{"id":"13f0b3bf-7820-4667-b2f6-8ffa13f6636e","arxiv_id":"2607.05015","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"dpti automates reversible HTI/TTI/pTI and Gibbs-Duhem workflows that connect analytic reference free energies to MLIP solids and liquids and produce phase boundaries with error estimates.","lead":"dpti is an open-source Python package that automates thermodynamic-integration free-energy and phase-boundary calculations for materials modeled by machine-learning interatomic potentials. It turns the many related MD tasks, path design, error checks, and Gibbs-Duhem tracing into JSON-driven workflows, demonstrated on silica and water.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper is a clear methods/software contribution whose strongest claim is operational (JSON-driven generation, execution, free-energy assembly with error bars, and GDI propagation for the supported classes). Evidence consists of open code, pedagogical JSON examples, quantitative free-energy tables with block-averaged statistical errors, integration-error-driven refinement, and dual-anchor consistency checks that close the free-energy cycle to ~1 meV/atom. The reader correctly flags the HTI-path reliability assumption as the softest point; that assumption is necessary for any TI free-energy difference to equal the true Helmholtz difference (Theory § HTI Paths; Discussions). However, the manuscript does not leave it untested: multi-step auxiliary potentials, spring-constant sensitivity, and independent-anchor agreement are supplied, and the demos recover physically plausible boundaries (silica triple point near experiment; water Tm offset attributed to the DFT functional). No stronger internal flaw (e.g., incorrect free-energy formulas in Appendices A–D, missing pV conversion, or non-reproducible automation) is evident. Therefore the ACCEPT / high-confidence verdict stands; the concrete spring-constant re-run is a low-cost verification that would still be worth performing but is not expected to overturn the claim.","tokens_in":27589,"tokens_out":699,"duration_ms":10588,"concrete_test":"Reproduce the refined coesite HTI free energy at 1600 K / 5 GPa (Table 5) from the public example inputs by (i) running the one-step path with spring_k = 0.15 and (ii) repeating with spring_k = 0.25; if |ΔG| exceeds the reported <0.4 meV/atom spring-constant sensitivity (Fig. S1) or the dual-anchor TTI consistency offset (~1 meV/atom), the path-reliability claim weakens for that solid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that dpti automates equilibrium TI (HTI + TTI/pTI + GDI) for MLIP atomic solids/liquids and water, producing free energies and phase boundaries from JSON inputs with statistical/integration-error estimates, as shown on silica and ice Ih–water. The reader’s weakest assumption (predefined multi-step HTI paths remain reversible and inside the MLIP’s reliable domain) is real in principle, but the manuscript already mitigates it: soft-core LJ intermediates (fitted for melt; auxiliary angle/bond for water), explicit path-independence consistency checks via dual HTI anchors + TTI/pTI (Figs. 6–7, 13; SI S2–S4), adaptive grid refinement with curvature-based integration-error estimates (Eqs. 26–27; Fig. 5; Table 5), and open example inputs/code. No internal inconsistency or hidden circularity appears; demos report meV-scale statistical errors that propagate sensibly into coexistence uncertainties, and scope limits (no auto mixing free energy; molecular crystals future) are stated. The automation claim therefore holds for the stated systems without a load-bearing failure mode that the evidence fails to address.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces dpti, an open-source Python package that automates equilibrium thermodynamic-integration workflows for phase-diagram calculations with machine-learning interatomic potentials. It implements Hamiltonian TI from analytically known reference states (Einstein crystal / molecule for solids, ideal gas or ideal molecules for liquids, with water-specific bond/angle auxiliaries), temperature and pressure TI to propagate Gibbs free energies, and Gibbs–Duhem integration to trace coexistence lines. Given JSON inputs, the package generates and dispatches LAMMPS MD tasks, evaluates free-energy contributions, reports statistical and numerical-integration errors, and supports adaptive λ-grid refinement. Two Deep Potential demonstrations—silica (β-quartz–coesite–melt) and ice Ih–liquid water—document per-contribution free energies, dual-anchor consistency checks, coexistence uncertainties, and assembled local phase boundaries.","tokens_in":27935,"tokens_out":1111,"duration_ms":18939,"significance":"If the automation works as claimed for the stated system classes, dpti fills a practical gap between general free-energy toolkits (e.g., CALPHY’s nonequilibrium framework, PLUMED, OpenFE) and the growing need for reproducible MLIP phase diagrams. Strengths that raise the contribution above a pure software note include: explicit Frenkel/Vega and water reference free energies (Appendices A–D), open example inputs and code, dual-anchor TTI/pTI path-independence checks, curvature-based integration-error estimates with adaptive refinement (Eqs. 26–27; Fig. 5; Table 5), and meV-scale free-energy uncertainties that propagate sensibly into coexistence errors. Prior applications to tin, lithium, minerals, and water further support utility. Scope limits (no automatic mixing free energy; molecular crystals not yet supported) are stated clearly.","major_comments":[{"comment":"Theory, “HTI Paths from Reference to Target States,” and Discussions: the central automation claim for liquids and ice relies on predefined multi-step soft-core/spring paths remaining reversible and inside the MLIP’s reliable domain. The silica and water demos mitigate this with dual-anchor consistency (Figs. 6–7, 13; SI S2–S4) and grid refinement, but the manuscript does not specify a routine user-facing check (e.g., force/energy outlier rates or reverse-path tests along λ) for a new MLIP. Adding a short recommended validation protocol would make the general-MLIP claim load-bearing rather than example-dependent.","section":null},{"comment":"Example: Silica, Stage 3 (melt) and Table 3: soft-core LJ ε,σ are obtained by an offline weighted energy/force fit that is not part of the automated dpti workflow. For atomic liquids this fitting step is effectively a free parameter of the HTI path. The paper should either document a minimal automated or semi-automated fitting utility, or state more explicitly that users must supply validated soft-core parameters and how sensitive G is to those parameters beyond the single silica case.","section":null}],"minor_comments":[{"comment":"Introduction and Discussions: CALPHY is cited as the closest automated free-energy tool; a short table or paragraph contrasting equilibrium TI (dpti) vs nonequilibrium/adiabatic-switching (CALPHY) for MLIP phase diagrams would help readers choose tools.","section":null},{"comment":"Eq. (26): the local integration-error estimate uses max absolute curvature from adjacent three-point stencils. A one-sentence note on known limitations (e.g., underestimation for non-smooth integrands near λ endpoints) would clarify how users should interpret ε_tot.","section":null},{"comment":"Fig. 10(a): the text correctly notes that triple-point error bars omit GDI accumulation and are therefore underestimated; consider stating this also in the figure caption.","section":null},{"comment":"Software Usage / Table S1: several JSON keys (protect_eps, soft_param activation, copies) appear only in SI; a one-line pointer in the main text to Table S1 after the first JSON mention would improve usability.","section":null},{"comment":"Typographical consistency: “dpti” is sometimes run-on with following words in the abstract/intro (e.g., “dpticonnects”, “dptiprovides”); fix spacing in the compiled PDF.","section":null},{"comment":"Appendix D, Eq. (39): the Gaussian bond integral form is standard; citing the same expression used in the prior DP water phase-diagram work more explicitly would help readers cross-check Amol_0.","section":null}],"recommendation":"minor_revision","confidential_remarks":"This is a solid methods/software paper for physics.comp-ph or a computational-materials methods venue. Novelty is primarily engineering and workflow completeness rather than new free-energy theory; that is appropriate if the journal accepts software-focused contributions. I see no circularity or hidden normalization to target diagrams. The two major comments are fixable with text and modest tooling documentation; I would not require new production phase diagrams for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a methods package paper, not a new free-energy theory. What is new is a purpose-built open workflow (dpti) that turns equilibrium TI for MLIPs into JSON-driven task generation, multi-step HTI paths for atomic solids/liquids and water, adaptive λ-grid refinement, statistical plus integration-error estimates, and GDI phase-boundary assembly. That is distinct from CALPHY’s nonequilibrium/LAMMPS focus and from the earlier manual DP phase-diagram work they cite.\n\nThey do the hard parts carefully. Theory restates Frenkel/Vega corrections, ideal-gas and molecular water references, and Clausius–Clapeyron GDI without inventing entities. The silica and ice Ih–water demos report free-energy contributions piece by piece, meV-scale statistical errors, grid-refinement ΔG, dual-anchor TTI/pTI consistency, and triple-point mutual agreement within the stated uncertainties. Code, example inputs, and prior applications are public. Scope limits are explicit: atomic systems plus water; no automatic mixing free energy; molecular crystals left for later.\n\nThe soft spot is real but already on the page: predefined multi-step HTI paths must stay reversible and inside the MLIP’s reliable domain. Soft-core LJ intermediates, water bond/angle auxiliaries, path-independence checks, and curvature-based refinement mitigate that for the systems they claim. Free parameters (spring κ, soft-core LJ, water springs, λ grids) are standard TI knobs, not hidden fitting of the phase boundaries. Citation pattern is appropriate; self-cites are prior use cases of the same line of work.\n\nWho gets value: people running MLIP phase diagrams for minerals, metals, and water who currently hand-assemble dozens of related MD tasks. Not a field-reorganizing advance, but a practical standard-candidate for that subfield. Math and data look solid for a computational-methods contribution. I would send it to peer review and engage with the package if I needed these calculations.","headline":"Solid methods/software paper: open equilibrium-TI automation for MLIP phase diagrams, with real demos and honest scope limits.","tokens_in":28517,"tokens_out":511,"would_cite":true,"duration_ms":5051,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"dpti automates thermodynamic integration so machine-learning potentials can yield free energies and phase boundaries without hand-built molecular-dynamics workflows.","keywords":["thermodynamic integration","machine learning interatomic potentials","phase diagrams","free energy","molecular dynamics","Gibbs-Duhem integration","silica","water"],"falsifier":"Recompute the same solid or liquid free energy with two independent HTI path designs (for example one-step versus three-step, or Frenkel versus Vega) and with two spring-constant or soft-core choices; if the Gibbs free energies differ by more than the reported statistical and integration errors, path independence fails for that system.","tokens_in":28521,"feed_emoji":"🌡️","tokens_out":928,"duration_ms":22430,"temperature":0.7,"pith_summary":"Thermodynamic integration is a standard way to get free energies and phase diagrams, but with machine-learning interatomic potentials it is hard: reversible paths must be designed carefully, many related molecular-dynamics runs must be managed, and small free-energy errors can move coexistence lines. This paper presents dpti, an open-source package that turns those steps into a JSON-driven workflow. It connects reference systems with known free energies to target solids and liquids through multi-step Hamiltonian paths, then propagates free energies in temperature and pressure and traces boundaries with Gibbs-Duhem integration. The authors demonstrate the pipeline on silica (beta-quartz, coesite, melt) and on ice Ih versus liquid water. A reader who needs reproducible free-energy phase diagrams from modern potentials would care because the package automates task generation, error estimation, grid refinement, and boundary propagation that are otherwise tedious and fragile.","feed_headline":"JSON workflow turns ML potentials into phase diagrams","feed_subtitle":"dpti links known reference free energies to solids, liquids and water, then traces coexistence with quantified errors.","key_machinery":"A five-stage thermodynamic-integration workflow: NpT/NVT equilibration, Hamiltonian TI from analytical references to the machine-learning target, temperature or pressure TI to locate crossings, and Gibbs-Duhem integration to trace boundaries, using multi-step reversible paths (springs, soft-core, molecular restraints) plus adaptive grid refinement and error propagation.","core_discovery":"Given JSON inputs, dpti generates and runs the molecular-dynamics tasks for thermodynamic integration with machine-learning potentials, computes free-energy contributions with statistical and integration-error estimates, and propagates coexistence points into phase boundaries for atomic solids, atomic liquids, and water, as shown for silica and ice–water.","pith_inferences":["Routine free-energy phase diagrams could become a standard validation test for new machine-learning potentials rather than a specialist project.","The water and ice reference constructions offer a template for other small molecular liquids if analogous bond and angle restraints are coded.","Finite-size corrections and substitutional mixing free energies still sit outside the automation, so alloy and disordered-solid diagrams will remain hybrid workflows.","Path reliability still depends on human choice of spring constants and soft-core parameters; automated path-validation diagnostics would be a natural next layer."],"forward_implications":["Absolute Gibbs free energies of machine-learning solids and liquids can be obtained from JSON specifications without hand-writing large sets of related MD inputs.","Statistical and quadrature errors are reported and used to refine integration grids, so free-energy uncertainty is quantified rather than guessed.","A single coexistence point plus phase-pair NpT runs can be expanded into local phase boundaries and assembled into multi-phase diagrams, as for silica’s triple-point region.","Dual-anchor consistency checks (two HTI temperatures or pressures, then TTI or pTI) become routine and can expose irreversible paths or under-sampling.","The same automation applies to other materials once suitable reference states and reversible paths exist for those system classes."],"fun_headline_variants":["dpti automates TI paths from JSON to MLIP phase boundaries","JSON inputs drive free-energy TI and coexistence for ML potentials","Open-source dpti links reference free energies to MLIP solids and liquids","Automated MD tasks compute MLIP phase diagrams with error estimates","dpti traces silica and ice–water boundaries via ML thermodynamic integration"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The package’s fixed multi-step integration paths must keep every intermediate state reversible and inside the region where the machine-learning potential remains accurate; if not, the integrated free-energy difference is not the true free-energy difference.","fun_headline_variants_meta":{"raw":{"variants":["dpti automates TI paths from JSON to MLIP phase boundaries","JSON inputs drive free-energy TI and coexistence for ML potentials","Open-source dpti links reference free energies to MLIP solids and liquids","Automated MD tasks compute MLIP phase diagrams with error estimates","dpti traces silica and ice–water boundaries via ML thermodynamic integration"]},"model":"grok-4.5","effort":"low","cost_usd":0.00328,"raw_usage":{"total_tokens":1096,"prompt_tokens":728,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":32800000,"prompt_tokens_details":{"text_tokens":728,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":276,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":728,"tokens_out":92,"duration_ms":3113,"temperature":1.0,"reasoning_tokens":276,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T10:04:16.987376+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Recompute the same solid or liquid free energy with two independent HTI path designs (for example one-step versus three-step, or Frenkel versus Vega) and with two spring-constant or soft-core choices; if the Gibbs free energies differ by more than the reported statistical and integration errors, path independence fails for that system.","supporting_citations":[],"review_version":1}