{"id":"becbd389-c70f-47df-97b8-9208e9167461","arxiv_id":"1908.00638","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors propagate MCMC-quantified CALPHAD parameter uncertainty through an elasto-chemical phase-field model of Mg2(SixSn1-x) and use machine learning to map the resulting microstructure distributions.","lead":"A materials science team built and demonstrated a pipeline that carries uncertainty from thermodynamic database parameters through microstructure evolution simulations in a thermoelectric alloy. The framework is meant to give engineers confidence bounds on predicted microstructures instead of single deterministic answers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MCMC calibration against a single computed CALPHAD phase diagram (Kozlov et al. [62]) makes the propagated credible intervals conditional on that diagram's bias; because experimental boundaries differ by tens of at%, the central uncertainty-propagation claim lacks external anchoring.","rationale":"The reader's weakest assumption is the same as the one I consider most load-bearing: the upstream CALPHAD posterior is anchored to a single computed phase diagram despite acknowledged experimental disagreement. I do not see an internal inconsistency in the propagation chain itself; the Gaussian-copula sampling and the 10,000-run phase-field campaign are a plausible demonstration of forward uncertainty propagation. The OPMD data link is a positive reproducibility gesture, though no dataset identifier or code is provided, and the paper does not report classifier accuracy or reconcile the 200,000 time series with 10,000 parameter combinations. These issues are secondary. Because the central claim is framed as propagating statistically quantified uncertainties, and the input calibration rests on one computational assessment rather than the experimental scatter the paper itself emphasizes, the conditional verdict remains appropriate without requiring rejection.","tokens_in":22828,"tokens_out":4655,"duration_ms":49075,"concrete_test":"Re-run the Section 4.1 MCMC with the likelihood based on the experimental phase-boundary data cited in refs. [38,39,62,63,70] (with the reported inter-experiment scatter of tens of at% as the likelihood variance) instead of the Kozlov et al. calculated diagram; then propagate both posteriors through the same phase-field sampling. If the posterior predictive 95% BCI of the phase diagram or the Table 2 QoI distributions (e.g., Cα, Cβ, area fraction) shift by more than the width of the original Kozlov-based band, the calibration reference is load-bearing and the paper's confidence bounds are conditional on that reference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is in Section 4.1: the MCMC posterior for the six CALPHAD parameters is fitted to 'calculated composition-temperature data sampled from the phase diagram proposed by Kozlov et al. [62]'—a computation, not a measurement. The introduction itself states that experimental phase boundary estimates for this system disagree by several tens of atomic percent ([39]). Since every downstream step (the Gibbs energy BCI bands in Figure 5, the Gaussian-copula samples in Section 3.3.3, and the microstructure QoI distributions in Table 2) is driven by this posterior, any systematic bias in the Kozlov diagram shifts the entire uncertainty band. The unknown likelihood variance treated as a hyperparameter can widen the band, but it cannot correct a biased mean. The paper therefore demonstrates propagation of uncertainty conditional on one computational reference, not the true experimental uncertainty that its motivating discussion invokes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a framework for uncertainty quantification and propagation through a multiscale model chain consisting of a CALPHAD thermodynamic description, a microelasticity model, and a Cahn-Hilliard phase-field model, applied to the Mg2(SixSn1-x) thermoelectric system. The CALPHAD interaction parameters are calibrated with an adaptive Metropolis-Hastings MCMC against the computed phase diagram of Kozlov et al., yielding posterior distributions that are propagated to Gibbs free energy curves and phase diagram credible intervals. These posteriors are then combined, via a Gaussian copula, with priors on microelastic and kinetic parameters to sample a high-dimensional input space for 10,000 phase-field simulations. The resulting microstructure ensemble is characterized through several quantities of interest, and machine-learning classifiers and clustering are used to identify decomposed versus not-decomposed microstructures. The authors also curate the generated dataset in the Open Phase-field Microstructure Database.","tokens_in":23093,"tokens_out":5106,"duration_ms":52642,"significance":"If the approach is sound, this is one of the few complete demonstrations of end-to-end uncertainty propagation through a CALPHAD-to-phase-field chain, and it adds a publicly available large microstructure dataset. The combination of Bayesian calibration, copula-based sampling, high-throughput phase-field simulation, and ML-based analysis is methodologically ambitious and could serve as a template for ICME uncertainty workflows. The main strength is the explicit treatment of correlated thermodynamic parameters and the transparent release of data. However, the external anchoring of the uncertainty is a central concern, because the MCMC posterior is conditioned on a single computational phase diagram rather than on experimental data, and the paper does not fully acknowledge the conditionality of its uncertainty bands.","major_comments":[{"comment":"The MCMC calibration in Section 4.1 fits the six CALPHAD parameters to 'calculated composition-temperature data sampled from the phase diagram proposed by Kozlov et al. [62]', which is a computational reference, not experimental data. The introduction (Section 2) itself notes that experimental phase boundary estimates for Mg2(SixSn1-x) disagree by several tens of atomic percent [39]. Since the posterior distributions drive the Gibbs energy BCIs in Figure 5, the Gaussian copula samples in Section 3.3.3, and the microstructure QoI distributions in Table 2, the reported credible intervals are conditional on the Kozlov diagram's bias. An unknown likelihood variance hyper-parameter can widen the intervals but cannot recenter a biased mean. Please either re-frame the central claim as propagation of uncertainty conditional on a chosen reference diagram, or anchor the posterior with experimental phase boundary data or multiple independent CALPHAD assessments.","section":"Section 4.1, Figure 5, Table 2"},{"comment":"The Gaussian copula is used to sample the 18-dimensional input space for phase-field simulations, but the paper does not specify how the correlation matrix R is constructed for the full input set, nor does it validate that the copula reproduces the actual joint posterior of the CALPHAD parameters. The joint frequency plot in Figure 4 suggests non-elliptical dependence for at least one parameter pair, and a Gaussian copula with pairwise correlations may not preserve higher-order or tail dependencies. Because the microstructure QoI distributions depend on the joint sampling, please add a quantitative validation step (for example, energy distance or scatter/QQ comparisons) between copula-generated CALPHAD parameter vectors and the MCMC posterior, and state explicitly how R is assembled from the MCMC covariance and the assumed independent priors.","section":"Section 3.3.3"},{"comment":"The 'Dispersion (sigma^2 mu)' column in Table 2 contains negative entries (e.g., -1012.1 for mu_chem = -156.06 and -6.01e-6 for mu_int = 1.59e-9), which are impossible if this column reports a variance or variance-to-mean ratio with positive variance. The header is also not defined consistently with Table 1, where the analogous column appears to report variances. These values undermine the quantitative QoI statistics that are central to the propagation claim. Please correct the definition and the values, or clearly relabel the column as the index of dispersion if that is the intent.","section":"Table 2"},{"comment":"Convergence of the MCMC sampler is supported only by a qualitative joint-frequency plot and a statement that 100,000 samples were generated. Since the posterior is the foundation for all downstream uncertainty propagation, quantitative convergence diagnostics (e.g., effective sample size, trace plots, or multiple-chain Gelman-Rubin statistics) should be reported to justify the use of the last 5,000 samples as representative of the stationary distribution.","section":"Section 3.3.1 and Section 4.1"}],"minor_comments":[{"comment":"The abstract mentions '200,000 time series of synthetic microstructures', while Section 4.2 reports 10,000 parameter combinations run through the phase-field solver; please clarify whether the 200,000 figure corresponds to 10,000 simulations at multiple time frames and state this explicitly.","section":"Abstract and Section 4.2"},{"comment":"The SFTS parameter row in Table 1 lists a negative dispersion value (-4.28), which is inconsistent with the variance interpretation used for other rows; this appears to be a sign or formatting error and should be corrected.","section":"Table 1, SFTS row"},{"comment":"References [61] and [70] are the same paper by Vives et al.; please consolidate the duplicate citation.","section":"References [61] and [70]"},{"comment":"The text states 'VMR = 0 is associated to a random data-set', which is incorrect: a variance-to-mean ratio of zero implies zero variance, not randomness; the intended statement is likely that VMR near 1 corresponds to a Poisson-like random distribution.","section":"Section 4.3, paragraph on VMR"},{"comment":"The sampling methodology is explained clearly with the two-dimensional example, but the extension to the actual 18-dimensional case should state how the marginal distributions for microelastic and kinetic parameters are specified (e.g., truncated normal fits, uniform bounds) and how the zero-correlation assumption for those parameters is reflected in R.","section":"Section 3.3.3 and Figure 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real attempt to push uncertainty through an entire CALPHAD → microelasticity → phase-field chain, and as far as I know the integration is genuinely new. The pipeline works, and the authors are honest about the mechanics. But the uncertainty they propagate is conditioned on a computed phase diagram, not experimental data, so the reported credible intervals should not be mistaken for the true uncertainty in the Mg2(Si,Sn) system.\n\nWhat is actually new: linking MCMC-based CALPHAD parameter inference with Gaussian copula sampling (which preserves pairwise correlations) and high-throughput phase-field simulation, then using ML to classify the resulting microstructure space. That combination is absent from the cited UQ-in-CALPHAD and UQ-in-phase-field literature. The paper also ships a large synthetic microstructure dataset via OPMD, which is a real resource. The phase-field physics is standard – Cahn-Hilliard plus microelasticity – but the contribution is the chain, not the individual models. The demonstration that QoIs are multimodal, with a large dispersion, is a useful warning against deterministic point sampling.\n\nThe soft spot is load-bearing: Section 4.1 fits the six CALPHAD parameters to 'calculated composition-temperature data sampled from the phase diagram proposed by Kozlov et al.' That is a computation, and the introduction tells you experimental boundaries differ by tens of at%. The unknown likelihood variance can inflate the band but cannot shift its center. Every downstream number – the Gibbs energy bands, the copula samples, the microstructure QoI distributions – inherits the Kozlov diagram's bias. The paper does not hide this, but the abstract and conclusion frame the output as uncertainty in the system, which overstates it. If the goal is a decision-support tool for this TE system, the calibration target needs to be experimental phase boundaries (or at least a sensitivity study over competing diagrams). If the goal is to demonstrate a general UP machinery, then the framing should say explicitly that the numbers are conditional on the chosen reference.\n\nMinor points: the ML classifiers are shown qualitatively with no accuracy metrics; '200,000 time series' vs '10,000 parameter combinations' is never reconciled; no code is released, only the data link. These are fixable with reporting effort.\n\nOverall: the framework is coherent, the math is straightforward and the demonstration is credible. The central limitation is external anchoring, not internal inconsistency. This paper deserves a serious referee; with revision it could be a solid methodological contribution to ICME-UQ. I'd send it out and ask the authors to recalibrate or reframe, add ML metrics, and clarify the dataset numbers.","headline":"A genuinely integrated UQ/UP pipeline for CALPHAD–phase-field modeling, but the uncertainty bands are conditional on a computed reference phase diagram, so treat the application numbers as a demonstration rather than predictive intervals.","tokens_in":23598,"tokens_out":3546,"would_cite":true,"duration_ms":35335,"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":"Thermodynamic parameter uncertainty can be propagated end-to-end through a CALPHAD and phase-field chain to yield probability distributions for microstructural descriptors in Mg2(SixSn1-x).","keywords":["uncertainty quantification","uncertainty propagation","CALPHAD","phase-field modeling","thermoelectric materials","Mg2(Si,Sn) system","Markov Chain Monte Carlo","Gaussian copula"],"falsifier":"Re-run the MCMC calibration using direct experimental composition-temperature data for the miscibility gap rather than the reference computed phase diagram, then compare the resulting 95% credible intervals on the phase boundary and on the microstructure QoIs with the intervals shown here; substantial non-overlap would show the propagated bands are conditional on the choice of calibration data.","tokens_in":22677,"feed_emoji":"🧊","tokens_out":8908,"duration_ms":80425,"temperature":0.7,"pith_summary":"This paper shows that uncertainty need not stop at the thermodynamic model: it can be carried all the way to microstructure predictions. The authors quantify the CALPHAD parameters of $\\mathrm{Mg}_2(\\mathrm{Si}_x\\mathrm{Sn}_{1-x})$ with a Bayesian MCMC calibration, sample the correlated posterior together with uncertainties in elastic and kinetic parameters, and feed those samples into an elasto-chemical phase-field model. The result is a distribution of microstructures rather than a single deterministic image, summarized by eight quantities of interest whose posterior statistics are mostly multimodal. A reader should care because decision-oriented materials design needs confidence bounds on process-structure predictions, and this work provides a concrete route to producing them through an expensive, high-dimensional simulation chain.","feed_headline":"Thermodynamic uncertainty reaches microstructure predictions","feed_subtitle":"Propagating CALPHAD error through phase-field runs gives probability bands for Mg2(Si,Sn) microstructures.","key_machinery":"The load-bearing object is the total free energy functional $F_{\\mathrm{tot}} = f_{\\mathrm{bulk}} + f_{\\mathrm{interfacial}} + f_{\\mathrm{elas}}$, in which the bulk chemical free energy is a sub-regular solution model with CALPHAD interaction parameters of the form $L^\\nu_{\\alpha\\beta} = {}^\\nu a_{\\alpha\\beta} + {}^\\nu b_{\\alpha\\beta} T$, and the microstructure evolution is governed by the Cahn-Hilliard equation $\\partial c/\\partial t = \\nabla \\cdot M\\nabla(\\delta F_{\\mathrm{tot}}/\\delta c)$ with microelasticity solved through the mechanical equilibrium and constitutive laws. This functional is the meeting point of the two uncertainty streams: the MCMC posterior on thermodynamic parameters enters through $f_{\\mathrm{bulk}}$, while the microelastic and kinetic priors enter through $f_{\\mathrm{elas}}$ and the mobility $M$. The argument is carried by treating this functional as a stochastic simulator: sample parameters, run the chain, collect microstructures, and summarize the output distribution through quantities of interest.","core_discovery":"The central claim is that statistically quantified uncertainty in thermodynamic parameters can be propagated through the Gibbs free energy, the equilibrium phase diagram, and then through a phase-field model that couples chemical and elastic driving forces, producing probability distributions for microstructural descriptors in the pseudobinary thermoelectric alloy $\\mathrm{Mg}_2(\\mathrm{Si}_x\\mathrm{Sn}_{1-x})$. The propagation is done in two stages: an MCMC calibration converts a prior on six CALPHAD interaction parameters into a posterior, and a Gaussian copula sampler draws parameter vectors from that posterior together with priors on elastic constants, stress-free transformation strain, molar volumes, mobility, gradient energy coefficient, and alloy composition. These samples drive 10,000 phase-field runs; the resulting synthetic microstructures are reduced to eight quantities of interest, and clustering separates them into decomposed versus not-decomposed classes. The paper's contribution is the end-to-end demonstration that meaningful uncertainty bands survive the chain and that the output space is not a single regime but a landscape with multiple modes.","pith_inferences":["Editorial inference: the same two-stage sampling scheme could be extended past microstructure to transport properties, turning the QoI distributions into confidence intervals for lattice thermal conductivity or the thermoelectric figure of merit; the paper stops at microstructure and lists phonon scattering as motivation.","Editorial inference: the strongest test of the calibration is to replace the reference phase diagram used in the MCMC likelihood with direct experimental phase boundary measurements of the same pseudobinary system; if the propagated credible intervals do not overlap the experimental boundaries, the posterior needs reweighting.","Editorial inference: because the output QoIs are multimodal, scalar summaries like variance-to-mean ratio may hide the regimes; a natural extension is to use the classifier decision boundaries as inverse maps that return the processing-parameter region for a target microstructure class."],"forward_implications":["If the framework is right, microstructure predictions from CALPHAD-based phase-field simulations can be reported as posterior distributions rather than point forecasts, with 95% credible intervals attached to phase boundary locations, phase compositions, characteristic lengths, and area fractions.","The high uncertainty seen in the Gibbs energy curves and phase diagram means that thermodynamic parameter uncertainty is not washed out by the subsequent simulation; downstream microstructure predictions inherit it and can change qualitatively, not just quantitatively.","The trimodal distributions of equilibrium phase compositions imply that deterministic min/mean/max sampling of inputs would miss whole microstructure regimes; sampling schemes that preserve correlations and marginal distributions are necessary.","Microstructure data can be organized into two broad classes, decomposed and not-decomposed, and classifiers can draw decision boundaries in input-parameter space that identify processing conditions favoring one or the other, connecting to different phonon-scattering regimes."],"supporting_citations":[{"why":"supplies the composition-temperature phase diagram data used as the calibration target for the MCMC inference.","marker":"[62]"},{"why":"provides the elasto-chemical phase-field model and the strain-induced suppression of the miscibility gap this work extends.","marker":"[39]"},{"why":"supplies the adaptive proposal scheme used to sample the CALPHAD parameter posterior.","marker":"[77]"},{"why":"supplies the semi-implicit Fourier-spectral solver for the Cahn-Hilliard evolution.","marker":"[72]"},{"why":"supplies the FFT-based iterative solver for the microelasticity equations with inhomogeneous elastic constants.","marker":"[29]"},{"why":"supplies the copula construction used to sample correlated parameter vectors.","marker":"[78]"}],"fun_headline_variants":["Propagating thermodynamic error to alloy microstructures","From CALPHAD to microstructure distributions with error bounds","Uncertainty propagation across phase-field model chains","End-to-end uncertainty quantification for thermoelectric alloys","Thermodynamic uncertainty shapes microstructure predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The chain is anchored by calibrating the CALPHAD parameters to one computed phase diagram, even though experimental phase boundary estimates for this system disagree by tens of atomic percent; if that reference diagram is biased, every downstream credible interval shifts with it.","fun_headline_variants_meta":{"raw":{"variants":["Propagating thermodynamic error to alloy microstructures","From CALPHAD to microstructure distributions with error bounds","Uncertainty propagation across phase-field model chains","End-to-end uncertainty quantification for thermoelectric alloys","Thermodynamic uncertainty shapes microstructure predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000951,"raw_usage":{"total_tokens":4104,"prompt_tokens":1041,"completion_tokens":3063,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2993}},"tokens_in":657,"tokens_out":3063,"duration_ms":21510,"temperature":1.0,"reasoning_tokens":2993,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:41:16.333950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the MCMC calibration using direct experimental composition-temperature data for the miscibility gap rather than the reference computed phase diagram, then compare the resulting 95% credible intervals on the phase boundary and on the microstructure QoIs with the intervals shown here; substantial non-overlap would show the propagated bands are conditional on the choice of calibration data.","supporting_citations":[{"cited_title":"Kozlov, J","cited_arxiv_id":null,"evidence_quote":"supplies the composition-temperature phase diagram data used as the calibration target for the MCMC inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the elasto-chemical phase-field model and the strain-induced suppression of the miscibility gap this work extends."},{"cited_title":"Haario, E","cited_arxiv_id":null,"evidence_quote":"supplies the adaptive proposal scheme used to sample the CALPHAD parameter posterior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the semi-implicit Fourier-spectral solver for the Cahn-Hilliard evolution."},{"cited_title":"Attari, A","cited_arxiv_id":null,"evidence_quote":"supplies the FFT-based iterative solver for the microelasticity equations with inhomogeneous elastic constants."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the copula construction used to sample correlated parameter vectors."}],"review_version":1}