{"id":"128ce275-b4f0-45f6-9600-b1ae7eba978e","arxiv_id":"2502.03727","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A microstructure-aware Bayesian optimization framework with active subspace dimension reduction finds optimal material configurations faster than latent-space-agnostic BO in synthetic and simulated thermoelectric cases.","lead":"This paper combines Bayesian optimization with microstructure descriptors and active subspace projection to design materials by linking processing, structure, and properties. A smart generalist might read it because the method promises faster discovery of materials such as thermoelectrics by using microstructural information rather than ignoring it.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control for extra features: reported convergence gain of latent-space-aware BO could stem from added input variables or ASM, not from microstructure content, so the central claim is underdetermined.","rationale":"The reader's weakest assumption concerns the cost and noise of microstructure characterization. That is a legitimate practical limitation, but the paper explicitly defers it to future work. My concern is more fundamental: the experimental comparison does not isolate the causal role of microstructure information. The latent-space-aware method differs from the baseline not only by having access to microstructure descriptors, but also by having additional input dimensions and by using the active-subspace transformation. Without a control that adds an equal number of non-informative features to the agnostic baseline, the observed gains in convergence cannot be uniquely attributed to the physical content of the microstructural features. This threatens the central claim as stated. The proposed control experiment is inexpensive and would settle the issue. If the control with noise features performs as well as the microstructure-aware version, the paper would need to reframe its contribution; if it does not, the claim is supported. I therefore agree with the CONDITIONAL verdict, but for a different reason than the reader's characterization-cost concern.","tokens_in":840,"tokens_out":1150,"duration_ms":68711,"concrete_test":"Re-run the synthetic benchmark with three additional conditions: (C1) agnostic BO with x plus the same number of random-noise features; (C2) agnostic BO with x plus non-informative deterministic functions of x (e.g., irrelevant polynomial terms); (C3) aware BO with x plus the five latent variables included directly (no active-subspace reduction). Compare average best-found value versus iteration over the same 100 seeds used in the paper. If C1 or C2 matches the latent-space-aware curve, the improvement is not attributable to microstructural content. If C3 matches the active-subspace version, ASM is not the source of the gain.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3 compares a GP on the four design variables x alone (latent space-agnostic BO) with a GP on x plus the active-subspace projection of the five latent variables f1-f5 (latent space-aware BO). The comparison changes two things at once: the set of input features and the use of gradient-based active-subspace preprocessing. No control condition supplies the agnostic BO with an equal number of extra inputs, such as noise variables or non-informative functions of x. Consequently, the convergence advantage in Figs. 4 and 6 could be an artifact of having more input dimensions or of the particular projection, rather than of the physical content of the microstructure descriptors. In the synthetic problem, f1-f5 are deterministic functions of x and y is a simple combination of them; giving these intermediate values to the GP trivially simplifies the regression task. A baseline with arbitrary intermediate variables would reveal whether the benefit is microstructure-specific. In the real case study, the descriptors are outputs of the same simulation that yields kappa, so they carry no independent measurement information; a control using the same number of raw simulation outputs (e.g., leading Fourier modes) could show similar gains, which would undermine the interpretation that microstructure awareness is the cause.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a Bayesian optimization (BO) framework that augments the conventional design-variable inputs to a Gaussian process with a low-dimensional active-subspace projection of microstructural descriptors, treated as latent variables. The method is tested on a synthetic problem with six latent variables and on a phase-field simulation of Mg2SnxSi1-x thermoelectric microstructures, where the effective thermal conductivity is minimized. The authors report that the latent-space-aware BO reaches better objective values after a few iterations than a conventional BO that uses only the design variables, and they use activity scores to identify which microstructural descriptors most influence the objective. The paper concludes that microstructure characterization should be integrated into automated materials discovery platforms.","tokens_in":15587,"tokens_out":12112,"duration_ms":107934,"significance":"The proposed framework is a plausible extension of the authors' earlier work and addresses a real gap: bringing process-structure-property linkages into active learning for materials discovery. The synthetic and phase-field demonstrations are clearly motivated, and the dynamic active-subspace update and activity-score diagnostic are useful ideas. The phase-field simulations with CALPHAD-reinforced elasto-chemical modeling and thermal conductivity computation are physically grounded, and the use of a constant seed number for consistency is good practice. If the convergence gains are robust, the work could inform the design of self-driving laboratories. The paper is also honest about some limitations, including the cost trade-off of microstructure characterization, but the evidence as presented does not yet isolate the cause of the improvement, and an internal inconsistency appears in the results discussion.","major_comments":[{"comment":"The comparison between latent space-agnostic BO and latent space-aware BO changes two factors at once: the GP receives additional input features (the active-subspace projection of f1–f5) and those features are obtained via a gradient-based active subspace method. There is no control in which the agnostic GP is given the same number of extra inputs that are non-informative (e.g., noise variables or arbitrary functions of x), nor a condition in which the latent-aware GP uses the raw latent descriptors without ASM. The convergence advantage in Figs. 4 and 6 could therefore be an artifact of the added input dimensions or of the particular projection, rather than of the microstructural content. Please add control experiments—these are inexpensive in the synthetic setting—to isolate the source of the improvement.","section":"Section 3 (synthetic and phase-field comparisons; Figs. 4 and 6)"},{"comment":"The active subspace method is introduced for a function f(x) and requires gradients ∇x f (Eqs. 4–5). In the application, the active subspace is computed in the latent (microstructure) space, so the required quantity is the gradient of the objective with respect to the descriptors f. The manuscript does not state how these gradients are obtained in the phase-field case, where y=κ is the output of a numerical heat-transfer solve and the descriptors include an Otsu threshold, a radially averaged FFT, and Shannon entropy. Please specify the gradient approximation scheme (e.g., analytic, finite-difference, adjoint, or GP-derivative) and discuss its accuracy and cost.","section":"Section 2.3 and Section 3"},{"comment":"The activity-score discussion for the real case is internally inconsistent. The text states that 'the radially averaged FFT structure function and Shannon entropy are the most influential' and then states 'According to Fig. 6(c), latent variables f2 and f5 have the lowest activity scores.' With the latent variables defined as F = [A_f, c_Mg2Sn, c_Mg2Si, PS_max, S_shannon], f5 is the Shannon entropy, so f5 cannot be both the most and the least influential. Please correct the indices or the summary, and make the figure legend and text consistent.","section":"Section 3, discussion of Fig. 6(c)"},{"comment":"The recommendation that microstructure characterization 'should be integral to automated—and eventually autonomous—platforms' goes beyond the evidence. In both demonstrations the descriptors are exact, noiseless, deterministic outputs of the same simulation that generates the objective, and the manuscript itself acknowledges that measuring structural features may impose additional costs and defers that trade-off to future work. To support the strong practical recommendation, the paper should either include a demonstration with noisy or partial descriptors, or temper the conclusion to the idealized setting with cost-free descriptors.","section":"Section 3 concluding paragraph and Section 4"}],"minor_comments":[{"comment":"The sentence 'we still use Eqs. 20 and 21 to calculate the objective function but the objective function information is extracted from the remaining 5 latent variables' should be reworded; as written it contradicts Eq. (21), which includes f6. The intended meaning is presumably that f6 is hidden from the model and excluded from the ASM, not that y is computed without f6.","section":"Section 3, Eqs. (20)-(21)"},{"comment":"The strain-displacement relation is written as ε_ij = 1/2(∂u_i/∂r_j − ∂u_j/∂r_i), which is the rotation tensor; the plus sign is required for the infinitesimal strain tensor.","section":"Section 2.5, Eq. (14)"},{"comment":"The expression for ηMax_local contains a malformed parenthesis and should be typeset carefully.","section":"Section 3, Eq. (23)"},{"comment":"The keyword line in the abstract contains a stray comma after 'Keywords:'.","section":"Keywords"},{"comment":"The data availability statement is minimal; consider depositing the phase-field and BO data in a public repository or providing a code repository to support reproducibility.","section":"Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a workmanlike extension of the authors' previous study (Molkeri et al., Acta Mater. 2022). The main technical issue is the lack of a control condition in the BO comparison; this is fixable but essential. There is also a clear internal inconsistency in the activity-score paragraph that suggests a final proofread is needed. I would be comfortable with a major revision rather than rejection, as the core idea is sound and the required additions are within scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a sensible incremental extension of the group's 2022 Acta Materialia paper. The new piece is automatic discovery of which microstructure descriptors matter, via a dynamically updated active subspace, instead of assuming that feature is known a priori. That is a real methodological step and the paper mostly delivers a clean demonstration.\n\nWhat works: the synthetic and the phase-field thermoelectric case are both run 100 times, compared against microstructure-agnostic BO and an exhaustive search, and the activity-score analysis is internally consistent. The authors are candid that real microstructure characterization costs are deferred to future work. The framing around process–structure–property is standard but appropriate.\n\nThe soft spots are real but not fatal. The headline comparison changes two things at once: the agnostic GP sees four design variables, the aware GP sees those plus the active-subspace projection. There is no control condition that gives the agnostic GP the same number of extra inputs (noise variables, say). In the synthetic problem, f1–f5 are deterministic functions of x and y is a simple combination of them; handing those intermediate values to the GP trivially simplifies regression. So the reported convergence advantage could stem from having more input dimensions or from the projection, rather than from the physical content of the microstructure descriptors. The interpretation \"microstructure awareness drives the gain\" is underdetermined by the experiments. That said, the paper is explicitly a proof-of-concept for a pipeline, and the missing controls are easy to add.\n\nTwo smaller issues: the gradient computation for the active subspace in the real case is opaque—no detail on how ∇_f κ is obtained from the phase-field/heat-transfer simulation. And the discussion of f6 in the synthetic setup (Eqs. 20–21) is confusing: y includes f6, yet the text says objective information is extracted from the remaining five latent variables. I think they mean f6 is simply unobserved, but the wording needs to be fixed. Also, no code is released and data are \"available upon request,\" which limits reproducibility.\n\nVerdict: worth a serious referee. The method is plausible, the paper is honest about its scope, and the confounds are fixable. A reviewer should push for the extra-feature control and gradient details; if those pass, this becomes a solid reference for microstructure-aware BO.","headline":"Sensible incremental extension of the group's earlier microstructure-aware BO; the convergence gain is real but not yet causally attributed to microstructure content.","tokens_in":16034,"tokens_out":3212,"would_cite":true,"duration_ms":32637,"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":"The paper claims that making Bayesian optimization microstructure-aware—feeding microstructural descriptors into the Gaussian process as latent variables compressed by the active subspace method—finds optimal material designs in fewer…","keywords":["microstructure-aware design","Bayesian optimization","Gaussian process","active subspace method","process-structure-property-performance","thermoelectric materials","thermal conductivity","materials discovery"],"falsifier":"A head-to-head campaign on a real material where each iteration includes an actual microstructure measurement would settle the claim: if the microstructure-aware optimizer's total cost including characterization time and noise-induced repeat measurements is not lower than a chemistry-only optimizer on the same budget, or if its convergence advantage vanishes when descriptor noise is introduced, the central claim fails.","tokens_in":15145,"feed_emoji":"🔬","tokens_out":9758,"duration_ms":90973,"temperature":0.7,"pith_summary":"Bayesian optimization is a standard way to search for material compositions and processing conditions with as few experiments as possible, but it usually ignores the microstructure that actually mediates between processing and properties. This paper tries to establish that explicitly feeding microstructural descriptors—such as phase fraction, scattering intensity, and entropy—into the optimizer as latent variables makes the search more efficient and accurate. It does so by compressing the descriptors with the active subspace method and conditioning a Gaussian process on both the design variables and the reduced microstructure representation. In a synthetic problem and in a simulated Mg$_2$Sn$_x$Si$_{1-x}$ thermoelectric thermal-conductivity design task, the microstructure-aware optimizer reaches the same optimum as the standard one after fewer iterations. If true, automated materials discovery platforms should treat microstructure characterization as a routine part of the optimization loop rather than as a post-mortem analysis.","feed_headline":"Microstructure data cuts experiments needed to find optimal materials","feed_subtitle":"Injecting structural descriptors as latent variables lets Bayesian design beat chemistry-only search, simulations show.","key_machinery":"The active subspace method (ASM) carries the argument. At observed points, the framework estimates the covariance matrix of the objective's gradient with respect to the microstructure descriptors and keeps the eigenvectors with the largest eigenvalues; projections onto those eigenvectors define a low-dimensional subspace that captures most of the objective's variability. The Gaussian process is conditioned on the union of the original design variables and this reduced latent representation, and the subspace is dynamically updated as new observations arrive. This turns a high-dimensional, partly unknown microstructure descriptor set into a small set of informative coordinates, which is what lets the optimizer exploit microstructure information without paying the full cost of the extra dimensions.","core_discovery":"The central claim is that latent-space-aware Bayesian optimization—conditioning the Gaussian process on an active-subspace projection of microstructure descriptors in addition to the controllable chemistry and processing variables—improves both the predictive model and the convergence of the design loop. In both demonstrations, the latent-aware optimizer matches or outperforms the latent-agnostic optimizer after the first few iterations, and all configurations converge to the same optimal region, so the gain is efficiency in reaching a known optimum. The active subspace is recomputed as observations accumulate, and its eigenvalues give activity scores that identify which microstructure features most control the objective; in the thermoelectric case, the radially averaged FFT structure function and Shannon entropy together explain nearly 90% of the total activity score. The paper presents this as a step toward treating microstructure as an explicit, addressable design variable instead of an emergent by-product of processing.","pith_inferences":["Extension: the paper does not price the cost of measuring microstructure, so a natural next step is to compare total campaign cost and let the optimizer decide adaptively whether the expected improvement justifies paying for a characterization step, using activity scores to choose which descriptors to measure.","Because the active subspace is estimated from local gradients of the objective, applying the method to real experimental data with noisy or sparse measurements may require gradient surrogates; it is an open quantitative question how much noise the convergence advantage survives.","The same latent-space-aware structure could apply to other intermediate variables beyond microstructure, such as stress state or defect density, whenever a measurable intermediate quantity links inputs to the objective.","A prospective test would be to run two parallel campaigns on a real alloy or thermoelectric system with equal budgets, one with and one without microstructure feedback, and compare how fast each reaches a target property."],"forward_implications":["Design campaigns that measure microstructure descriptors during optimization should reach target compositions and processing conditions in fewer iterations than campaigns that record only chemistry and property.","The activity scores give designers a data-driven ranking of which microstructural features to control, so characterization effort can be focused on the features that actually drive the property.","Because the active subspace is updated online, the method can start with no prior knowledge of which microstructure feature matters and discover it during the campaign, unlike the earlier approach that required that knowledge in advance.","The efficiency gain in the examples is convergence speed to the same optimum, not a better optimum, so the practical payoff is reduced experimental cost per campaign.","For self-driving laboratories, the framework implies that real-time microstructure characterization should be built into the optimization loop because the information it provides accelerates subsequent decisions."],"supporting_citations":[{"why":"Supplies the prior demonstration that microstructure features accelerate Bayesian materials design, which the present framework extends by discovering rather than prescribing the critical features.","marker":"[29]"},{"why":"Provides the Gaussian-process regression machinery that the framework conditions on.","marker":"[36]"},{"why":"Establishes the active subspace method used for dimensionality reduction.","marker":"[40]"},{"why":"Bases the adaptive active-subspace optimization loop that the framework builds on.","marker":"[45]"},{"why":"Supplies the sequential-optimization variant of active-subspace reduction with Gaussian processes.","marker":"[47]"},{"why":"Provides the phase-field modeling framework that generates the simulated microstructures in the case study.","marker":"[49]"},{"why":"Provides the semi-implicit Fourier-spectral numerical solver used for microstructure evolution.","marker":"[50]"},{"why":"Gives the micro-elasticity equations used to model stress-strain interactions in the simulated microstructures.","marker":"[51]"},{"why":"Supplies the Otsu thresholding method used to compute the phase-fraction microstructure descriptor.","marker":"[52]"},{"why":"Frames thermal conductivity as the optimization objective and Fourier's law as the heat-transfer model.","marker":"[54]"}],"fun_headline_variants":["Microstructure-aware Bayesian design finds materials faster","Bayesian materials design gets a microstructure boost","Microstructure data speeds up Bayesian materials search","Active subspace microstructures cut experiments in design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that microstructural descriptors can be obtained exactly and without meaningful cost at every optimization step, because in the demonstrations the descriptors come from the same simulation that produces the property; the paper itself notes that real characterization could add cost, and if measurements are slow, noisy, or expensive, the convergence advantage could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Microstructure-aware Bayesian design finds materials faster","Bayesian materials design gets a microstructure boost","Microstructure data speeds up Bayesian materials search","Active subspace microstructures cut experiments in design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000121,"raw_usage":{"total_tokens":1085,"prompt_tokens":928,"completion_tokens":157,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":102}},"tokens_in":544,"tokens_out":157,"duration_ms":2498,"temperature":1.0,"reasoning_tokens":102,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:57:35.136695+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A head-to-head campaign on a real material where each iteration includes an actual microstructure measurement would settle the claim: if the microstructure-aware optimizer's total cost including characterization time and noise-induced repeat measurements is not lower than a chemistry-only optimizer on the same budget, or if its convergence advantage vanishes when descriptor noise is introduced, the central claim fails.","supporting_citations":[{"cited_title":"Molkeri, D","cited_arxiv_id":null,"evidence_quote":"Supplies the prior demonstration that microstructure features accelerate Bayesian materials design, which the present framework extends by discovering rather than prescribing the critical features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian-process regression machinery that the framework conditions on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the active subspace method used for dimensionality reduction."},{"cited_title":"Khatamsaz, et al., Adaptive active subspace-based e fficient multifidelity materials design, Mater","cited_arxiv_id":null,"evidence_quote":"Bases the adaptive active-subspace optimization loop that the framework builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sequential-optimization variant of active-subspace reduction with Gaussian processes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the phase-field modeling framework that generates the simulated microstructures in the case study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the semi-implicit Fourier-spectral numerical solver used for microstructure evolution."},{"cited_title":"Attari, A","cited_arxiv_id":null,"evidence_quote":"Gives the micro-elasticity equations used to model stress-strain interactions in the simulated microstructures."},{"cited_title":"Otsu, A threshold selection method from gray-level his- tograms, IEEE transactions on systems, man, and cybernetics 14 9 (1) (1979) 62–66","cited_arxiv_id":null,"evidence_quote":"Supplies the Otsu thresholding method used to compute the phase-fraction microstructure descriptor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Frames thermal conductivity as the optimization objective and Fourier's law as the heat-transfer model."}],"review_version":1}