reformulation is credible. The residual risk is entirely empirical: the software's n2p2 compatibility is asserted more broadly than the single tested model supports. Table 2's n2p2 comparison at LAMMPS output precision is weaker than the Chebyshev comparisons at ~1e-14 eV/Å, and Section 4.1 lists many supported conventions that could each hide an ordering/scaling bug. Because backward compatibility for supported n2p2 functions is a central contribution, not a peripheral detail, I would make acceptance conditional on the additional compatibility suite rather than rejecting or unconditionally accepting. If the suite passes, the verdict should revert to ACCEPT.","tokens_in":14066,"tokens_out":10364,"duration_ms":129032,"concrete_test":"Run the released AccelNet converter and LAMMPS pair style against a generated suite of small n2p2 models spanning: zeta=1,2,5,10; lambda=±1; scaling on/off; normalize_nodes on/off; cutoff types 0,4,8,9; the fractional cutoff; two-element systems; and element-specific vs global network topologies. Compare energies and forces against n2p2's own high-precision prediction output (not LAMMPS log/dump files) on random configurations, requiring max|ΔE|<1e-12 eV and max|ΔF|<1e-12 eV/Å. Any mismatch would localize the unsupported convention; a clean pass would settle the backward-compatibility claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The Cartesian-moment factorization itself is exact: Eq. (23) follows from Eqs. (18)-(22), and Eq. (36) correctly handles the species masks, the 1/(1+δ) normalization, and the self-term subtraction for the n2p2 wide angular function. The load-bearing weakness is not algebraic but the breadth of the backward-compatibility assertion for n2p2. Section 4.1 claims support for symmetry-function types 2, 3, and 9, cutoff types 0-8 and the fractional cutoff, documented scaling modes, normalize_nodes, global or element-specific topologies, and direct directory loading. Yet the only end-to-end n2p2 check is a single 192-atom TiO2 model with 54 type-9 angular functions, and Table 2/Section 3.2 reports agreement only at the precision of LAMMPS log/dump output, not full machine precision. If any supported convention (e.g., a different cutoff type, scaling mode, normalize_nodes setting, or multi-element descriptor ordering) is misread or reordered, energies and forces would deviate despite the exact algebra. The paper's claim to be an exact drop-in evaluator for supported n2p2 polynomial angular functions is therefore under-supported by the current verification, even though the central derivation is sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper derives an exact algebraic identity for angular descriptor components of the form of Eq. (15), where the angular kernel is a finite polynomial of cos(theta) and the remaining weight factorizes into one-neighbor terms. Under these conditions, the unordered pair sum is shown to equal one-half the contraction of one-neighbor Cartesian moments (Eq. (23)) minus a diagonal self-term. The multinomial expansion in Eq. (18) is correct, as are the specializations to the aenet Chebyshev descriptor (Eq. (28)) and to n2p2 wide angular functions with integer zeta (Eq. (36)), including species masks, the 1/(1+delta) normalization, and the (1+lambda)^zeta self-term subtraction. The authors implement this in the AccelNet inference engine, read trained aenet and n2p2 models, and validate energies and forces against the external packages and by finite differences. Single-core LAMMPS benchmarks carefully decompose the observed speedup into engine and algorithmic contributions; the moment algorithm is 3.629x faster than optimized direct evaluation for the dense TiO2 Chebyshev model and 1.371x faster for the TiO2 n2p2 model.","tokens_in":14346,"tokens_out":7059,"duration_ms":82861,"significance":"If the breadth claims can be supported, this is a practically valuable contribution: it accelerates existing trained neural-network potentials without retraining, with an exactness guarantee that is established algebraically rather than by fitting. The derivation is self-contained and parameter-free, the timing decomposition in Eq. (49) is honest about which gains come from the algorithm and which from the implementation, and the open-source release of the code, converters, and LAMMPS interfaces strengthens reproducibility. The main evidential weakness is the narrowness of the external verification for n2p2: the central exact identity is proved, but the claim that AccelNet is a drop-in evaluator across the full advertised set of n2p2 conventions is supported by only one end-to-end model and by output-precision-limited comparisons.","major_comments":[{"comment":"The paper claims support for n2p2 symmetry-function types 2, 3, and 9, cutoff types 0-8 plus the fractional cutoff, documented scaling modes, energy normalization, atomic reference energies, global or element-specific topologies, and normalize_nodes. The only end-to-end n2p2 comparison in Table 2 is a single 192-atom TiO2 model with 16 type-2 radial and 54 type-9 angular functions. No end-to-end test is reported for other cutoff types, other scaling modes, normalize_nodes enabled, element-specific topologies, or type-3 functions. The descriptor-level tests in Section 3.3 are useful but do not exercise the full model-loading, descriptor-ordering, and network-weight path. Because exact backward compatibility is the central claim, this breadth gap is load-bearing. I would be satisfied either by adding a small compatibility matrix of n2p2 models covering these conventions or by explicitly re","section":"Section 4.1 vs Section 3.2, Table 2"},{"comment":"The abstract states that AccelNet reproduces descriptors, energies, and forces to floating-point roundoff, while Section 3.2 reports for the n2p2 model that the comparisons agree at the precision of LAMMPS log and dump files. The zero entries in Table 2 therefore do not demonstrate full machine-precision agreement for the n2p2 path; they only show agreement at the printed precision. This is a weaker statement than the abstract makes. The Chebyshev rows, with force differences around 1e-14 eV/A, do support the roundoff wording, but the n2p2 comparison should either be repeated with higher-precision output (for example, many-digit ASCII output) or the claim should be qualified as holding to the printed precision for this model.","section":"Abstract and Section 3.2, Table 2"}],"minor_comments":[{"comment":"The sentence 'The H2O and Chebyshev TiO2 systems were propagated with NVE integration and canonical stochastic velocity rescaling at 300 K' combines two incompatible ensembles. Please clarify whether the thermostat is applied after NVE integration or whether these are separate protocols.","section":"Section 4.5"},{"comment":"Minor typo: 'A n2p2 model is specified' should read 'An n2p2 model is specified.'","section":"Section 4.4"},{"comment":"The phrase 'second-generation TiO2 model' is vague; replacing it with 'the n2p2 TiO2 model' would make the table and surrounding text easier to follow.","section":"Section 3.1 / Table 1"}],"recommendation":"major_revision","confidential_remarks":"The algebraic core of the paper is sound and the software release is a genuine contribution. The main reason for major revision rather than acceptance is that the backward-compatibility claim for n2p2 is stated very broadly while the external verification is narrow and, in the n2p2 case, limited by output precision. This is fixable with additional tests or a more careful caveat, so I do not view it as a fatal flaw."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Briefly: this paper is worth a serious referee. The Cartesian moment factorization in Eqs. (18)-(23) is exact, the derivation is self-consistent, and the numerical checks agree with aenet to roundoff. The timing decomposition is honest—they separate engine gains from algorithmic gains, and the 3.6x algorithmic speedup for the high-neighbor TiO2 Chebyshev case is credible. The n2p2 speedup of 1.37x is more modest but still real.\n\nWhat's new: the multinomial expansion of a dot product is standard, and the paper properly credits MTP, ACE, CACE, and polynomial-kernel feature maps. The contribution is the careful application to trained legacy models: preserving species channels, ordered-pair normalization, self-term subtraction, scaling, and network weights, so the descriptor values are equivalent up to floating-point order. That is the hard part, and they do it.\n\nSoft spots: the 'backward-compatible' claim for n2p2 is broader than the tests. Section 4.1 lists types 2, 3, 9 for symmetry functions, cutoff types 0-8 and fractional, scaling modes, normalize_nodes, global or element-specific topologies, direct loading. But the only end-to-end n2p2 check is a single 192-atom TiO2 model with 54 type-9 functions, and agreement is reported at the precision of LAMMPS output, not full machine precision. If any supported convention is misread or reordered, energies would shift despite the exact algebra. That doesn't invalidate the central result, but it does mean the drop-in claim for n2p2 is under-verified. The same concern applies to aenet: the two Chebyshev models cover one descriptor family, but aenet has Behler2011 G1-G5 inputs as well, and those aren't tested end-to-end.\n\nThe paper's own limitations section (6.5) is clear about what cannot be accelerated: rjk-dependent terms, noninteger powers, general nonpolynomial kernels. That's honest. And unsupported n2p2 settings are rejected explicitly, which mitigates the compatibility risk.\n\nWho is this for: anyone doing production MD with trained aenet or n2p2 potentials, especially high-neighbor systems. It's not a conceptual breakthrough in ML potential design, but it's a well-executed algorithmic improvement with code released. A serious referee should engage; the main request should be broader compatibility tests, especially for n2p2 conventions.","headline":"A clean, exact algebraic reformulation of standard angular descriptors that does what it claims; the n2p2 drop-in claim is slightly broader than the tests support.","tokens_in":14825,"tokens_out":3485,"would_cite":true,"duration_ms":34690,"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":"AccelNet shows that polynomial angular descriptors used by trained aenet and n2p2 neural-network potentials are exactly equal to one-neighbor Cartesian moment contractions, and uses this identity to accelerate molecular-dynamics evaluations","keywords":["machine-learning interatomic potential","neural-network potential","angular descriptor","Cartesian moment factorization","backward-compatible acceleration","Chebyshev descriptor","n2p2 symmetry functions","molecular dynamics"],"falsifier":"Load a supported n2p2 model with positive-integer angular symmetry functions and compare AccelNet's moment path against n2p2 at full double precision on configurations with all species channels active; any energy or force difference above about $10^{-12}$ eV or eV/Å would show a compatibility gap not visible in LAMMPS output precision. Likewise, evaluate an aenet Chebyshev model with angular order beyond four and compare descriptor derivatives against direct pair-sum evaluation.","tokens_in":13953,"feed_emoji":"⚛️","tokens_out":6086,"duration_ms":66110,"temperature":0.7,"pith_summary":"Angular descriptors in Behler–Parrinello style neural-network potentials compare a central atom to every pair of its neighbors, so their cost grows with the square of the local neighbor count. This paper proves that whenever the angle enters through a finite polynomial of the cosine and the radial and chemical factors split into one-neighbor weights, the entire neighbor-pair sum is exactly equal to a contraction of one-neighbor Cartesian moments. The proof is purely algebraic: each power of a dot product is expanded by the multinomial theorem, the per-neighbor sums are formed first, and the diagonal self terms are subtracted explicitly. AccelNet implements this identity for the aenet Chebyshev angular descriptor and for supported n2p2 polynomial angular functions, preserving descriptor ordering, scaling, network weights, energies, and analytic forces. In a dense 1152-atom TiO2 Chebyshev model, the moment path is 3.629 times faster than an optimized direct pair-loop evaluation.","feed_headline":"Moment identity makes trained neural-network potentials 3.6x faster","feed_subtitle":"Trained aenet and n2p2 models run unchanged in MD, matching energies and forces to roundoff","key_machinery":"The Cartesian moment factorization identity. For $P(c)=\\sum_{p=0}^{P_{\\max}} a_p c^p$ with $c_{ijk}=u_{ij}\\cdot u_{ik}$, the descriptor $G_{i\\gamma}=\\sum_{j<k} h_{ij}^{(\\gamma)} h_{ik}^{(\\gamma)} P(c_{ijk})$ equals $\\frac{1}{2}\\left[\\sum_p a_p \\sum_{\\ell_x+\\ell_y+\\ell_z=p} \\frac{p!}{\\ell_x!\\ell_y!\\ell_z!}(M_{i,\\ell}^{(\\gamma)})^2 - P(1)D_i^{(\\gamma)}\\right]$. This replaces $O(z_i^2)$ neighbor-pair enumeration with $O(z_i)$ moment accumulation, while the contraction step costs depend on the polynomial degree but not on the neighbor count.","core_discovery":"For angular descriptors of the form $G_{i\\gamma}=\\sum_{j<k} h_{ij}^{(\\gamma)} h_{ik}^{(\\gamma)} P(c_{ijk})$ with $P$ a finite polynomial and $c_{ijk}=u_{ij}\\cdot u_{ik}$, the unordered neighbor-pair sum is exactly equal to the moment contraction $\\frac{1}{2}\\left[\\sum_p a_p \\sum_{\\ell_x+\\ell_y+\\ell_z=p} \\frac{p!}{\\ell_x!\\ell_y!\\ell_z!}(M_{i,\\ell}^{(\\gamma)})^2 - P(1)D_i^{(\\gamma)}\\right]$, where $M_{i,\\ell}^{(\\gamma)}=\\sum_j h_{ij}^{(\\gamma)} u_{ij,x}^{\\ell_x}u_{ij,y}^{\\ell_y}u_{ij,z}^{\\ell_z}$ and $D_i^{(\\gamma)}=\\sum_j (h_{ij}^{(\\gamma)})^2$. No approximation enters; only the summation order changes. The paper applies this identity to the Artrith–Urban–Ceder Chebyshev angular descriptor us","pith_inferences":["The same factorization should apply to any legacy angular descriptor whose angle dependence is a finite polynomial in $\\cos\\theta$ with separable one-neighbor weights, so other radial/angle symmetry-function families may also have hidden finite-rank structure exploitable without refitting.","The identity is the same kind of density trick used by moment tensor potentials and atomic cluster expansions, which suggests that other existing descriptor families could be reformulated as internally evaluated moments rather than as new fitted bases.","A production code could extend AccelNet's static crossover by routing per atom or per frame based on instantaneous neighbor counts, which the paper does not implement.","The 3.629x factor is for a single core; combining the moment formulation with multi-core or GPU neighbor loops may yield larger end-to-end gains in production simulations."],"forward_implications":["Existing trained aenet and n2p2 models can be run through AccelNet in LAMMPS without retraining or refitting, with energies and forces preserving their original values to roundoff.","The speedup from moment evaluation grows with local neighbor count relative to polynomial degree, so dense environments with large cutoffs benefit most.","For the supported n2p2 polynomial angular functions, moment evaluation is 1.371 times faster than the optimized direct path even when the reference is an independently optimized modern implementation.","Angular functions that do not satisfy the separability or polynomial conditions are detected and evaluated directly, so unsupported terms never change the model's semantics.","The direct-versus-moment crossover is model-dependent; AccelNet's automatic policy is a heuristic threshold, not a universal optimal point."],"supporting_citations":[{"why":"Defines the Artrith–Urban–Ceder Chebyshev angular descriptor whose finite polynomial form is the object being factorized.","marker":"[18]"},{"why":"Defines the aenet implementation and trained-model conventions that AccelNet must reproduce exactly.","marker":"[12]"},{"why":"Describes the n2p2 parallel training code and its symmetry-function conventions, including the type 9 angular functions targeted here.","marker":"[14]"},{"why":"Introduces atom-centered symmetry functions, the framework in which these angular descriptors are defined.","marker":"[10]"},{"why":"Establishes the atomic-energy decomposition that lets descriptor-level equality imply energy and force equality.","marker":"[5]"},{"why":"Provides LAMMPS, the molecular-dynamics engine through which the accelerated potentials are benchmarked and integrated.","marker":"[19]"},{"why":"Earlier library-based LAMMPS implementation of n2p2 potentials that the AccelNet LAMMPS interface parallels.","marker":"[13]"}],"fun_headline_variants":["Exact moments replace neighbor loops in NN potentials for speedup","Trained NN potentials run faster without retraining via moment trick","Polynomial angular terms compressed to moments for exact acceleration","Moment identity yields exact speedup for aenet and n2p2 models","No retraining: exact acceleration of neural-network potentials"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"That the three tested models (H2O Chebyshev, TiO2 Chebyshev, and TiO2 n2p2) exercise every descriptor, scaling, and network convention a supported aenet or n2p2 model can use, so the roundoff-level agreement extends to all compatible models.","fun_headline_variants_meta":{"raw":{"variants":["Exact moments replace neighbor loops in NN potentials for speedup","Trained NN potentials run faster without retraining via moment trick","Polynomial angular terms compressed to moments for exact acceleration","Moment identity yields exact speedup for aenet and n2p2 models","No retraining: exact acceleration of neural-network potentials"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00107,"raw_usage":{"total_tokens":4324,"prompt_tokens":751,"completion_tokens":3573,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":3501}},"tokens_in":495,"tokens_out":3573,"duration_ms":30103,"temperature":1.0,"reasoning_tokens":3501,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:47:23.976811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Load a supported n2p2 model with positive-integer angular symmetry functions and compare AccelNet's moment path against n2p2 at full double precision on configurations with all species channels active; any energy or force difference above about $10^{-12}$ eV or eV/Å would show a compatibility gap not visible in LAMMPS output precision. Likewise, evaluate an aenet Chebyshev model with angular order beyond four and compare descriptor derivatives against direct pair-sum evaluation.","supporting_citations":[],"review_version":1}