{"id":"d57ad893-73ad-474b-8cf1-164482fd63d2","arxiv_id":"1908.01610","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Parametric constraints on atomic and lattice degrees of freedom, generated from crystal prototypes, preserve or selectively break symmetry during DFT relaxations, reducing average relaxation steps by 33 to 54 percent on a 359-material benchmark.","lead":"This paper presents a new way to speed up computer simulations of crystal structures by using symmetry-based rules to reduce the number of variables that are optimized. The method is implemented in two simulation tools and shown to cut the number of relaxation steps by an average of 33 to 54 percent, and up to 96 percent for a 216-atom defect supercell, while also reaching metastable phases that ordinary relaxations miss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Step-count savings may compare different stopping criteria: constrained runs stop on projected forces, free runs on full forces, which could inflate the reported average savings.","rationale":"In good faith, the paper's conceptual contribution is substantial: a parametrized constraint formalism implemented in FHI-aims and ASE, exact symmetry preservation by construction, and credible demonstrations on ZrO2 and Bi2O3. The AFLOW prototype mapping and the 359-material benchmark are independent support. My concern targets only the headline step-count claim, not the method's existence. The reader identified the linear-subspace spanning assumption as weakest; I agree that is a genuine limitation, and the MgO energy offset is acknowledged. But the more load-bearing issue for the central acceleration claim is the apparent mismatch between the constrained and free convergence metrics. If the constrained runs stop on projected forces while the free runs stop on full forces, the savings in Eq. (12) are not apples-to-apples; this could inflate all the quoted percentage reductions and even the polaron factor-of-24 claim. The test is simple and should be required for acceptance. Because the data are public and the concern is falsifiable, I do not move the reader's verdict; I keep it CONDITIONAL and add this specific condition.","tokens_in":17992,"tokens_out":5646,"duration_ms":60203,"concrete_test":"Use the archived NOMAD dataset for all 359 benchmark materials. For each constrained-relaxation endpoint, run a single FHI-aims SCF at the reported geometry (or read the last SCF from outputs) and compute max|F_i| and max stress component in full Cartesian space. Compare these residuals to the 0.005 eV/A threshold used for free relaxations. Then rerun the constrained relaxations with an added condition that the full-space forces and stresses also fall below the same threshold and recompute Eq. (12). If the average savings drops substantially or many endpoints fail the full-space threshold, the headline acceleration claim must be revised; if residuals are already below threshold, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim in Section III.B is the 33.11% (PBE) and 52.43% (PBEsol) reduction in relaxation steps, Eq. (12). These numbers are only meaningful if the constrained and free arms are stopped under the same convergence criterion. The text states that constrained relaxations stop when 'the total forces on the free parameters' are below 0.005 eV/A; the generalized forces in Eqs. (5c)-(5d) are projections onto the reduced parameter space. Components of the physical force or stress orthogonal to this subspace are not part of the stopping condition, and the paper does not report full-space residuals at the constrained endpoints. At an identical threshold, a constrained endpoint can therefore have finite unrelaxed full forces that a free relaxation would continue to minimize. The MgO polaron results show the bias is real in at least one case: the constrained minimum lies 69-78 meV above the free minimum (Section III.C). For the benchmark, structure matching via AFLOW-XTAL-MATCH uses tolerances (m<=0.1) and does not quantify force residuals, so 'same final structure' does not remove this issue. If the constrained runs are stopped early with respect to full forces, part or most of the claimed step savings is an artifact of comparing different stopping rules. This is directly testable because the full input/output data are archived in NOMAD.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a parametrically constrained geometry-relaxation scheme for DFT, in which fractional atomic coordinates and lattice vectors are mapped through linear Jacobians onto a low-dimensional parameter space derived from AFLOW crystal prototypes, with forces and stresses transformed to generalized forces in that space. The authors implement the scheme in FHI-aims, demonstrate it on metastable ZrO2 and Bi2O3 polymorphs, benchmark it on 359 materials across 13 prototypes, and apply it to a polaronic distortion in MgO. The central claims are that the constraints exactly preserve symmetry, reduce the number of relaxation steps by about 33% (PBE) and 52% (PBEsol) for converging cases, and cut the steps for a 216-atom polaron supercell by 96% relative to free relaxation, at the cost of a small (69-78 meV) energy bias in the MgO case.","tokens_in":18327,"tokens_out":6397,"duration_ms":66740,"significance":"If the quantitative claims are sound, the method is a useful and general tool for high-throughput materials screening and for defect and metastable-phase calculations, because it automates symmetry-preserving or symmetry-breaking constraints and couples them with existing optimizers. The paper's data and code availability are concrete strengths: the complete input/output files are deposited in NOMAD (DOI:10.17172/NOMAD/2019.10.19-1), the implementation is in a released version of FHI-aims, and an ASE implementation is referenced. The central issue is whether the reported step-count savings are a fair comparison of equivalently converged calculations; this must be resolved before the headline numbers can be accepted.","major_comments":[{"comment":"The claimed average savings compare free relaxations stopped on full Cartesian force components with constrained relaxations stopped on generalized forces in the reduced parameter space. In the workflow of Fig. 2, the convergence check is applied to the forces on the active degrees of freedom, and in Section III.C the MgO runs explicitly stop when 'the total forces on the free parameters' are below threshold. Because the generalized forces in Eqs. (5c)-(5d) are projections, components of the physical forces and stress orthogonal to the constraint subspace are never monitored; at the same threshold value, a constrained endpoint may therefore have large full-space residuals that a free relaxation would continue to reduce. The MgO results, where the constrained minima are 69.1-78.4 meV above the free minima (Section III.C), show that this bias can be numerically significant. The authors should report the full-space force/stress residuals at the constrained endpoints for the benchmark set (the NOMAD archive makes this possible) and should quote step savings only for cases where the endpoints are equivalently converged under the full-space criterion.","section":"III.B, Eq. (12), and Fig. 2"},{"comment":"Equation (12) defines S = (N_free - N_constrained)/N_constrained x 100%, but the text interprets S as a percentage reduction in relaxation steps. This is internally inconsistent: for the 216-atom MgO case, Section III.C reports a 96% reduction, which corresponds to (234-10)/234 = 95.7% with the free count in the denominator, whereas Eq. (12) gives 2240%. With Eq. (12), the reported averages of 33.11% and 52.43% correspond to conventional reductions of only 24.9% and 34.4% relative to the free relaxation. The savings definition should be changed to use N_free in the denominator, or explicitly relabeled as a speedup factor minus one, and all reported values in the abstract, Section III.B, and Table IV should be recomputed consistently.","section":"III.B, Eq. (12)"},{"comment":"The 'same final structure' classification relies on AFLOW-XTAL-MATCH with the loose tolerance m <= 0.1, which permits finite deviations in lattice vectors and atomic positions. Since the constrained stopping criterion only controls reduced-space forces, structures classified as matching can still differ substantially in their full-space force residuals: the free endpoint is fully converged, whereas the constrained one may not be. The discussion of negative-savings cases such as PtS2 (Fig. 5) attributes the extra steps to 'unproductive' or near-threshold behavior, but without full-force residuals the comparison remains ambiguous. Please provide the distribution of maximum full-space atomic forces and stress residuals at the constrained endpoints for the 359 materials, or restart each constrained endpoint unconstrained to demonstrate that no significant relaxation remains.","section":"III.B, Table IV and Fig. 5"}],"minor_comments":[{"comment":"The abstract and conclusions state that the method reduces the number of relaxation steps by about 50%, but the actual reported averages are 33.11% and 52.43% for the converging subset; the conclusions should reflect the range and the dependence on the functional and prototype.","section":"Abstract and Section III.B"},{"comment":"The column 'Average Savings' in Table IV uses the nonstandard denominator of Eq. (12); the text and tables should make the metric definition explicit and avoid terms like 'percent reduction' when the denominator is the constrained step count.","section":"III.B, Eq. (12) and Table IV"},{"comment":"The sentence 'the structures were relaxed until the total forces on the free parameters were below 10^-4 eV/A' should be clarified to state explicitly that this is a projected-force criterion and to give the corresponding full-space force residuals at the reported endpoints.","section":"III.C"},{"comment":"The generalized left inverse in Eq. (7) is defined with (A^T A)^(-1), which requires A to have full column rank; the paper should state this rank condition and comment on how rank-deficient parameterizations are handled in practice.","section":"II, Eq. (7)"},{"comment":"The paper cites AFLOW-XTAL-MATCH as 'in preperation' (Ref. [59]); since this tool is used in the benchmark to define structural matches, the authors should state which version was used and provide the matching parameters in the archive, so that the classification is reproducible.","section":"IV, Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the implementation is carefully described, with good data availability. The main obstacle is the comparability of convergence criteria between the constrained and free relaxations, which directly affects the headline savings numbers. The savings definition in Eq. (12) is also misleading as stated and should be corrected. These issues are addressable within the scope of the manuscript, so I do not recommend rejection, but the current version should not be accepted until the benchmark comparison is made fair and the reported numbers are recomputed. I also note that the heavy reliance on AFLOW tools and the in-preparation XTAL-MATCH reference is a mild self-citation burden, though it does not lead to circularity in the central result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid methods paper. The parametric constraint scheme is genuinely useful: it maps full atomic and lattice degrees of freedom onto a low-dimensional linear subspace defined by AFLOW prototypes, transforms forces and stresses with the Jacobian, and preserves symmetry exactly by construction. The implementation is in FHI-aims and ASE, and the full input/output dataset is archived in NOMAD. That is real, checkable work, and the zirconia and bismuth-oxide demonstrations show a clear benefit for metastable and dynamically stabilized phases.\n\nWhat the paper does well: the linear algebra is standard but applied cleanly; the benchmark covers 359 materials across 13 prototypes; the authors report divergent cases and even negative cases where constrained relaxation takes more steps; and they explicitly acknowledge that the MgO constrained minima lie 69–78 meV above the free minima. That level of transparency is not typical and deserves credit.\n\nThe soft spot is the central quantitative claim. The constrained runs stop when forces on the free parameters fall below threshold, and those are projected forces. The free runs stop on full forces. For a constrained endpoint, components orthogonal to the constraint subspace are not part of the stopping condition, and the paper does not report full-space residuals at those endpoints. So the reported 33–53% average step savings, and the 96% saving on the 216-atom MgO cell, may partly be an artifact of comparing two different stopping rules. The MgO energy gap shows the bias is real in at least one case. XTAL match with m ≤ 0.1 confirms structural similarity but not force convergence. This is directly testable because the NOMAD dataset is available; the authors should report full-space force and stress residuals at constrained endpoints, or rerun a subset to the same full-force threshold.\n\nThere are also minor gaps: no wall-clock time, so step savings do not map directly to compute savings; the prototype selection is described but not justified as representative; and the structure-matching metric relies on an unpublished tool.\n\nBottom line: the method itself is sound, and the symmetry-preservation benefit stands independently of the step-count claim. The paper deserves peer review and will be useful to anyone doing high-throughput relaxations, metastable phase searches, or supercell defect calculations. I would not desk-reject it. I would ask for the residual analysis before accepting the speedup numbers.","headline":"A clean, useful constrained-relaxation scheme with honest benchmarks, but the headline step-count savings may be inflated by comparing projected-force stopping with full-force stopping.","tokens_in":18751,"tokens_out":2439,"would_cite":true,"duration_ms":25392,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Parametrically constrained relaxations map atomic and lattice degrees of freedom onto a lower-dimensional space derived from crystal prototypes, cutting DFT relaxation steps by an average of 33% (PBE) and 52% (PBEsol), and by 96% for a…","keywords":["parametric constraints","geometry relaxation","density functional theory","crystal prototypes","symmetry preservation","high-throughput materials discovery","lattice distortion","metastable phases"],"falsifier":"Take a material with a known equilibrium distortion that lies outside the prototype parameter space, such as a strong Jahn-Teller mode in a defective supercell, constrain the relaxation to the prototype's reduced set, and compare the final energy against an unconstrained relaxation; if the constrained minimum sits more than the target convergence tolerance above the free minimum, the subspace was too small and the scheme's savings come at an unacceptable accuracy cost.","tokens_in":17855,"feed_emoji":"⚛️","tokens_out":4420,"duration_ms":46549,"temperature":0.7,"pith_summary":"This paper establishes that geometry relaxations in density functional theory can be dramatically accelerated and made symmetry-exact by restricting the optimizer to a low-dimensional parameter space instead of the full set of atomic and lattice coordinates. The authors implement this parametric constraint scheme in an all-electron code and show that it reliably relaxes metastable and dynamically stabilized phases that free relaxations miss or distort. On a benchmark of 359 materials across 13 crystal prototypes, constrained relaxations that converge to the same final structure as free ones use about one-third fewer steps with PBE and half with PBEsol. For a 216-atom MgO polaron supercell, the constrained relaxation converges in 10 steps versus 234 unconstrained. A sympathetic reader should care because this offers a general, automated route to speed up and stabilize high-throughput materials searches and supercell defect calculations.","feed_headline":"Parametric relaxations cut DFT steps by up to 96%","feed_subtitle":"Lower-dimensional mapping keeps crystal symmetry exact and speeds high-throughput materials discovery.","key_machinery":"The central object is a linear map between the full set of fractional atomic coordinates and lattice vectors and a small vector of parameters, defined through Jacobian matrices $J_R$ and $J_L$. Coordinates are mapped into the reduced space using the generalized left inverse $A^{-1,L}=(A^TA)^{-1}A^T$, and forces and stresses are pulled back with the transposed Jacobians, $F_r=J_R^T F_R$ and $F_l=J_L^T F_L$. The Hessian is likewise transformed blockwise into reduced space so the BFGS/truncated-Newton optimizer runs entirely on the parameters. The parameter sets themselves come from the AFLOW crystallographic prototype definitions, so the procedure is automated for any structure class; additional parameters can be added to capture local distortions such as Jahn-Teller modes.","core_discovery":"The paper claims that parametrically constrained relaxations, which map atomic positions and lattice vectors onto a reduced parameter set defined by crystallographic prototypes, preserve the target symmetry exactly while substantially lowering the number of relaxation steps. The mapping is linear: fractional coordinates and lattice vectors are expressed through Jacobians, forces and stresses are pulled back into the reduced space via the transposed Jacobians, and the optimizer acts only on the few parameters. The scheme is general because the parameter sets are generated automatically from the AFLOW prototype library, and it can be extended to local distortions by adding parameters along known soft modes. The authors demonstrate that this approach converges metastable phases (e.g., cubic ZrO2 and γ-Bi2O3) that free relaxations would lose, and that it cuts the step count for relaxing a large polaronic supercell by an order of magnitude. The cost is a small energy bias when the true minimum lies outside the chosen parameter subspace, quantified as 69–78 meV for the MgO polaron tests.","pith_inferences":["The same parameter mapping could double as a coordinate system for transition-state searches along soft modes, since the reduced space already encodes the relevant distortion direction.","Comparing the full forces with the back-transformed reduced forces at each step would give a per-structure diagnostic for whether the chosen parameter subspace is rich enough; large residuals would flag materials needing extra parameters.","The linear-subspace restriction suggests a natural generalization: piecewise-affine or curvilinear parameterizations built from phonon eigenvectors could recover the true minimum while retaining most of the step savings.","If the speedups transfer to denser k-point grids and larger basis sets, symmetry-exact constrained relaxation could make phonon calculations on defective supercells practical at scales where free relaxations currently dominate the cost."],"forward_implications":["Symmetry is preserved exactly during relaxation, eliminating the small symmetry drift that free relaxations introduce and that would otherwise corrupt symmetry-based downstream calculations like finite-difference phonons.","Dynamically stabilized and metastable phases become routinely accessible in high-throughput searches because the optimizer can be confined to the intended polymorph.","Large supercell defect relaxations, such as polarons, can be converged in an order of magnitude fewer steps when the dominant distortion mode is known.","The method requires no change to the underlying optimizer, so it can be dropped into existing electronic-structure codes and relaxation workflows.","The step-count savings grow as the number of free parameters shrinks relative to the full degrees of freedom, making the biggest gains in highly symmetric or tightly constrained systems."],"supporting_citations":[{"why":"Supplies the crystallographic prototype definitions from which the reduced parameter spaces are generated.","marker":"[36]"},{"why":"Extends the prototype library to additional crystal structures, widening the automatic parameter generation.","marker":"[37]"},{"why":"Describes the FHI-aims all-electron code in which the parametric relaxation scheme is implemented.","marker":"[45]"},{"why":"Provides the symmetry analysis used to define and verify crystal prototypes.","marker":"[52]"},{"why":"Source of the initial AFLOW geometries used in the 359-material benchmark.","marker":"[56]"},{"why":"Source of initial Materials Project geometries used in the benchmark.","marker":"[57]"},{"why":"Provides the structural misfit measure used to judge whether constrained and free relaxations reach the same structure.","marker":"[59]"},{"why":"Defines the MgO polaron supercell setup used to demonstrate the order-of-magnitude speedup for local distortions.","marker":"[66]"}],"fun_headline_variants":["Parametric constraints cut DFT relaxation steps by 96%","Symmetry-preserving relaxations speed DFT by 10x","Parametric relaxations: exact symmetry, 10x fewer steps","High-throughput DFT accelerated with parametric symmetry locks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The true relaxed structure must lie within the linear subspace spanned by the chosen parameters; otherwise the constrained relaxation converges to a slightly higher-energy, biased geometry.","fun_headline_variants_meta":{"raw":{"variants":["Parametric constraints cut DFT relaxation steps by 96%","Symmetry-preserving relaxations speed DFT by 10x","Parametric relaxations: exact symmetry, 10x fewer steps","High-throughput DFT accelerated with parametric symmetry locks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000625,"raw_usage":{"total_tokens":2863,"prompt_tokens":884,"completion_tokens":1979,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1912}},"tokens_in":500,"tokens_out":1979,"duration_ms":15713,"temperature":1.0,"reasoning_tokens":1912,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:08:04.594321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a material with a known equilibrium distortion that lies outside the prototype parameter space, such as a strong Jahn-Teller mode in a defective supercell, constrain the relaxation to the prototype's reduced set, and compare the final energy against an unconstrained relaxation; if the constrained minimum sits more than the target convergence tolerance above the free minimum, the subspace was too small and the scheme's savings come at an unacceptable accuracy cost.","supporting_citations":[{"cited_title":"Mehl, David Hicks, Cormac Toher, Ohad Levy, Robert M","cited_arxiv_id":null,"evidence_quote":"Supplies the crystallographic prototype definitions from which the reduced parameter spaces are generated."},{"cited_title":"Mehl, Eric Gossett, Cor- mac Toher, Ohad Levy, Robert M","cited_arxiv_id":null,"evidence_quote":"Extends the prototype library to additional crystal structures, widening the automatic parameter generation."},{"cited_title":"Ab initio molecular simu- lations with numeric atom-centered orbitals","cited_arxiv_id":null,"evidence_quote":"Describes the FHI-aims all-electron code in which the parametric relaxation scheme is implemented."},{"cited_title":"Taylor, Cormac Toher, Michael J","cited_arxiv_id":null,"evidence_quote":"Provides the symmetry analysis used to define and verify crystal prototypes."},{"cited_title":"Taylor, Lance J","cited_arxiv_id":null,"evidence_quote":"Source of the initial AFLOW geometries used in the 359-material benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of initial Materials Project geometries used in the benchmark."},{"cited_title":"Mehl, and Stefano Curtarolo","cited_arxiv_id":null,"evidence_quote":"Provides the structural misfit measure used to judge whether constrained and free relaxations reach the same structure."},{"cited_title":"First-principles su- percell calculations of small polarons with proper ac- count for long-range polarization effects","cited_arxiv_id":null,"evidence_quote":"Defines the MgO polaron supercell setup used to demonstrate the order-of-magnitude speedup for local distortions."}],"review_version":1}