{"id":"0daec193-45cb-4309-9733-89e5465a6a23","arxiv_id":"2412.10405","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Bayesian optimization calibrates a strain-gradient crystal plasticity model of Hastelloy X from cyclic stress-strain data and links simulated fatigue damage hotspots to large grains, high Schmid factor, and about 42 degrees misorientation.","lead":"The paper fits a detailed material model (nine adjustable parameters) for 3D-printed Hastelloy X metal using Bayesian optimization with a Gaussian process surrogate, then uses the fitted model to look for where fatigue cracks are likely to start. A generalist might read it because it promises cheaper, more automated calibration of fatigue simulation tools and identifies large, favorably oriented grains as probable failure sites.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 42-degree failure-site claim rests on a non-unique parameter set: macroscopic calibration constrains only C*sqrt(rho_SSD), leaving local GND/FIP fields undetermined.","rationale":"The strongest claim has two components: (1) BO achieves calibration with as few as 50 initial simulations within 75 iterations, and (2) the calibrated model identifies larger, high-Schmid-factor grains with about 42-degree misorientation as probable failure sites. Component (2) is the paper's principal scientific contribution and appears in the abstract, title, and conclusion, so an unaddressed threat to it is the most load-bearing. The reader's weakest assumption correctly identifies the transferability gap: macroscopic stress-strain calibration does not by itself determine local slip and GND fields. My analysis sharpens that gap with a concrete degeneracy visible in the paper itself. Equation 5 depends on C and rho_SSD mainly through C*sqrt(rho_SSD), and Section 3.5/Figure 11(e) explicitly shows the product converging while individual parameters keep fluctuating. The objective function (Eq. 15) therefore cannot distinguish parameter combinations along this valley from macroscopic stress-strain data. Since the local fatigue indicator field (maximum W) is computed from slip and GND fields that depend on the split, the Table 4 locations and the 42-degree plus/minus 1.67-degree statistic may be artifacts of the chosen point in the valley rather than material behavior. The proposed test perturbs C and rho_SSD along this valley and reruns the exact Section 4.3 analysis; if the failure-site ranking and misorientation statistics are stable, the concern is resolved, and if not, the microstructural conclusion requires additional constraints or experimental validation. This does not change the reader's CONDITIONAL verdict: the paper remains a competent calibration study, but the headline microstructural relationship should be read as conditional on parameter identifiability and validation. The efficiency claim is also weakened by the biased construction of the 50/25 initial sets (removing the best Delta-sigma cases from the 100-set), but that issue affects the methodological efficiency claim rather than the physical finding that is the focus of the paper.","tokens_in":24052,"tokens_out":6172,"duration_ms":60810,"concrete_test":"Fix C*sqrt(rho_SSD) at the calibrated product (about 0.915) and select two alternative parameter sets within the Table 2 bounds, e.g. (C=0.2, rho_SSD about 20.9) and (C=0.4, rho_SSD about 5.2), leaving all other parameters at Table 3 values. Confirm that each set reproduces the plus/minus 0.5% and plus/minus 0.75% cyclic stress-strain curves with Delta-sigma (Eq. 15) within about 1 MPa of the reported 15.06 MPa. Then rerun the five-cycle twin-RVE simulations of Section 4.3 for all 10 RVEs and recompute the grain with highest plastic strain energy density, the GND/diameter CDF values, and the average misorientation and Schmid factor statistics. If the failure-site grains change or the mean misorientation departs from 42 degrees plus/minus 1.67 degrees beyond the stated uncertainty, the microstructural claim is not determined by the calibration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central microstructural finding in Section 4.3—that maximum plastic strain energy density identifies grains with large diameter, high Schmid factor, and average misorientation roughly 42 degrees plus/minus 1.67 degrees (Table 4)—is not robustly connected to the calibrated parameters. The objective function (Eq. 15) fits only the homogenized stress-strain response; it does not constrain the individual values of C and rho_SSD except through their product in Eq. 5 (effective CRSS = tau_c^0 + C*G*b*sqrt(rho_for) + tau_c). In the small-strain regime the GND contribution to rho_total is secondary, so the yield stress depends essentially on C*sqrt(rho_SSD). Section 3.5 explicitly observes this coupling (Figure 11(e)), reporting convergence of the product near 1 while C and rho_SSD continue to fluctuate. The BO optimum (C=0.1, rho_SSD=83.7, Table 3) is therefore one point on a valley of near-equal macroscopic fits. Local fields are not invariant along this valley: rho_SSD enters rho_total and hence the forest density (Eqs. 6-8), and GND densities derive from slip gradients that depend on the local hardening. No demonstration is given that alternative parameter sets with equivalent Delta-sigma produce the same maximum-W grains, the same CDF rankings, or the same 42-degree average misorientation. Since the paper does not validate predicted failure sites against experimentally observed crack initiation, the reported relationship remains conditional on the non-uniqueness of the calibration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a Bayesian optimization (BO) framework with a Gaussian-process surrogate to calibrate a nine-parameter strain-gradient crystal plasticity model for laser powder bed fusion Hastelloy X using cyclic stress-strain data at ±0.5% and ±0.75% strain amplitudes and 500 °F. A modified objective function that penalizes cycle-to-cycle hardening is introduced, and the authors report that calibration is achieved within 75 iterations with as few as 50 initial simulations. SHAP-based sensitivity analysis is used to rank parameter influence along the hysteresis loop, and synthetic RVEs with and without twins are compared. The authors conclude that twin boundaries alter local fields but not the macroscale response, and that grains with larger diameter, high Schmid factor, and average misorientation of about 42°±1.67° are probable fatigue failure sites.","tokens_in":24354,"tokens_out":4085,"duration_ms":46848,"significance":"If the claims are established, the work would be a useful contribution to efficient crystal-plasticity calibration, which is a recognized bottleneck. The use of cyclic data at multiple strain amplitudes, the explicit treatment of cycle-to-cycle hardening in the objective, and the inclusion of twins in synthetic microstructures address relevant gaps in the literature. The paper also benefits from concrete computational infrastructure: the OXFORD-UMAT is public, the RVE generation is detailed, and 10 RVEs are used. These strengths make the calibration workflow potentially reproducible and transferable. However, the central efficiency claim and the microstructural failure-site claim are each weakened by load-bearing issues: the reduced-initial-data experiment is not a fair test, and the calibrated parameters demonstrably lie on an identifiability valley, so the grain-scale predictions are not shown to be robust.","major_comments":[{"comment":"The conclusion that 'as few as 50 initial simulations' suffice is not supported by the experiment as conducted. The text states that the 50- and 25-point initial datasets are taken from the 100-point set 'by removing the best performing, i.e., simulations that have low Δσ values.' This is not a realistic 50-point initial design: it removes the most informative points, so the retained set is biased to look better than a fresh LHS-50 sample. The comparison therefore does not establish the stated efficiency claim. The experiment should be repeated with independent LHS draws of size 50 and 25, or at least with repeated random subsampling of the 100-point set, reporting the distribution of achieved Δσ.","section":"Sec. 3.4, Fig. 9"},{"comment":"The calibration is overparameterized with respect to the macroscopic data, and the authors themselves observe the coupling in Figure 11(e): the objective constrains essentially the product C*sqrt(rho_SSD), while C and rho_SSD individually continue to fluctuate. This is consistent with Eq. (5), because at the small applied strains (<0.0075) the GND contribution to rho_total is secondary and the yield stress depends mainly on C*sqrt(rho_SSD). The optimized values C=0.1 and rho_SSD=83.7 in Table 3 are therefore one point on a near-equal-fit valley, not a uniquely identified parameter set. Since rho_SSD enters the forest density through Eq. (8) and hence affects the GND-influenced local fields used in Section 4.3, the microstructural failure-site statistics in Table 4 are conditional on that arbitrary point. The paper should either demonstrate that alternative parameter sets with equivalent macroscopic Δσ produce the same maximum-W grains, the same CDF rankings, and the same ~42° misorientation statistic, or it should reframe the microstructural findings as conditional on the selected point. The fact that C converges to the lower bound of the search range further suggests an active boundary rather than an identified optimum.","section":"Sec. 3.5, Eq. (5), Table 3"},{"comment":"The central microstructural claim—that larger grains with high Schmid factor and average misorientation of about 42°±1.67° are probable failure sites—is not validated against experimental observations of crack initiation. The simulations use the optimized parameters from Table 3 and the maximum plastic strain energy density to identify failure sites, but no comparison is made to measured crack locations, EBSD-based damage maps, or fractography from the experimental program cited as [37]. A macroscopic stress-strain fit cannot by itself establish that grain-scale FIP fields are correct, especially given the parameter non-uniqueness noted above. At minimum, the paper should present the 42° result as a model prediction requiring validation, and ideally it should compare against available experimental failure-site data for the same or similar specimens.","section":"Sec. 4.3, Table 4"},{"comment":"The synthetic RVEs contain a twin fraction of only 0.4–0.5, whereas the EBSD measurements show a twin fraction of 0.62–0.73. The authors acknowledge this mismatch but do not assess its effect on the twin-related conclusions. Since Sections 4.1–4.3 rely on the presence and placement of twin boundaries to make claims about strain energy density accumulation, TB-based failure sites, and the 42° misorientation statistic, the systematic under-representation of twins is a load-bearing limitation. A sensitivity study with higher twin fractions, or at least a discussion of whether the conclusions are expected to be insensitive to twin fraction, is needed before those claims can be accepted.","section":"Sec. 2.2.1 and Secs. 4.1–4.3"}],"minor_comments":[{"comment":"Equation numbers are duplicated: Eq. (12) is used for both the Gaussian process prior and the effective plastic strain rate, and Eq. (13) is used for both the expected improvement acquisition function and the plastic strain energy density. Please renumber.","section":"Sec. 2.3.1 and Sec. 2.2.2"},{"comment":"The text reports R² scores but does not explicitly state whether the quoted values are for the held-out 30% test set or for the training set. Please clarify, and if possible report both training and test R².","section":"Sec. 3.2, Fig. 6"},{"comment":"The caption says 'number of iterations to achieve optimal solution,' but the algorithm is run for a fixed budget of 75 iterations and the reported value is the best found within that budget. Rephrase to 'best value found within the iteration budget' to avoid implying that a stopping criterion detected a global optimum.","section":"Sec. 3.4, Fig. 9"},{"comment":"The definition of 'average misorientation of a grain' should be made more precise: it is defined as a number-weighted average of misorientation with neighboring grains, but the weighting rule and angular cutoff used should be stated explicitly.","section":"Sec. 4.3"},{"comment":"The symbol tau_c^a appears to denote both the current slip resistance in Eq. (9) and a term in Eq. (5) described as 'strength due to statistical hardening.' The notation is confusing and should be disambiguated.","section":"Sec. 2.2.2, Eq. (5)"},{"comment":"Reference [51] is a duplicate of [34]; please consolidate or cite the original source once.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid practical motivation and the BO/GP workflow is mostly clearly presented. My concern is that the headline efficiency claim is based on an unfair comparison, and the microstructural failure-site claim is built on a non-unique calibration without validation. Both issues are fixable in principle: re-running the reduced-initial-data study properly, and adding a parameter-invariance or experimental-validation check. This fits the journal's scope, but I would not accept without those changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on arXiv:2412.10405. The Bayesian optimization calibration of the strain-gradient CP model is competent, and the two-strain objective with the inter-cycle hardening penalty is a genuinely useful addition. The SHAP sensitivity analysis and the twins-vs-no-twins comparison are also reasonable computational exercises. The GP surrogate R² around 0.92 and the convergence behavior are plausible. So there is real value here for people doing CP parameter identification.\n\nThe soft spots are where the claims outrun the evidence. The 'as few as 50 initial simulations' result is weakened by the construction: the authors took the 100-point LHS set and removed the best-performing points. That is not a realistic 50-point initial design, so the efficiency claim is partly an artifact. The stress-strain agreement is a fitting error, not an independent validation, and the elastic constants were hand-adjusted.\n\nThe bigger issue is the Section 4.3 conclusion about failure sites. Larger grains with high Schmid factor and average misorientation around 42° are identified as likely failure sites. But the calibration only constrains the macroscopic homogenized response. The paper itself shows that C and rho_SSD are coupled: the product C*sqrt(rho_SSD) converges while the individual parameters keep fluctuating. Local fields like GND density and plastic strain energy density are not invariant along that valley. The BO optimum in Table 3 is one point on a near-flat direction, and there is no demonstration that alternative parameter sets with the same macroscopic fit produce the same failure-site predictions. Since the predicted sites are not compared to experimentally observed crack initiation locations, the 42° finding is conditional on one non-unique parameter choice. The SHAP result that hardening parameters dominate the inter-cycle stress difference is also partly circular, because that exact difference was inserted into the objective function.\n\nNone of this kills the core contribution. The calibration framework works and the objective term is useful. But the microstructural fatigue claim is over-sold. The paper should be read as a calibrated-model study with plausible but unvalidated microstructure predictions. It deserves a serious referee, but the authors should be pushed to either validate the failure-site predictions against experiment or substantially soften the claim and discuss the non-uniqueness. I would not cite the 42° result as established. The BO calibration approach itself I would cite.","headline":"Competent BO calibration study with a useful new objective term, but the headline microstructural failure-site claim is built on a non-unique parameter set and needs to be walked back or validated.","tokens_in":24909,"tokens_out":3395,"would_cite":true,"duration_ms":30338,"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":"Bayesian optimization calibrates a nine-parameter crystal-plasticity model of additively manufactured Hastelloy X and identifies large, favorably oriented grains at twin boundaries as fatigue-failure sites.","keywords":["crystal plasticity","Bayesian optimization","Gaussian process surrogate","Hastelloy X","low-cycle fatigue","twin boundaries","additive manufacturing","fatigue indicator parameter"],"falsifier":"Interrupt fatigue tests of the same L-PBF Hastelloy X at 500°F, locate the first cracks by post-test EBSD or in-situ imaging, and compare those sites with the grains the model flags as having the highest accumulated plastic strain energy density; if early cracks appear in small, poorly oriented, low-misorientation grains, the proposed $42^\\circ$ large-grain criterion is falsified.","tokens_in":23776,"feed_emoji":"🔬","tokens_out":11485,"duration_ms":102665,"temperature":0.7,"pith_summary":"This paper establishes that a nine-parameter strain-gradient crystal-plasticity model of additively manufactured Hastelloy X can be calibrated to measured fatigue data by Bayesian optimization with a Gaussian-process surrogate, using as few as 50 initial simulations and at most 75 optimization iterations. The authors add a second term to the usual stress-error objective that penalizes the small hardening difference between consecutive loading cycles, which is what lets the fit match two stable experimental cycles at both 0.5% and 0.75% strain amplitude. The calibrated model is then used to ask which microstructural features concentrate fatigue damage, measured by accumulated plastic strain energy density. It finds that the predicted failure sites are almost always at twin boundaries in the largest grains, with high Schmid factor (favorable slip orientation) and an average neighboring-grain misorientation of about 42°±1.67°, giving a concrete microstructural fingerprint for low-cycle fatigue in this alloy.","feed_headline":"Calibration finds fatigue failure sites in printed Hastelloy X","feed_subtitle":"Fits nine crystal-plasticity parameters in 75 simulations; links fatigue damage to large, favorably oriented grains.","key_machinery":"The load-bearing machinery is the coupling of a Gaussian-process surrogate model with expected-improvement Bayesian optimization, wrapped around a strain-gradient crystal-plasticity finite-element model of a representative volume element (a 200-µm cubic patch of about 300 grains). The Gaussian process learns the map from the nine constitutive parameters to an objective $\\Delta\\sigma$ that combines the root-mean-square error between simulated and experimental stress at 44 points on two consecutive cycles with a $\\lambda$-weighted term penalizing cycle-to-cycle hardening mismatch. The strain-gradient formulation matters because it computes geometrically necessary dislocation densities from the curl of the plastic deformation gradient, and those densities feed both the effective slip resistance and the accumulated plastic strain energy density $W$ used as the fatigue indicator; this is the channel through which a macroscopically calibrated model produces grain-scale failure-site predictions.","core_discovery":"The central claim is that a single set of nine strain-gradient crystal-plasticity parameters, found by Bayesian optimization against macroscopic stress-strain loops, reproduces the 500°F cyclic response of L-PBF Hastelloy X at ±0.5% and ±0.75% strain and provides trustworthy grain-scale fatigue-damage indicators. The optimized fit reaches an average $\\Delta\\sigma$ of 15.06 MPa in 61 iterations once the objective includes a $\\lambda$-weighted hardening-mismatch penalty, with as few as 50 initial Latin-hypercube simulations; SHAP-based sensitivity analysis shows the yield-related parameters ($\\rho_{SSD}^a$, $\\tau_c^0$, $C$) dominate the stress response, with backstress parameters gaining influence during compressive loading. In ten 200-µm RVEs with and without inserted twins, the homogenized stress-strain curves are nearly identical, but the location of maximum plastic strain energy density shifts from a normal grain boundary to a twin boundary when twins are present, and the overall strain-energy distribution shifts to higher values. Across all RVEs, the grain with the highest accumulated $W$ sits near the top of the grain-size distribution, has a Schmid factor typically near 0.46, and has an average misorientation of $42^\\circ \\pm 1.67^\\circ$ with its neighbours, which the paper proposes as the probable fatigue-failure criterion.","pith_inferences":["An implication left implicit is that the $42^\\circ$ misorientation band should be checked against measured grain-boundary character distributions; the paper's EBSD shows a high twin fraction, so the 39°–45° range may partly reflect the prevalence of $\\Sigma3$ boundaries rather than a damage-specific preference.","A testable extension would be to calibrate the model on one strain amplitude and predict the other: the paper calibrates both amplitudes together, so it does not demonstrate that the parameters generalize beyond the fitted loading conditions.","The twin fraction in the RVEs (0.4–0.5) is below the EBSD-measured range (0.62–0.73), so re-running the failure-site analysis at the measured twin fraction would show whether the large-grain/$42^\\circ$ criterion is robust to twin density.","The sensitivity ranking suggests a cheaper variant of the method: fix low-influence parameters at reasonable values before optimization and use the remaining parameters in Bayesian optimization, then compare fit quality; the paper identifies which parameters matter but does not run this ablation."],"forward_implications":["Calibrating a new crystal-plasticity model needs no more than about 50 initial simulations and 75 optimization rounds, so inverse calibration becomes feasible on a single workstation.","Adding a $\\lambda$-weighted penalty for cycle-to-cycle hardening mismatch makes the simulated loops reproduce the near-overlapping experimental cycles, which a plain stress-error objective fails to do.","Synthetic microstructures without twins can reproduce the macroscopic stress-strain curve yet mislocate the predicted fatigue-initiation site, so twin boundaries must be included for micromechanical fatigue studies.","The identified failure-site fingerprint—large grain diameter, high Schmid factor, and average misorientation near $42^\\circ$—can serve as a screening criterion for fatigue-prone microstructures in L-PBF Hastelloy X at 500°F.","Because the optimization and sensitivity analysis are formulated at the level of the constitutive parameters and the stress-strain objective, the same Bayesian optimization procedure transfers to other alloys and other crystal-plasticity laws."],"supporting_citations":[{"why":"Supplies the L-PBF Hastelloy X specimens, heat-treatment history, and strain-controlled fatigue data at 500°F and ±0.5/±0.75% strain that the BO objective targets.","marker":"[37]"},{"why":"Provides the strain-gradient crystal-plasticity implementation in which the calibrated constitutive equations and dislocation-density fields are computed.","marker":"[45]"},{"why":"Formulates the geometrically necessary dislocation density calculation from the plastic deformation gradient used in the flow rule and fatigue indicator.","marker":"[47]"},{"why":"Earlier cyclic-calibration benchmark for IN718 by genetic algorithm that motivates the Bayesian optimization efficiency claim and supplies the parameter search ranges.","marker":"[6]"},{"why":"Prior microstructure-sensitive Hastelloy X crystal-plasticity fatigue-life model without twins, which the twin-insertion study directly extends.","marker":"[35]"},{"why":"Open-source synthetic microstructure generator used to create the 3D RVEs with grain size, orientation, and twin statistics derived from EBSD.","marker":"[41]"},{"why":"Provides the Gaussian-process Bayesian optimization implementation, including surrogate training and expected-improvement acquisition, used for parameter identification.","marker":"[54]"},{"why":"Prior demonstration of Bayesian optimization for fatigue-related crystal-plasticity backstress parameters, the closest methodological baseline the paper builds on.","marker":"[30]"}],"fun_headline_variants":["Bayesian calibration finds fatigue hotspots in printed superalloy","75 runs fit crystal model, reveal fatigue sites in Hastelloy X","Big grains with high Schmid factor fail in 3D-printed metal","Optimized crystal model pinpoints fatigue failure in printed alloy","Twin boundaries shift fatigue damage in L-PBF Hastelloy X"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The grain-scale failure predictions assume that parameters fitted to the overall stress-strain loops also give the correct local deformation and dislocation patterns inside grains, since the model is never checked against measured crack-start locations; the twin fraction in the RVEs is also below the measured EBSD range.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian calibration finds fatigue hotspots in printed superalloy","75 runs fit crystal model, reveal fatigue sites in Hastelloy X","Big grains with high Schmid factor fail in 3D-printed metal","Optimized crystal model pinpoints fatigue failure in printed alloy","Twin boundaries shift fatigue damage in L-PBF Hastelloy X"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1617,"prompt_tokens":1108,"completion_tokens":509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":724,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":724,"tokens_out":509,"duration_ms":5753,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:50:02.273298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Interrupt fatigue tests of the same L-PBF Hastelloy X at 500°F, locate the first cracks by post-test EBSD or in-situ imaging, and compare those sites with the grains the model flags as having the highest accumulated plastic strain energy density; if early cracks appear in small, poorly oriented, low-misorientation grains, the proposed $42^\\circ$ large-grain criterion is falsified.","supporting_citations":[{"cited_title":"Fatigue life prediction of rough Hastelloy X specimens fabricated using laser powder bed fusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the L-PBF Hastelloy X specimens, heat-treatment history, and strain-controlled fatigue data at 500°F and ±0.5/±0.75% strain that the BO objective targets."},{"cited_title":"The role of defects and critical pore size analysis in the fatigue response of additively manufactured IN718 via crystal plasticity,","cited_arxiv_id":null,"evidence_quote":"Earlier cyclic-calibration benchmark for IN718 by genetic algorithm that motivates the Bayesian optimization efficiency claim and supplies the parameter search ranges."},{"cited_title":"DREAM.3D: A Digital Representation Environment for the Analysis of Microstructure in 3D,","cited_arxiv_id":null,"evidence_quote":"Open-source synthetic microstructure generator used to create the 3D RVEs with grain size, orientation, and twin statistics derived from EBSD."},{"cited_title":"GitHub - SheffieldML/GPyOpt: Gaussian Process Optimization using GPy","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian-process Bayesian optimization implementation, including surrogate training and expected-improvement acquisition, used for parameter identification."},{"cited_title":"Identifying material parameters in crystal plasticity by Bayesian optimization,","cited_arxiv_id":null,"evidence_quote":"Prior demonstration of Bayesian optimization for fatigue-related crystal-plasticity backstress parameters, the closest methodological baseline the paper builds on."}],"review_version":1}