{"id":"af16623c-7f26-403f-a3e6-8b984fcd1c1c","arxiv_id":"2501.00012","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A direct-sampling path planner with a 'bad phase purge' transformation finds CALPHAD-valid composition gradients in 5 to 9 element systems using orders of magnitude fewer thermodynamic calculations than surrogate modeling.","lead":"Materials scientists often design composition gradients between alloys by building a computer model of the whole alloy space, which becomes slow when many elements are involved. This paper tests a faster approach that queries the thermodynamic model only along candidate paths, then uses a phase-rule transformation to remove unwanted phases, reporting up to a million-fold reduction in thermodynamic calculations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bad Phase Purge continuity is unproven and the method admits possible bad-phase incursions between discrete points; feasibility of generated gradients therefore rests on an untested conjecture. A dense revalidation of transformed paths would settle it.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the BPP conjecture guarantees no bad phase at each transformed point but does not guarantee continuity, and the method's own footnote admits possible small incursions between discrete points. I agree with that assessment. The paper is otherwise careful: it labels the continuity statement as unproven, provides a partial proof, and gives empirical evidence across six alloy pairs. Those are real supporting data, but they do not close the gap. The central efficiency claim is only meaningful if the final path actually avoids bad phases everywhere along the continuous gradient, not merely at sampled vertices. Because the paper openly leaves this unresolved, the appropriate verdict remains CONDITIONAL: accept only after the dense revalidation described above, or after a rigorous proof of continuity. I do not recommend changing the reader's verdict, since the reader already arrived at CONDITIONAL for exactly this reason.","tokens_in":14946,"tokens_out":4647,"duration_ms":46294,"concrete_test":"Perform a dense re-check on at least one relaxed+BPP path from each alloy pair, especially the 7- and 9-element cases (316L–AlSiMg, IN718–AlSiMg). After applying Eq. (1) at the original 0.01 discretization, subdivide every straight edge of the transformed path at 0.001 (and at 0.0001 wherever the purged point moves by more than 0.05) and run the same single-equilibrium CALPHAD check with the same good/limited/bad classification. Record the maximum jump ||sigma*(alpha_{i+1}) - sigma*(alpha_i)|| between adjacent purged points. If any interior point has total bad fraction above the paper's own threshold (<10^-2), or if any adjacent jump exceeds the collision-check resolution, the generated gradient is not feasible as claimed; if all interior points are bad-free and jumps are below resolution, the practical conjecture holds for these cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that relaxed-RRT plus BPP produces feasible (zero bad phase) paths with orders-of-magnitude fewer CALPHAD calls—depends on the BPP transformation preserving path continuity and on interpolated segments staying bad-phase-free. The paper proves only pointwise absence of bad phases (Supplementary Notes, item 1) and explicitly states: 'We have been unable to prove (2). Thus, it remains to be rigorously shown that the transformed path is continuous (does not contain discrete jumps in composition).' The Section 3.3 footnote concedes 'small incursions into bad or limited may occur between the discrete points' because BPP is applied at resolution 0.01 and edges are linearly interpolated. Since H1 and H2 are discontinuous across phase-boundary surfaces, the pointwise map sigma*(alpha) in Eq. (1) can jump when adjacent original points lie in different multiphase regions; even if each purged vertex is feasible, a straight segment between two feasible vertices can re-enter bad-phase territory. The reported speed-ups (8x10^2 to 10^6) are computed for paths whose feasibility is exactly what is in question. This is an internal admitted gap, not an external-consensus dispute.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an on-the-fly path planning method for composition gradients in high-dimensional alloy systems, replacing upfront surrogate modeling of CALPHAD with direct equilibrium calculations during RRT sampling. To overcome narrow feasible corridors, the authors relax the bad-phase constraint and then apply a pointwise 'Bad Phase Purge' (BPP) transformation, based on the lever rule, to move each point along its equilibrium tie-simplex until no bad phases remain. The method is benchmarked on six pairs of structural alloys at 400 °C and on optimal planning with RRT*. The authors prove pointwise elimination of bad phases, but explicitly state they cannot prove continuity of the transformed path. Reported speed-ups over surrogate modeling range from 8x10^2 to 10^6.","tokens_in":15226,"tokens_out":6672,"duration_ms":56035,"significance":"If the continuity of the BPP-transformed path can be established or validated, the paper offers a practically valuable alternative to surrogate-based planning in 5-9 element systems: the transformation contains no fitted parameters, is derived from equilibrium thermodynamics, and the pointwise algebra in the Supplementary Notes is correct and machine-checkable. The paper is also honest in flagging the unproven part of the conjecture. However, the headline efficiency numbers are relative to an estimated, not measured, surrogate training cost, and the feasibility of the returned paths rests on the unproven continuity property, so the significance is conditional at this stage.","major_comments":[{"comment":"The central feasibility claim depends on continuity of the transformed path, which the authors explicitly state they have not proven: 'We have been unable to prove (2).' Consequently, a straight-line segment between two adjacent purged points at the 0.01 discretization can re-enter bad-phase regions, and the footnote in §3.3 concedes that 'small incursions into bad or limited may occur between the discrete points.' Because the speed-ups in Figure 5 are reported for paths whose feasibility rests on this property, please either prove continuity (or a bound on incursion size) or perform a dense revalidation for all six paths and random seeds, reporting the maximum bad-phase fraction on interpolated sub-intervals between every pair of adjacent purged points.","section":"Appendix A / Supplementary Notes"},{"comment":"The efficiency benchmark compares RRT+purge query counts against a surrogate-model training cost estimated as 5×10^d, cited to Ref. [17] as 'one example using at least 5×10^d samples.' This is an external estimate rather than a measured surrogate cost for the six alloy systems considered here; for d=9 it implies 5×10^9 CALPHAD calls, which this paper cannot verify. The 8×10^2 to 10^6 speed-up factors are therefore ratios to an unmeasured quantity. Please state explicitly that this is a literature-derived estimate, cite the exact location in Ref. [17], and, if possible, directly measure the surrogate construction cost for at least the 5-element Ti64-AlSi10Mg system.","section":"§3.3, Figure 5"},{"comment":"Equation (1) is undefined whenever H1(sigma(alpha))·V_good = 0, i.e., when the equilibrium at a point contains no good phase. The conjecture assumes only f_undesirable < 1, which does not exclude such points because 'limited' phases count neither as good nor bad. In addition, the encoding of limited phases in V_good is not defined: Eq. (1) says 'good (1) or bad (0)', yet §3.3 and Figure 7 indicate that BPP removes limited phases as well. Please specify the encoding of limited phases, add the condition H1·V_good > 0 to the conjecture (or prove it for the generated paths), and report the minimum good-phase fraction along each of the six paths.","section":"Eq. (1), Appendix A"}],"minor_comments":[{"comment":"The steering distance is given as '√2/10, representing 10% of the maximal distance within any composition space'; please clarify the units and the definition of maximal distance, since the raw value is ambiguous in a simplex with normalized composition coordinates.","section":"§2"},{"comment":"The sentence 'the speed-up provided by the new approach ranges from 8x10^2 (Ti64 to AlSiMg) to 10^6 (Ti64 to AlSiMg)' lists the same alloy pair twice; one of these should likely be a different pair, such as IN718-AlSiMg.","section":"§3.3"},{"comment":"The line 'Similarly, we can show H1(sigma*)·Vgood =' is incomplete and lacks the right-hand side of the identity; please complete this displayed equation.","section":"Supplementary Notes"},{"comment":"The caption refers to 'ten unique pairs of five alloys (the four alloys described previously along with NiTi)', but NiTi is not defined in Table 1 with a composition or limited-phase list; please add this information or remove NiTi from the figure.","section":"Figure 8"},{"comment":"The manuscript does not include a data or code availability statement; providing the Thermo-Calc scripts, the relaxation threshold used for each run, and the random seeds would greatly improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and is candid about the unproven conjecture. The main risk is that the headline speed-up is computed for paths whose feasibility depends on that conjecture. I would be willing to reconsider after the authors supply either a proof of continuity or a dense revalidation, and after the surrogate cost estimate is clearly separated from measured numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick take: this is a real advance for practical composition-gradient design in 5–9 element systems, and the authors are refreshingly upfront about the one mathematical gap. The OTF approach—sampling CALPHAD directly during RRT instead of building a surrogate—is a natural but genuinely useful departure, and the relaxed-constraint + Bad Phase Purge trick is clever. Pointwise, the purge provably removes bad phases using lever-rule tie-simplex geometry, and the paper shows convincingly that it cuts CALPHAD queries by orders of magnitude on real alloy pairs.\n\nWhere it gets soft: the transformation is only conjectured to preserve path continuity, and the authors say so plainly in the Supplementary Notes. If adjacent purged points jump, the interpolated segment can re-enter bad-phase regions; Section 3.3 even footnotes small incursions. That is a genuine gap between “each sampled point is clean” and “the physical gradient is feasible.” For a manufacturing design tool, that matters. The speed-up numbers are also computed against an estimated surrogate-model cost (5×10^d) rather than a measured baseline for the high-dimensional cases, so the 10^6 figure is an upper-bound-ish estimate, not a direct measurement.\n\nThat said, the empirical evidence is decent: six alloy pairs, three seeds, paths verified to have <10^-2 bad phase at discrete points. The authors don't overclaim; they call the BPP a conjecture and invite counterexamples. The pointwise algebra is correct, there are no fitted parameters, and the citation pattern to the surrogate-model literature [10,11,17] is appropriate.\n\nWho is this for? Someone building computational design tools for additive manufacturing or doing CALPHAD-based path planning. They'd get a new method worth trying. A serious referee should engage, mainly to push for either a proof of continuity under reasonable assumptions (e.g., within a single tie-simplex region or across boundaries with non-degenerate good-phase fractions) or a dense revalidation protocol that checks interpolated segments and quantifies incursions. I'd take it to peer review; it deserves referee time.","headline":"A genuinely useful engineering method, honestly presented, but the bad phase purge rests on an unproven continuity conjecture and the headline speedups lean on an estimated surrogate baseline.","tokens_in":15706,"tokens_out":1826,"would_cite":true,"duration_ms":17777,"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":"This paper claims that replacing machine-learned surrogates with direct on-the-fly thermodynamic sampling—plus a relaxed bad-phase constraint and the Bad Phase Purge transformation—makes composition-gradient path planning tractable in…","keywords":["composition gradients","functionally graded materials","CALPHAD","rapidly-exploring random trees","Bad Phase Purge","lever rule","high-dimensional path planning","on-the-fly sampling"],"falsifier":"Take one of the relaxed paths from the paper, apply Bad Phase Purge to adjacent discretized points, and evaluate the equilibrium phase fractions along the straight interpolation between each purged pair: if any interpolated segment shows a nonzero bad-phase fraction, or if adjacent purged points jump across a bad-phase region, then the purge does not guarantee a feasible gradient.","tokens_in":14767,"feed_emoji":"🧭","tokens_out":14425,"duration_ms":121692,"temperature":0.7,"pith_summary":"The paper argues that planning composition gradients in multi-element alloys does not require building a machine-learned surrogate of thermodynamic phase predictions. Instead, querying the thermodynamic calculation directly during a rapidly-exploring random tree search, while temporarily allowing bad phases, finds feasible paths with orders of magnitude fewer calculations. The relaxed path is then pushed out of bad-phase regions by the Bad Phase Purge formula, a lever-rule reweighting of equilibrium phases. In benchmarks on four structural alloys, the approach was up to about a million times faster than surrogate modeling, and it succeeded in cases where strict planning stalled.","feed_headline":"Finding alloy gradients is a million times faster","feed_subtitle":"Direct thermodynamic queries replace months of surrogate-model training for 5-9 element alloys.","key_machinery":"The load-bearing object is the Bad Phase Purge (BPP) transformation, defined as $$\\$\\sigma$^*(\\$\\alpha$) = H_2(\\$\\sigma$(\\$\\alpha$)) \\cdot \\frac{1}{H_1(\\$\\sigma$(\\$\\alpha$)) \\cdot \\vec{V}_{\\mathrm{good}}} \\left( H_1(\\$\\sigma$(\\$\\alpha$)) \\odot \\vec{V}_{\\mathrm{good}} \\right),$$ where $H_1$ maps a composition to its equilibrium phase fractions, $H_2$ maps a composition to a matrix of the equilibrium phase compositions, and $\\vec{V}_{\\mathrm{good}}$ is a 0/1 vector marking acceptable phases. The formula uses the lever-rule fact that, within a multiphase region, all compositions on a tie-simplex share the same phase compositions, so renormalizing the good-phase fractions moves the point out of bad-phase territory without altering the good phases' compositions. Applying this pointwise to a discretized path needs no additional thermodynamic queries. The paper proves each transformed point has zero bad phase by a short algebraic identity; it states but does not prove that the transformed path remains continuous.","core_discovery":"The central discovery is that the bottleneck in computational gradient design is the upfront surrogate, not the planner: in systems with more than four elements, the number of ground-truth thermodynamic (CALPHAD) calculations needed to build a surrogate grows exponentially with element count, whereas the number of direct queries an RRT needs to find a first feasible path depends on the visibility of the free space, not the dimension. The paper's method therefore samples the equilibrium calculation on the fly during RRT growth, relaxes the bad-phase constraint to widen narrow corridors, and then applies the Bad Phase Purge transformation pointwise to eliminate bad phases from the resulting path. For the six alloy pairs tested, this relaxed-plus-purge pipeline found paths with effectively zero bad phase everywhere, including two Al-rich pairs where the strict constraint found no path within two weeks. The reported speedup over surrogate modeling ranges from about 8×$10^{2}$ to $10^{6}$ across the test systems.","pith_inferences":["The paper leaves implicit that the same relax-and-purge strategy could apply to other equilibrium constraints: any quantity defined through phase fractions and phase compositions can be assigned a good/bad mask, so the recipe is not tied specifically to 'bad phases'.","A natural extension is to cache the $H_1$ and $H_2$ outputs from earlier collision checks and reuse them for nearby queries; because Bad Phase Purge is pointwise and costs no extra thermodynamic evaluations, this hybrid could preserve the speedup while cutting duplicate calculations.","The reported speedup depends on the surrogate training-cost estimate used for comparison; if future active-learning surrogates need far fewer than $5\\times10^d$ samples, the dimension crossover at which direct sampling wins will shift.","If the unproved continuity of the purged path fails in some phase diagram, an adaptive-resolution repair is plausible: refine only segments whose purged endpoints lie in different phase regions, testing the interpolation at higher resolution before accepting the path."],"forward_implications":["For the four structural alloys tested, every relaxed path after Bad Phase Purge had effectively zero bad phase at all sampled points, including alloy pairs where a strict RRT could not find a path within two weeks.","The number of thermodynamic calculations for the first feasible path scales with the geometry of the feasible region rather than with the number of elements, so the advantage over surrogate modeling grows as systems become more element-rich.","With RRT*, Bad Phase Purge can be applied on-the-fly, so optimized paths for path length or integrated chemical-potential gradient are evaluated on the purged, strict-constraint path rather than on the relaxed path.","Because Bad Phase Purge works pointwise and uses only data already returned by the equilibrium calculation, the total thermodynamic cost of the relaxed-plus-purge pipeline is essentially the number of RRT collision checks."],"supporting_citations":[{"why":"Establishes the CALPHAD-coupled path planning approach and the surrogate-model baseline that the on-the-fly method is compared against.","marker":"[10]"},{"why":"Supplies the rapidly-exploring random tree (RRT) algorithm used for stochastic sampling and connection of feasible points.","marker":"[14]"},{"why":"Supplies RRT*, the asymptotically optimal variant used for the path-optimization extension with on-the-fly Bad Phase Purge.","marker":"[13]"},{"why":"Provides the subspace-inclusive sampling method and the surrogate scaling estimate of $5\\times10^d$ ground-truth samples used to benchmark efficiency.","marker":"[17]"},{"why":"Gives the probabilistic-completeness and free-space-visibility rationale for why RRT sample counts need not scale with dimension.","marker":"[18]"},{"why":"Supplies the CALPHAD software and thermodynamic database used to evaluate equilibrium phases for each collision check.","marker":"[19]"},{"why":"Defines the tie-simplex formalism of multiphase equilibrium that underpins the lever-rule argument behind Bad Phase Purge.","marker":"[21]"}],"fun_headline_variants":["Direct thermodynamic queries speed alloy gradient design 1e6x","Skip the surrogate: on-the-fly CALPHAD sampling is 1e6x faster","Lever rule corollary cuts gradient design from months to minutes","No surrogate needed: on-the-fly path planning for alloy gradients"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the premise that applying the Bad Phase Purge separately to each sampled point of a relaxed path leaves a connected, bad-phase-free path, but the paper proves only that each transformed point is free of bad phases, not that the transformed path is continuous.","fun_headline_variants_meta":{"raw":{"variants":["Direct thermodynamic queries speed alloy gradient design 1e6x","Skip the surrogate: on-the-fly CALPHAD sampling is 1e6x faster","Lever rule corollary cuts gradient design from months to minutes","No surrogate needed: on-the-fly path planning for alloy gradients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1348,"prompt_tokens":936,"completion_tokens":412,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":333}},"tokens_in":552,"tokens_out":412,"duration_ms":4023,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:55:07.722394+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one of the relaxed paths from the paper, apply Bad Phase Purge to adjacent discretized points, and evaluate the equilibrium phase fractions along the straight interpolation between each purged pair: if any interpolated segment shows a nonzero bad-phase fraction, or if adjacent purged points jump across a bad-phase region, then the purge does not guarantee a feasible gradient.","supporting_citations":[{"cited_title":"21 5 Acknowledgements This work is supported by NASA grant number ECF 80NSSC21K1810, and the Department of Defense through the NDSEG fellowship","cited_arxiv_id":null,"evidence_quote":"Establishes the CALPHAD-coupled path planning approach and the surrogate-model baseline that the on-the-fly method is compared against."},{"cited_title":"Reichardt, A.A","cited_arxiv_id":null,"evidence_quote":"Supplies RRT*, the asymptotically optimal variant used for the path-optimization extension with on-the-fly Bad Phase Purge."},{"cited_title":"Hofmann, S","cited_arxiv_id":null,"evidence_quote":"Supplies the CALPHAD software and thermodynamic database used to evaluate equilibrium phases for each collision check."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the tie-simplex formalism of multiphase equilibrium that underpins the lever-rule argument behind Bad Phase Purge."}],"review_version":1}