{"id":"102332d1-aef3-409b-9bf6-8fa6cd77ffa6","arxiv_id":"1908.02038","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"An improved FIRE minimizer (fire 2.0) with half-step backtracking and symplectic integration is implemented in LAMMPS and shown to outperform standard FIRE and conjugate gradient on eight materials science benchmarks.","lead":"This paper presents an improved version of the FIRE energy-minimization algorithm, called fire 2.0, and its implementation in the LAMMPS molecular dynamics code. It shows that the choice of time integrator and minimization parameters strongly affects how quickly atomistic simulations reach low-energy states, and that fire 2.0 often beats the standard FIRE and conjugate-gradient methods on materials science test problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Parameter recommendations rely on one test case and speedup ratios use case-specific stopping points, so the claimed robustness of fire 2.0 is not yet established.","rationale":"The paper's core algorithmic contribution is credible: it clearly identifies the integrator dependence of FIRE, provides a LAMMPS implementation with a public commit, and shows consistent gains over the previous LAMMPS implementation on eight diverse problems. Independent work by Shuang et al. also supports the integrator recommendation. However, the load-bearing empirical claims are conditional. The parameter recommendations are the practical output of the paper, and they are derived from a single case with an unproven assertion of problem-independence. The speedup ratios are not all measured to a common convergence threshold, and the 'lower energy' claim lacks energy data. These issues do not invalidate the algorithm, but they justify a conditional verdict pending release of benchmark inputs, additional parameter sweeps, and a uniform stopping-criterion analysis. The reader's weakest assumption already identified the representativeness and single-case parameter study; my analysis agrees and sharpens it by pointing to the heterogeneous stopping points and the missing energy comparison.","tokens_in":17887,"tokens_out":7420,"duration_ms":78704,"concrete_test":"Run the alpha0/tmax sweeps from Fig. 3 on at least two additional representative cases: case 1 (EAM dislocation in Al) and case 6 (MEAM dislocation-precipitate in Mg), using the same protocol. If the optimal ranges for alpha0 and tmax fall outside 0.10–0.25 and 2–12 in either case, the Section 5.4.4 generalization fails. Separately, re-express Table 2 using a single target f2norm for all algorithms (or explicitly labelling failures) and report the final potential energies for case 6 for fire, CG, and fire 2.0. If the speedup ratios change materially or the energy ordering reverses, the Summary's claims need qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that fire 2.0 is significantly faster than fire or CG and can find lower-energy structures — is supported by Table 2 and Fig. 2, but the evidence is less uniform than the Summary suggests. In cases 3 and 4, CG never reaches the stated f2norm threshold (10^-8 eV/A); the speedup is measured at the lowest f2norm CG achieved, not at a common convergence target. In case 2, fire 2.0 itself stops at a plateau via the MAXVDOTF criterion, and the speedup vs. fire is defined by matching that plateau, not by reaching the threshold. Thus the quoted ratios are not all 'force evaluations to reach target tolerance', which weakens the quantitative force of 'significantly faster'. More importantly, Section 5.4.4 states 'the observed trends do not depend on the problem' and then generalizes alpha0 in 0.10–0.25 and tmax in 2–12, but the parameter study was run only on case 5 (Si vacancies, SW potential). No sweeps are shown for EAM, BKS, or MEAM systems, or for NEB cases, and no repeats or error bars are reported. Since the raw data are withheld ('cannot be shared at this time'), this generalization is unfalsifiable from the paper alone. Finally, the 'lower energy structure' claim rests on case 6, where fire 2.0 places the dislocation differently than CG/fire; the paper reports this qualitatively but gives no final potential energies, so 'lower energy' is inferred from lower force norms rather than demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FIRE 2.0, a modified version of the FIRE energy-minimization algorithm, and its implementation in LAMMPS. The main algorithmic changes are the use of semi-implicit Euler or Velocity-Verlet integration instead of the explicit Euler integrator used by the earlier LAMMPS implementation, a half-step backtracking correction after uphill motion is detected, a delay in adjusting timestep and mixing factor, and an additional stopping criterion based on repeated downhill failures. The authors benchmark FIRE 2.0 against conjugate gradient and the standard LAMMPS FIRE on eight atomistic test cases spanning EAM, MEAM, SW, and BKS force fields, including NEB path optimizations, and report speedups measured in force-evaluation counts. They recommend specific parameter ranges (alpha0 in 0.10–0.25, tmax in 2–12) and conclude that FIRE 2.0 is significantly faster than FIRE and CG and can find lower-energy structures.","tokens_in":18234,"tokens_out":2830,"duration_ms":29370,"significance":"If the performance claims hold, the paper has clear practical value for the atomistic simulation community: it provides a drop-in improved minimizer for LAMMPS, gives explicit guidance on time integration for FIRE-style methods, and documents the algorithm at the level needed for reproduction. The source code is made available, and the benchmark methodology using force-evaluation counts is a standard and reasonable basis for comparison. The identification of the explicit Euler integrator as a major performance bottleneck in the original LAMMPS FIRE is a concrete and useful contribution. However, the strength of the general conclusions is currently limited by the uneven benchmarking protocol and the single-case parameter study.","major_comments":[{"comment":"The speedup ratios versus CG in cases 3 and 4 are not measured at a common convergence threshold: the text states that CG fails to reach the 1e-8 eV/Å f2norm target, so the comparison is made at the lowest f2norm CG achieved, while FIRE 2.0 continues to the full target. Similarly, in case 2 the speedup versus FIRE is defined by matching the MAXVDOTF plateau rather than the threshold. Because the quoted ratios therefore correspond to different stopping points, the summary claim that FIRE 2.0 is 'significantly faster' than CG and FIRE is quantitatively underdetermined. Please report, for every case and every method, the final f2norm value, the number of force evaluations, and the stopping criterion used, and recompute ratios at a common target wherever possible.","section":"Table 2 and Section 5.4.1"},{"comment":"The statement that 'the observed trends do not depend on the problem' is not supported by the evidence shown. The parameter sweeps for alpha0 and tmax are performed only on case 5 (vacancies in Si with the SW potential); no sweeps are shown for EAM, BKS, MEAM, or NEB cases, and no repeat runs or error bars are reported. Since this single case is used to justify the general recommendations alpha0 in 0.10–0.25 and tmax in 2–12, those recommendations are not yet established for the broader class of materials-science problems the paper targets. Please run the same sweep on at least one representative case from another force-field class, or explicitly restrict the recommendation to covalent SW-type systems.","section":"Section 5.4.4 and Figure 3"},{"comment":"The claim that FIRE 2.0 'can result in lower energy structures not found by other algorithms' is not demonstrated. In case 6, the paper reports that FIRE 2.0 produces a different dislocation position than CG and FIRE, but it does not report the final potential energies of the competing configurations. A lower final force norm does not by itself imply a lower energy, and no energy comparison is provided anywhere in the paper. Please either report the final potential energies (or energy differences) for the configurations obtained by each method in case 6, or soften the claim to state that FIRE 2.0 finds qualitatively different configurations.","section":"Section 5.4.1, Figure 2(6), and Summary"}],"minor_comments":[{"comment":"The text says the evolution of f2norm is shown in Fig. 1, but the convergence curves appear in Fig. 2; please correct the cross-reference.","section":"Section 5.4"},{"comment":"The text states the parameter study used a timestep of '1 ps' while the Figure 3 caption says '1 fs'; the later discussion of 'optimum ∆t being 1 fs' suggests the caption is correct, so please fix the text.","section":"Section 5.4.4 and Figure 3 caption"},{"comment":"The statement that raw data 'cannot be shared at this time' makes it difficult to verify the reported force-evaluation counts and convergence curves; providing at least a table of final f2norm values and force-evaluation counts for all cases and methods would improve reproducibility.","section":"Data availability"},{"comment":"The terms 'Euler Implicit' and 'Euler Semi-implicit' are used interchangeably in places (e.g., Section 5.4.3 vs. Section 5.4.4); please use one consistent name to avoid confusing the semi-implicit Euler scheme with a fully implicit integrator.","section":"Section 3.1 and Algorithms 3–6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reports a genuinely useful algorithmic improvement and provides code, but the quantitative evidence needs strengthening before publication. The speedup comparisons at non-common stopping thresholds and the single-case parameter study are the two load-bearing issues; both are fixable within the scope of a revision. I would not reject, but I would not accept in current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before recommending FIRE to anyone doing atomistic minimization. It does three concrete things: it specifies three algorithmic modifications to FIRE (half-step backtracking after uphill motion, delayed timestep growth, and moving the velocity/force mixing to the end of the step), it compares four integrators, and it ships a public LAMMPS implementation with a commit hash. The algorithm descriptions are clear enough to reimplement, and the eight benchmark cases cover EAM, MEAM, SW, and BKS potentials, including surfaces, dislocations, melts, and NEB. That is real, citable work.\n\nThe integrator finding is the strongest part: explicit Euler is bad, semi-implicit Euler is robust, and Velocity Verlet is usually a bit better but sometimes unstable. This overlaps with Shuang et al. (2019), and the paper says so. The LAMMPS implementation is a practical contribution on its own, and the additional MAXVDOTF stopping criterion is a sensible safeguard.\n\nWhere the evidence wobbles: the recommended ranges for alpha0 and tmax come from a parameter sweep on only case 5 (Si vacancies, SW potential), yet the text claims the trends do not depend on the problem. That generalization is unsupported; no sweeps are shown for EAM, MEAM, or BKS. The speedup ratios in Table 2 are also not all measured to the same tolerance. In cases 3 and 4, CG stalls at a higher f2norm and the comparison uses the lowest f2norm CG reaches; in case 2, fire 2.0 itself stops at a plateau. The table and text are transparent about this, so the numbers are not misleading, but the headline claim of \"significantly faster\" is softer than it looks at first glance. Finally, the claimed lower-energy structure in case 6 (dislocation near a precipitate) is reported only as a different configuration; no final potential energies are given, so \"lower energy\" is inferred, not demonstrated. No repeats or error bars are provided anywhere, and the raw data are explicitly withheld.\n\nNone of this is fatal. The algorithmic ideas are sensible, the implementation is available, and the benchmark suite is reasonably broad. The paper deserves a serious referee, but a good referee should ask for the input files, a parameter sweep on at least one non-SW system, and final energies for case 6. I would cite this if I used FIRE in LAMMPS.","headline":"A genuinely useful methods paper: FIRE 2.0 is clearly specified, implemented in LAMMPS, and mostly faster than FIRE/CG, but the parameter recommendations rest on a single test case and a few speedup ratios are measured at non-identical stopping points.","tokens_in":18780,"tokens_out":1543,"would_cite":true,"duration_ms":16642,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K05","65K10"],"pacs":["02.70.Ns"],"model":"deepseek-v4-flash","headline":"A modified fast inertial relaxation engine, fire 2.0, reaches a fixed force threshold in fewer force evaluations than FIRE in all eight test cases and beats conjugate gradient in five of six direct minimizations.","keywords":["FIRE algorithm","energy minimization","atomistic simulation","LAMMPS","time integration","conjugate gradient","nudged elastic band","dislocation relaxation"],"falsifier":"Relax the Mg dislocation–precipitate system (case 6) with the dislocation started at several distances from the precipitate, using the same $f_{2\\,\\mathrm{norm}}=10^{-8}$ eV/$\\AA$ stopping criterion for FIRE, CG, and fire 2.0. The claim that fire 2.0 reaches lower-energy structures would fail if fire 2.0 does not consistently move the dislocation toward the precipitate while the other algorithms do, or if conjugate gradient with different line-search settings reaches that configuration in no more force evaluations.","tokens_in":17727,"feed_emoji":"⚛️","tokens_out":15747,"duration_ms":143196,"temperature":0.7,"pith_summary":"The paper sets out to show that the fast inertial relaxation engine, a pseudo-dynamics minimizer widely used in atomistic simulation, is limited less by its mixing idea than by how it is integrated and how it handles overshoot. It presents fire 2.0, a LAMMPS implementation that switches from explicit Euler to semi-implicit Euler or Velocity Verlet, moves the configuration back by half a timestep whenever the discrete power $P = F \\cdot v$ turns non-positive, and delays the timestep and mixing adjustments after such events. On eight materials-science test cases covering dislocations, surfaces, ionic melts, and nudged elastic band paths, the paper reports that fire 2.0 reaches the chosen force threshold in fewer force evaluations than the standard FIRE in every case, and beats conjugate gradient in five of six direct minimizations. If this transfers to other systems, routine energy minimization of defects and reaction paths becomes faster and less likely to stall in shallow valleys.","feed_headline":"Energy minimization runs up to 30x faster with FIRE 2.0","feed_subtitle":"The upgraded LAMMPS minimizer finds lower-energy states in fewer force evaluations.","key_machinery":"The load-bearing mechanism is the half-step back correction. FIRE accelerates along the force direction but must detect the turn where the power $P = F \\cdot v$ changes sign; by the time a discrete step sees $P \\le 0$, the trajectory has overshot uphill. FIRE 2.0 responds with $x(t) \\leftarrow x(t) - 0.5 \\, \\Delta t \\, v(t)$ before setting $v = 0$, which counters most of that overshoot while staying closer to the turning point than a full-step reversal. The other essential component is the time integrator: with explicit Euler the corrected algorithm behaves like steepest descent, whereas semi-implicit Euler or Velocity Verlet make the variable-timestep descent robust, and the paper identifies this choice as the most important parameter affecting FIRE's performance.","core_discovery":"fire 2.0 is a modification of the FIRE algorithm built on four changes: a symplectic integrator (semi-implicit Euler or Velocity Verlet) instead of the explicit Euler used by LAMMPS's standard FIRE; a half-step backward coordinate correction when the discrete power $P = F \\cdot v$ becomes non-positive, before velocities are zeroed; a short delay before timestep growth and mixing-coefficient decay resume after a negative-power step; and a stopping rule that exits after too many consecutive negative-power steps. With these changes the paper reports that fire 2.0 reaches the target force norm in fewer force evaluations than standard FIRE in all eight tests, with finite speedups from 1.8x to more than 30x, and outperforms conjugate gradient in five of six direct minimizations, the exception being the non-periodic nanoporous gold pillar. In the Mg dislocation–precipitate test, FIRE and CG leave the dislocation in the matrix while fire 2.0 relaxes it toward the precipitate; the paper presents this as evidence that fire 2.0 can reach lower-energy structures that other algorithms do not find.","pith_inferences":["The half-step back correction is a generic remedy for discrete-time overshoot, so a natural (untested) extension is to apply the same correction to other damped-dynamics minimizers such as Quickmin and see whether they gain similarly.","Force-evaluation counts ignore communication and overhead, so on massively parallel machines the wall-clock speedup of fire 2.0 over FIRE could be smaller than the force-evaluation ratio; benchmarking wall-clock time on target hardware would settle this.","Since the parameter sweep was run on one system only, the recommended alpha0 and tmax ranges may need case-specific retuning for energy landscapes with very different curvature scales; this is an inference beyond the paper's own generalization.","The case-6 result implies that a published relaxed structure can depend on the minimizer, not just the force field and threshold; reporting the minimizer, integrator, and force norm alongside structures would make such results reproducible."],"forward_implications":["LAMMPS users can expect fire 2.0 to reach a given force norm in fewer force evaluations than the standard FIRE on typical defect, surface, and NEB relaxation tasks; the reported gain grows with the complexity of the NEB path.","On systems with long-range electrostatics or long-range elastic fields, conjugate gradient can terminate early with 'line search alpha is zero' while fire 2.0 continues to reduce the forces, making fire 2.0 a more reliable choice for such problems.","The time integration scheme is the dominant algorithmic choice: switching from explicit Euler to semi-implicit Euler is enough to change FIRE from a steepest-descent-like method into a robust minimizer.","Users should cite the exact force norm in published results, because LAMMPS's Euclidean-norm criterion is several orders of magnitude stricter than maximum-force-component criteria used in some other codes.","As default recommendations, alpha0 should lie in 0.10–0.25 and tmax in 2–12, with the simulation timestep set to the same value used in low-temperature MD; reducing tmax improves stability."],"supporting_citations":[{"why":"Introduces the original FIRE algorithm whose velocity–force mixing and adaptive timestep fire 2.0 modifies, and provides the baseline FIRE behavior used in the comparisons.","marker":"[17]"},{"why":"The LAMMPS code in which both the existing FIRE and the new fire 2.0 are implemented; the paper's speedups are measured in LAMMPS force evaluations.","marker":"[9]"},{"why":"Prior study of how integration formulations affect FIRE performance; motivates the paper's integrator comparison and independently recommends semi-implicit Euler.","marker":"[31]"},{"why":"The NEB framework used in test cases 7 and 8, which requires damped-dynamics minimizers and therefore sets the comparison basis for those cases.","marker":"[7]"},{"why":"Mishin EAM potential for aluminum used in the edge-dislocation relaxation (case 1) and vacancy-migration NEB (case 7) tests.","marker":"[65]"},{"why":"Stillinger–Weber potential for silicon used in the vacancy relaxation (case 5), the case chosen for the parameter study.","marker":"[51]"},{"why":"BKS potential for silicate glass with long-range Coulomb interactions in case 2, where CG fails and fire 2.0 reaches a plateau.","marker":"[58]"},{"why":"Kim MEAM potential for Mg–Al used in the dislocation–precipitate case 6, where fire 2.0 finds a different dislocation position.","marker":"[68]"},{"why":"Kim MEAM potential for Mg–Ca used in the synchroshear NEB case 8, where fire 2.0 is 2.9x faster than FIRE.","marker":"[49]"},{"why":"EAM potential for gold used in the nanoporous bulk and pillar cases 3 and 4, including the one case where CG beats fire 2.0.","marker":"[66]"}],"fun_headline_variants":["FIRE 2.0: up to 30x faster energy minimization in LAMMPS","New FIRE algorithm finds deeper minima, runs up to 30x faster","LAMMPS gets FIRE 2.0: 30x speedup, better minima","Improved FIRE for LAMMPS: fewer force calls, lower-energy results","FIRE 2.0 vs CG: faster and reaches lower states in atomistic runs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance claims rest on the assumption that the eight chosen test systems and the comparison by force-evaluation counts represent typical energy-minimization problems; in particular, the paper's recommended parameter ranges are extrapolated from a parameter sweep run on only one of the eight cases (silicon vacancies).","fun_headline_variants_meta":{"raw":{"variants":["FIRE 2.0: up to 30x faster energy minimization in LAMMPS","New FIRE algorithm finds deeper minima, runs up to 30x faster","LAMMPS gets FIRE 2.0: 30x speedup, better minima","Improved FIRE for LAMMPS: fewer force calls, lower-energy results","FIRE 2.0 vs CG: faster and reaches lower states in atomistic runs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000761,"raw_usage":{"total_tokens":3332,"prompt_tokens":851,"completion_tokens":2481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":2366}},"tokens_in":467,"tokens_out":2481,"duration_ms":74383,"temperature":1.0,"reasoning_tokens":2366,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:55:17.417878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Relax the Mg dislocation–precipitate system (case 6) with the dislocation started at several distances from the precipitate, using the same $f_{2\\,\\mathrm{norm}}=10^{-8}$ eV/$\\AA$ stopping criterion for FIRE, CG, and fire 2.0. The claim that fire 2.0 reaches lower-energy structures would fail if fire 2.0 does not consistently move the dislocation toward the precipitate while the other algorithms do, or if conjugate gradient with different line-search settings reaches that configuration in no more force evaluations.","supporting_citations":[{"cited_title":"Shuang, P","cited_arxiv_id":null,"evidence_quote":"Prior study of how integration formulations affect FIRE performance; motivates the paper's integrator comparison and independently recommends semi-implicit Euler."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Kim MEAM potential for Mg–Al used in the dislocation–precipitate case 6, where fire 2.0 finds a different dislocation position."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Kim MEAM potential for Mg–Ca used in the synchroshear NEB case 8, where fire 2.0 is 2.9x faster than FIRE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EAM potential for gold used in the nanoporous bulk and pillar cases 3 and 4, including the one case where CG beats fire 2.0."}],"review_version":1}