acceptodds
Under review as a conference paper at ICLR 2027

CRYSTALMO-TURBO: MULTI-OBJECTIVE TRUST- REGION BAYESIAN OPTIMIZATION FOR HIGH- PRECISION JOINT CRYSTAL STRUCTURE REFINEMENT

Abstract

Crystal structure refinement is a fundamental inverse problem in materials characterization, where structural parameters are optimized to reproduce experimental diffraction data. Conventional approaches, such as least-squares and likelihood-based optimization, rely on local search and often struggle with non-convex, noisy, and highly correlated parameter landscapes, particularly when integrating multiple diffraction modalities. Joint refinement of X-ray and neutron data is especially challenging due to their complementary but competing sensitivities, which are typically combined through scalarized objectives requiring manual weighting and leading to suboptimal solutions. We propose CrystalMO-TuRBO, a multi-objective trust-region Bayesian optimization architecture for joint crystal structure refinement. The method models X-ray and neutron discrepancies as separate objectives and transforms the problem into a normalized maximization setting. A two-phase optimization strategy is introduced: Phase 1 performs global exploration using parallel trust-region Bayesian optimization across multiple scalarizations to identify promising regions of the parameter space, while Phase 2 conducts localized refinement within a shrinking region to achieve high-precision solutions. This design explicitly separates global search from fine-grained optimization, addressing the unique accuracy requirements of refinement tasks. We evaluate the proposed method on experimentally collected X-ray and neutron diffraction data from single-crystal HoTiO. Results demonstrate improved convergence, robustness, and parameter precision compared to classical refinement methods and Bayesian optimization baselines on refinement of a single-crystal pyrochlore material system.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Bayesian optimization, crustallographic refinement, materials characterization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rbccOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Crystal structure refinement is an inverse problem in which atomic positions, occupancy (Occ: the proportion of atoms present at a site), and atomic displacement parameters (U: the magnitude of thermal vibrations) are fitted to match the reflection intensities measured by diffraction. This paper formulates the simultaneous refinement of single-crystal X-ray diffraction (XRD) and neutron diffraction (ND) as a two-objective Bayesian optimization (BO) problem aimed at minimizing two intensity residuals, fXRDf_{XRD} and fNDf_{ND}, and proposes CrystalMO-TuRBO (CMT). In Phase 1, CMT employs four linear scalarizations with different weights (combining the two objectives into a single objective via a weighted sum; λ = 0.15, 0.45, 0.65, 0.85) , each with a different weight (combining the two objectives into a single objective via a weighted sum; λ = 0.15, 0.45, 0.65, 0.85). In Phase 2, CMT performs local optimization of the scalarization with λ = 0.5 using qEI (the batch version of the expected improvement) within a box of fixed width around the best point (width specified as 1/2/5/10% of the range).

Using a single set of experimental XRD and ND data for Ho2Ti2O7 single crystals, the authors examined two conditions, primary extinction correction (PE) and secondary extinction correction (SE) on the neutron side, and compared the results of simultaneous least-squares (LS) refinement using the standard crystal structure refinement software GSAS-II with those of qEHVI (Expected Hypervolume Improvement; a method that directly improves standard metrics for multi-objective optimization), ParEGO (random-weight scalarization), MORBO (multi-objective Bayesian optimization for high dimensions using multiple confidence regions; Daulton et al., UAI 2022; here, the authors’ reimplementation), and CMT, within a budget of 400 evaluations. The authors report that, based on single-run results (Table 1) and the averages of iterative experiments (Tables 2 and 3), CMT using a 10% box achieved the smallest fXRD+fNDf_{XRD} + f_{ND} (PE 23.79, SE 23.55 on average across iterative experiments), and they also present ablation studies on box width and the inclusion or exclusion of Phase 2.

Critical strengths
  1. The authors introduced a problem formulation into machine learning, treating the simultaneous refinement of XRD and ND as two separate objectives rather than a weighted sum, along with a set of measured data for the same substance. As motivation, this paper presents an example where, when optimized separately, the Ho occupancy splits into 1.100 for XRD and 0.827 for ND (§4.2, Tables 6 and 7). However, caution is required when interpreting this discrepancy. There is the possibility that the parameters are not fixed (Critical Weaknesses 2) and the question of whether the XRD and ND samples are the same (Presentation, Questions For Authors 5); the 1.100 value from the XRD side appears to be the upper limit of the search range in Table 6. Reason for Impact on Overall Evaluation: This paper presents a new real-world problem in scientific applications; if properly re-evaluated, this problem and the data could serve as a benchmark for machine learning.
Critical weaknesses
  1. In the repeat experiments for the main results, the results for the proposed method and MORBO appear to have been swapped when saved, meaning that the conclusion that the proposed method outperforms MORBO (§4.1: “CMT also consistently outperforms the representative MORBO baselines”) is not supported by the authors’ data. Even within the main text, the MORBO row in Table 2 (fXRD+fNDf_{XRD} + f_{ND} = 28.77 ± 2.51) shows values in each column that are nearly identical to those in the row for CMT’s Phase 1 alone (28.69 ± 2.43–2.47) (the same applies to Table 3), the MORBO convergence curve (Fig. 2 B1) is also nearly identical to the curve for CMT Phase 1 alone (Fig. 3 A), and the x-axis extends to approximately 420 evaluations, the same as for CMT, rather than the 400 used for other methods. In the anonymous OSF (Open Science Framework, a service for publishing research data) repository provided by the authors in their Reproducibility Statement (containing code, diffraction data, and results for each seed from repeat experiments), the results in the MORBO folder are identical to those from the CMT run consisting solely of Phase 1; conversely, some runs in the CMT 10% folder are labeled as MORBO. When reviewers recalculate the data as recorded, CMT’s advantage over MORBO disappears (there is no significant difference in PE, and the mean is actually lower for MORBO; in SE, MORBO performs better. See Questions For Authors 1 for the numerical values). Reason for impact on the overall evaluation: The paper’s central claims (§4.1, §5) do not hold when the publicly available data is correctly categorized.

  2. The claim in the title, “high-precision joint crystal structure refinement,” is supported solely by a comparison of residuals; no evidence is provided to demonstrate that the parameters were determined with high precision. Neither the standard crystallographic fit metrics (R1: relative residual of the structure factor; wR2: relative residual of the intensity weighted by measurement error; GoF: goodness of fit calculated by dividing the weighted residual by the degrees of freedom) nor the standard uncertainties of the parameters are reported in this paper. Furthermore, in the adopted solution (the CMT 10% Phase 1+2 solution for PE, row labeled “CMT 10% Ph 2” in Table 1), the occupation fractions of both Ti and Ho are at the lower limit of the search range, 0.800. At the best points of the iterative experiments, even among the 12 runs recorded as CMT (PE, seeds 0–11), the occupancy rates are scattered across the entire range of 0.8 to 1.2 (CMT_10__Ph_1_2/PE_best_by_sum_per_seed.csv). Since the computational intensity is proportional to the scale k, and the occupancy also affects the overall intensity through the structure factor, there are combinations where increasing one parameter while decreasing the other results in virtually no change in the residuals. Due to this strong correlation (a relationship approaching degeneracy), it is highly likely that the parameters are not well-defined (this correlation is addressed in “Questions for Authors 4”). Furthermore, the primary metric fXRD+fNDf_{XRD} + f_{ND} is a quantity of the same form as the scalarized (normalized, equally weighted sum of two objectives) λ = 0.5 that Phase 2 minimizes; there is no comparison with multi-objective metrics such as hypervolume, nor with single-objective TuRBO for the same sum (Questions For Authors 6). Reason for impact on the overall evaluation: Even if Critical Weaknesses 1 is addressed, the fact that the residuals are small does not necessarily mean that the structural parameters were determined with high precision, and thus the claim in the title cannot be substantiated. Furthermore, the fact that CMT outperforms others in terms of the sum of residuals could simply be a consequence of CMT itself optimizing a quantity that has the same form as that sum.

Questions for authors
  1. [Clarification] Please explain why the results in OSF’s Repeated Run Results/MORBO/ are identical to those in CMT_2__Ph_1/ (CMT Phase 1 only), and why "method": "MORBO" is recorded in the summary.json files for CMT_10__Ph_1_2/ starting from seed_13 onward. If categorized as recorded, the CMT values in Tables 2 and 3 (23.79, 23.55) match the average of this mixture; however, based solely on the runs recorded as CMT (421 evaluations), the PE is 24.13 and the SE is 25.08. In the reviewer’s recalculation, compared to what appear to be MORBO runs (PE n=7, SE n=8), there was no significant difference in PE (Mann–Whitney U: two-sample test based on ranks, p ≈ 0.70. The means are 24.13 for CMT and 23.23 for MORBO), while MORBO performs better in terms of SE (p ≈ 0.010). The main text (§4.1, p. 7; §4, p. 5, conveys the same meaning) states that “the improvements achieved by CMT over the competing methods are statistically significant at the 0.05 significance level”; however, in the authors’ test CSV, even after Holm correction (correction for multiple comparisons), the p-value for SE versus MORBO is p = 0.056, and the number of iterations in the aggregated CSV ranges from 9 to 21 per method. Could you correct the transposition and recreate Tables 2 and 3, clearly indicating the number of iterations (n) for each row and the p-values with the correction method specified? If CMT performs significantly better even when compared to the correct MORBO, the concern in Critical Weaknesses 1 will be resolved, and the evaluation will be upgraded (regarding the claim of high accuracy, see Questions For Authors 2 and Questions For Authors 3). (Critical Weaknesses 1)

  2. [Clarification] Can you demonstrate the extent to which each method can recover the true parameters using data (with noise) synthesized from a known structure? If CMT performs better in terms of recovery error, the claim of high accuracy will be supported by the parameters rather than the residuals, and the evaluation will be upgraded. (Critical Weaknesses 2)

  3. [Clarification] Please explain the conditions for LS simultaneous refinement. The LS (PE) in Table 1 is fXRD+fND=57.2f_{XRD} + f_{ND} = 57.2, but the LS from the iterative experiments in Table 2 is 102.35 ± 27.48. As a definitive method, both the difference and the variability are large, and it is unclear what the value of 400 in the “NIterations” column of the table represents. In the publicly available data, the LS for the iterative experiments appears to start randomly across the entire search range and then clip the solution to the range after refinement ("start_mode": "random", "clipped_to_mobo_bounds": true in Repeated Run Results/LS/*/seed_k/summary.json). I would like to know the initial values, the weights based on σ (measurement error) and the histogram weights, the constraints, and how the starting points are determined. While GSAS-II’s simultaneous least squares typically minimizes residuals weighted by σ, this paper evaluates each method using ff without σ weights. Can you present a comparison that evaluates LS, which starts from nominal values and applies constraints, using the weighted residuals minimized by LS itself, as well as R1/wR2/GoF? If CMT outperforms LS in that comparison as well, the claim of high accuracy will be supported by standard crystallographic metrics, and the evaluation will be elevated. (Critical Weaknesses 2)

  4. [Clarification] Do you consider the Ti and Ho occupancy of 0.800 (lower limit) in the adopted solution to be physically reasonable? Can you present the mean and standard deviation of each parameter from repeated experiments, as well as the correlation between scale and occupancy? The paper cites Agada et al. (2025), which refined the XRD data for the same material, Ho$_2$Ti$_2$O$_7$, using BO, sets the occupancy range at 0.9 to 1.1. Furthermore, in Table 6 of this paper (which presents BO applied to XRD data alone following the procedure in that paper), the TuRBO-1 solution yields OccTi 0.900 and OccHo 1.100; if the range is the same as that of Agada et al., these values correspond to the upper and lower limits of that range. I would also like to know why the authors expanded the range to 0.8–1.2 through simultaneous refinement. If it can be demonstrated that the occupation ratio is stably determined within this range, some of the concerns raised in Critical Weaknesses 2 would be alleviated. (Critical Weaknesses 2)

  5. [Clarification] Are the crystals used for XRD and ND from the same sample? According to Appendix D, the XRD crystal is a 0.1 mm crystal grown using the TSFZ method (solvent-based floating-zone method), while §3.1 states that the ND crystal is a 2 mm crystal grown using the FZ method (floating-zone method); however, the main text treats them as a single crystal in §4.2 and §4.4. If they are different samples, the discrepancy in occupancy in §4.2 could be explained by the difference in samples, and the reviewer would reduce the weight assigned to Critical Strengths 1. (Critical Strengths 1・Presentation)

  6. [Clarification] If possible, please present the results of running the single-objective TuRBO-1 on 400 evaluations for the sum at λ = 0.5 (using the same standardization), along with the hypervolume for all methods (the size of the region in the objective space dominated by the set of obtained solutions, measured at a common reference point). If CMT outperforms the former and does not fall short in hypervolume, the benefits of multi-objective optimization and the two-stage structure will be demonstrated, and the evaluation will improve. (Critical Weaknesses 2)

  7. [Clarification] Among the discrepancies between the main text and the implementation, I would like to confirm two points that directly affect the results. (1) The authors’ implementation (gsas_nd_9d.py) replaces |U| ≤ 10⁻⁶ with 0.01 only for the ND evaluation; what is the reason for this? There are many solutions where U is close to 0 (in Table 1, many U values are 0.000; even among the best results from the published iterative experiments, a total of 20 components in the 21 best runs from the SE’s CMT 10% folder, including the run recorded as MORBO, have |U| ≤ 10⁻⁶. This is an issue with the ND evaluation function common to all methods) . In such solutions, XRD and ND are evaluated at different U values. If the rankings remain unchanged after re-evaluating the results excluding this substitution, some of the concerns regarding the validity of the evaluation metrics (residuals) themselves, as discussed in Critical Weaknesses 2, would be alleviated. (2) CMT has 421 evaluations, while the baseline has 400 evaluations (including the initial point), so the budgets do not match (Section 4 sets the budget to 400 for all methods). If CMT still outperforms the baseline even after aligning the budgets to 400, this confirms that the comparison with MORBO in “Critical Weaknesses 1” was conducted under the same budget conditions. (Critical Weaknesses 1, Critical Weaknesses 2)

4 comments
AuthorsOpenReview ↗

Response to Reviewer rbcc

Thank you for the detailed feedback on this manuscript. For clarity, we have labelled the response as Clarifications 1 to 7, corresponding to the same clarifications 1 to 7 in the reviewer's comments.

Response to Clarification 1: We thank the reviewer for carefully identifying the apparent inconsistency between the implementation records and the reported CMT/MORBO results. We investigated the archived runs and confirmed that the "method": "MORBO" entries identified by the reviewer in the summary.json files from seed_13 onward were caused by a metadata-labeling error. The method tag in the notebook was not updated when the underlying script being executed was changed to the CMT implementation. This occurred because seeds 1–12 and seeds 13 onward were executed on different machines by different individuals. Importantly, this affected only the recorded method label in summary.json; it did not change the algorithm that was executed, the parameter values generated, or the objective values obtained. On the observation on similarity of CMP Phase 1 and MORBO results, the similarity is expected given the closely related structure of the two methods. Both approaches use multiple trust regions to conduct surrogate-based multi-objective Bayesian optimization in parallel, allowing different regions of the search space and different objective trade-offs to be explored simultaneously. MORBO uses several trust regions with scalarized objective vectors, while CMT Phase 1 similarly decomposes the XRD–ND problem into multiple scalarized trust-region subproblems. Thus, the comparable convergence behavior between MORBO and CMT Phase 1 is not unexpected: both methods are designed to perform broad multi-objective exploration through parallel trust-region surrogate optimization. In CMT, the four scalarizations provide simultaneous exploration of different XRD–ND trade-offs, while the trust-region mechanism focuses the search within the promising regions of the refinement landscape. The similarity in the Phase 1 results should therefore be understood as a consequence of the related exploration strategies used by the two methods, rather than as evidence of an implementation or result mix-up. The full CMT framework subsequently applies its Phase 2 local refinement stage to further exploit the promising region identified during this exploration. We have also re-checked the results used to generate Tables 2 and 3 against the underlying repeated-run result files. The reported CMT and MORBO summary values were based on the intended experiment outputs rather than on the erroneous method labels in the summary.json metadata. For the comparison of the full CMT (Phase 1 + Phase 2) implementation and MORBO, the raw and corrected p-values for refinement with PE correction are 0.0000205 and 0.001068 respectively, implying the difference is significant. However, for refinement with SE correction, the raw and corrected p-values are 0.001375 and 0.056 respectively. Hence, the significance for SE is in the borderline. The results (for PE and SE) are in serial 170 and 368 (respectively) of significance_pairwise_within_family.csv. Please note that refinement with PE and SE corrections are conducted for the same problem/data, the only distinction is in the extinction parameters being refined in each case. Considering how significant a little perturbation in the parameter values can on material properties, the difference observed is not negligible. We will revise Tables 2 and 3 to explicitly report the number of repeated runs (n) for each method, together with the corresponding statistical tests and p-values.

Response to Clarification 2: We thank the reviewer for this suggestion. We did attempt this approach using synthetic XRD/ND data generated from known structural parameters, which provided a direct ground truth for evaluating parameter recovery. However, we encountered an important limitation: the conventional LS refinement repeatedly failed because of high parameter correlations, even when different levels of noise were introduced into the synthetic data. In contrast, the BO methods were able to recover the actual parameter values very easily, essentially reproducing the known ground-truth parameters across the different BO methods. Consequently, this synthetic-data experiment did not provide a meaningful basis for distinguishing the methods: the BO methods recovered the parameters accurately, while the LS comparison was dominated by the parameter-correlation failure. For this reason, we chose to evaluate the methods using the actual experimental XRD and ND data, where the structure contains realistic experimental complexity and parameter correlations. This provides a more practically relevant benchmark for evaluating the proposed joint-refinement methodology.

AuthorsOpenReview ↗

Additional Response to Reviewer rbcc

The responses continue with Clarification 3 here:

Response to Clarification 3: The LS results reported in Tables 1 and 2 correspond to different experimental settings. We clarify the initialization, refinement procedure, constraints, objective functions, and the meaning of the 400-iteration value below. Distinction between the LS results in Tables 1 and 2: The LS results reported in Tables 1 and 2 correspond to two different experimental protocols. The Table 1 LS(PE) result represents the conventional crystallographic least-squares joint refinement starting from a nominal, well-chosen crystallographic parameter set. For the primary-extinction (PE) experiment, this refinement gives f_XRD = 28.50 and f_ND = 28.74. This is the result from the standard local crystallographic refinement. In contrast, the Table 2 LS result is part of the repeated-run experiment designed to evaluate the robustness of LS relative to CMT and the other optimization methods. In this experiment, LS is initialized from the same randomized parameter-space construction used for the CMT comparison. Consequently, different random seeds can converge to different local minima. The reported 102.35 ± 27.48 is therefore the mean ± standard deviation of the CMT evaluation objective across the repeated LS runs. It is not a second estimate of the Table 1 LS refinement and should not be interpreted as an average of R1, wR2, or GoF. Thus, the two tables answer different questions: • Table 1: How does conventional LS perform from a standard crystallographic starting model? • Table 2: How robust is local LS when initialized from the same broad parameter-space distribution used for the global optimization methods? Difference between the Table 1 LS result and the Table 2 LS mean is therefore a consequence of the different experimental designs rather than an inconsistency in the LS implementation. Table 1 uses a conventional crystallographic starting model and reports the resulting local LS solution. Table 2 deliberately perturbs the starting point across the parameter space and evaluates the resulting local LS solutions using the common CMT objective. The relatively large standard deviation in Table 2 therefore reflects the sensitivity of local LS to initialization and the presence of multiple local solutions in the refinement landscape. This behavior is consistent with the motivation for introducing CMT as a global, bounded multi-objective optimization framework. For the primary-extinction experiment, conventional LS produces f_XRD = 28.50 and f_ND = 28.74, whereas the CMT solutions achieve substantially lower values for both objectives. In the repeated-run experiment, the CMT methods likewise exhibit substantially lower mean objective values than LS How the LS starting points are determined: For the repeated-run experiments, each LS run is initialized independently using the same parameter-space construction used for the CMT/MOBO search. Each seed generates a random starting vector within the prescribed search region. The starting points therefore differ across seeds and are not repeated evaluations from the nominal parameter vector used for the Table 1 LS refinement. The relevant search region is defined by the parameter bounds used in the CMT formulation. These bounds constrain the physically meaningful ranges of the structural parameters, including the atomic occupancies, displacement parameters, and O1 fractional coordinate. The purpose of this randomized initialization is to evaluate the sensitivity of the local LS solver to starting conditions. Because LS is a local optimization method, different starting points can converge to different local solutions. Objective function minimized by LS versus the objective used for comparison: During LS refinement, GSAS-II minimizes its conventional least-squares residual using the experimental measurement uncertainties (σ) as weights. The joint refinement uses the XRD and ND histograms with their corresponding histogram weighting factors. Thus, the LS optimization itself is based on the conventional σ-weighted crystallographic residual. The CMT framework, however, uses the two diffraction objectives defined in Eq. (1) of the manuscript. These objectives are based on the discrepancy between the observed and calculated diffraction intensities, with the XRD and ND objectives scaled to place the two modalities on comparable numerical ranges. Comparing between LS and BO methods, the final solutions produced by LS runs are evaluated using the same objectives used to evaluate CMT and the other MOBO methods. In other words, we took the refined parameters from LS in each case, evaluate the diffraction intensity values for those set of parameters, then calculate the objectives for BO methods using Eqn. (1).

This response continues on the next page/comment

AuthorsOpenReview ↗

Continuation of Response to Reviewer rbcc (Page 3)

Response to Clarification 3 continues:

Therefore, the LS values reported in Table 2 are not the internal GSAS-II σ-weighted residuals. They represent the performance of the LS-derived solution under the common objective used for the method comparison. Constraints during LS refinement: The LS solver itself performs a conventional local least-squares refinement. The structural parameters are initialized within the defined search region, but the CMT parameter bounds are not imposed as a projection step during every LS iteration. Consequently, an unconstrained LS refinement can move outside the CMT/MOBO search region during optimization. For comparison, the final LS solution is subsequently mapped back into the common CMT search region before its CMT objective is evaluated. This ensures that LS and CMT are evaluated under the same parameter-domain definition and prevents an LS solution that leaves the CMT search region from receiving an advantage or penalty solely because it explored outside the domain available to CMT. Meaning of NIterations = 400: The value 400 represents the number of iterations for BO methods and the number of cycles for LS refinement. Even though it is practically not sufficient for LS to converge within 400 cycles, we included 400 for uniformity in comparison. Comparison using standard crystallographic metrics: We thank the reviewer for the suggestion. We agree that R1, wR2, and GoF are standard crystallographic metrics. However, our study compares all methods using the same XRD/ND objective functions defined in Eq. (1). Therefore, we retain these objectives as the primary evaluation criteria to ensure a consistent and fair comparison across LS and the BO methods. Although LS internally minimizes the conventional σ-weighted GSAS-II residual, its final solution is evaluated using the same objective function used for the BO methods.

Response to Clarification 4: We thank the reviewer for raising this important question regarding the physical interpretation of the occupancy parameters and the choice of optimization bounds. We agree that an occupancy of 0.800, corresponding to a nominal 20% deficiency on a crystallographic site, would represent a substantial defect concentration that cannot be justified without additional experimental evidence. The occupancy range of 0.8–1.2 was not intended to represent the expected physical defect concentration in Ho₂Ti₂O₇. Instead, we deliberately adopted a relatively broad parameter space to test the robustness of the optimization algorithms and examine their behavior under less restrictive bounds. In particular, although previous structural studies in the literature, predominantly based on powder diffraction, provide useful reference values, we chose not to impose these values as strict constraints in the present methodological benchmark. The same parameter bounds were applied to all compared algorithms to ensure a consistent comparison.

We recognize that a solution reaching the imposed occupancy boundary does not necessarily represent a physically reliable determination of site occupancy. Such behavior may instead reflect limited sensitivity of the diffraction data to individual parameters or correlations among refinement parameters. In particular, scale factors and occupancies can be correlated, and simultaneous XRD/ND refinement does not necessarily eliminate this ambiguity. Although we have not quantitatively evaluated these correlations in the present study, we agree that such an analysis would be valuable for establishing the physical reliability of the refined occupancies. We have revised the manuscript to clarify the rationale for the selected parameter bounds and explicitly distinguish optimization performance from the physical interpretation of the resulting structural parameters. We also acknowledged that occupancy values approaching the search boundaries should not be interpreted as quantitative evidence of defect concentrations without further validation. The present study primarily aims to develop and benchmark a joint XRD/ND optimization framework, while a detailed determination of site occupancies and their uncertainties would require additional crystallographic analysis.

Response to Clarification 5:

The crystals used for XRD and ND were cut from the same cylindrical rod grown by the TSFZ method, as reported in Ref. [Ghasemi 2018]. Crystals of different sizes were selected to accommodate the distinct sample-size requirements of X-ray and neutron diffraction experiments. Therefore, the discrepancy in the refined occupancies is unlikely to arise from differences in sample growth or composition. Instead, it may reflect differences in extinction effects and uncertainties in the integration of diffraction intensities used to extract the squared structure factors, (F_{hkl}^2), in the two experiments. We will also correct “FZ” to “TSFZ” in §3.1 to ensure consistency and avoid confusion.

AuthorsOpenReview ↗

Response to Reviewer rbcc Continues (Page 4)

Response to Clarification 6: We thank the reviewer for the suggestion. However, the statement of the problem is to provide joint refinement of the structure parameters based on both XRD and ND data, and not either one of those individually. It is evident to obtain a better/lower chisq value (between experimental and simulation) when we refine using either XRD or ND data (ignoring the other data), which is not the goal of the application. Here, we want to jointly refine the structure parameters for which identifying Pareto optimal solutions is necessary. Unlike single-objective optimization, multi-objective BO provides a set of Pareto-optimal solutions, allowing domain scientists to select appropriate trade-offs among competing objectives such as chisq calculated from XRD and ND data. Thus, comparing methods between single and multi-objective BO is not feasible and thus to maintain consistency, we have compared with different multi-objective BO methods.

Response to Clarification 7: We thank the reviewer for identifying this implementation detail. We confirm that gsas_nd_9d.py replaces |U| ≤ 10⁻⁶ with 0.01 only for the ND evaluation. This was introduced as a numerical safeguard to ensure that GSAS-II receives a nonzero Debye–Waller factoratomic displacement parameter at the U = 0 boundary. We note that the present XRD and ND datasets were collected from a single crystal at 100 K and 110 K, respectively, and the refinement uses integrated Bragg intensities. At these relatively low temperatures, the atomic displacement parameters may be weakly constrained by the available data, particularly due to correlations with other refinement parameters. Thus, near-zero optimized U values do not necessarily indicate physically vanishing atomic displacements. Importantly, this treatment is implemented in the common GSAS evaluation script and is therefore applied identically to all methods, including CMT, MORBO, the other BO methods, and the LS re-score. For any given parameter vector, all methods are evaluated by the same XRD/ND functions and therefore receive the same objective values. Consequently, the U-floor does not introduce a method-specific advantage under the adopted evaluation protocol, although we acknowledge that removing this substitution could affect the quantitative residuals and potentially the relative rankings. or alter the relative comparison between methods; changing the floor would change the common objective function rather than selectively change the ranking of one method against another. As with the parameter bounds adopted for other variables in this study, the U-floor is a user-defined implementation choice that can be readily adjusted based on the requirements of a particular application. Our primary focus is the development and evaluation of the optimization methodology, rather than prescribing a universal U-floor or parameter range. Users can therefore modify the bounds and numerical safeguards according to their material system, prior knowledge, and domain-specific requirements. We acknowledge that the CMT and MORBO implementations included 400 optimization iterations in addition to the initial samples, whereas the other methods used a total budget of 400 evaluations. As shown in Fig. 2, the best results for MOBO and ParEGO were obtained before 200 evaluations, and their performance does not improve with additional evaluations, hence this discrepancy is not expected to change the conclusion. Nevertheless, to ensure a uniform and fair comparison across all methods, we will exclude the final 21 evaluations from the CMT and MORBO results and update the reported results using a common budget of 400 evaluations for all methods.

Reviewer_pSksOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose CrystalMO-TuRBO, which is a multi-objective BO framework for joint crystal structure refinement using XRD and ND data. The method employs a two-phase optimization strategy: the first phase uses multiple parallel TuRBO-1 optimizers with different scalarization weights to explore promising regions of the parameter space, while the second phase performs localized refinement to improve solution accuracy. The framework is evaluated on experimental diffraction data from a single Ho₂Ti₂O₇ crystal, and the authors compare the results against conventional least-squares refinement and existing multi-objective BO methods, including ParEGO, qEHVI, and MORBO. The experimental results showcase improved objective values and robust convergence, highlighting the potential of trust-region BO for high-precision joint crystal structure refinement.

Critical strengths
  1. Practical Application: The paper addresses an important real-world problem in crystal structure refinement by jointly optimizing XRD and ND data.

  2. Effective Two-Phase Optimization: The proposed framework combines global exploration using multiple trust regions with local refinement, improving optimization performance and convergence accuracy.

  3. Relatively Strong Experimental Performance: The proposed method achieves better overall refinement results than conventional least-squares optimization and several established multi-objective BO methods, including ParEGO, qEHVI, and MORBO.

  4. Evaluation on Real Experimental Data: The experiments use real diffraction measurements rather than purely synthetic data, demonstrating the practical applicability of the proposed approach.

  5. Robustness Analysis: The authors conduct repeated experiments with 12 random seeds and investigate the effects of different local refinement region sizes, providing additional evidence of the method's effectiveness.

Critical weaknesses
  1. Limited Methodological Novelty: The proposed framework mainly combines existing techniques, including TuRBO-1, weighted scalarization, and local refinement. The paper does not introduce a fundamentally new BO algorithm, surrogate model, or acquisition function.

  2. Limited Experimental Evaluation: The experiments are conducted on a single material system (Ho₂Ti₂O₇). Although multiple experimental configurations and random seeds are considered, the evaluation lacks diversity in benchmarks. Additional material systems or standard multi-objective optimization benchmarks would strengthen their claims.

  3. Insufficient Justification for Design Choices: Several important hyperparameters, such as the number of trust regions (they mainly use TurBO-1), scalarization weights, and local refinement region sizes, appear to be heuristically selected. The paper does not sufficiently justify these choices or investigate their sensitivity.

  4. Limited Ablation Studies: Although the authors compare Phase 1 with the complete two-phase framework, more detailed ablation studies are needed to isolate the contributions of different components, particularly the use of multiple scalarization weights and parallel trust regions.

  5. Unclear Advantages Over Existing Methods: MORBO already incorporates trust-region BO for multi-objective problems. The paper could better explain why the proposed framework is preferable to MORBO beyond its performance on the specific crystal refinement problem.

  6. Limited Evidence of Physical Parameter Accuracy: Lower diffraction fitting errors do not necessarily imply more accurate crystal structure parameters. Additional validation against independently established structural parameters would strengthen the claim of high-precision refinement.

Questions for authors
  1. The proposed framework combines existing techniques, including TuRBO-1, weighted scalarization, and local refinement. Could the authors clarify the key methodological contributions beyond integrating these established approaches? In particular, how does the proposed method fundamentally differ from MORBO?

  2. The experimental evaluation is limited to a single material system (Ho₂Ti₂O₇). Could the authors provide additional experiments on other material systems or standard multi-objective optimization benchmarks to demonstrate the generalizability of the proposed framework? This may help me better assess my judgement in the submission.

  3. The framework uses four trust regions with manually selected scalarization weights, followed by a local refinement stage. Could the authors provide more comprehensive ablation studies to justify these design choices? In particular, how sensitive is the performance to the number of trust regions, the scalarization weights, and the allocation of the evaluation budget between the two phases?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pSks

Question 1. The proposed framework combines existing techniques, including TuRBO-1, weighted scalarization, and local refinement. Could the authors clarify the key methodological contributions beyond integrating these established approaches? In particular, how does the proposed method fundamentally differ from MORBO?

Response to Question 1: TuRBO-1, weighted scalarization, and local Bayesian optimization are established techniques individually. The contribution of CrystalMO-TuRBO (CMT) is therefore not the invention of these individual components, but the development of a problem-specific two-phase multi-objective optimization architecture that combines them in a way tailored to the joint XRD–ND crystal-structure refinement problem. Relationship of the proposed CMT to MORBO and Phase 1: CMT Phase 1 shares several important ideas with MORBO, particularly the use of multiple trust-region searches to explore different trade-offs between competing objectives. However, the two implementations are not identical. CMT Phase 1 decomposes the XRD–ND problem into parallel scalarized TuRBO-1 states, each with its own Gaussian-process surrogate, adaptive trust region, and q-Expected Improvement acquisition function. Four fixed scalarizations, λ = {0.15, 0.45, 0.65, 0.85}, are used to simultaneously explore different XRD–ND preferences, while all evaluated solutions are retained in a common archive. Our MORBO implementation, in contrast, uses random scalarization together with Thompson sampling within its multi-region search and uses hypervolume-based criteria for trust-region adaptation. Thus, although CMT Phase 1 is conceptually related to MORBO, it should not be interpreted as simply reproducing MORBO. Primary methodological contribution: Phase 2. The principal methodological contribution of CMT is the explicit transition from multi-objective exploration to localized refinement. After Phase 1 has identified promising regions of the XRD–ND objective landscape, Phase 2 selects the current best region and constructs a localized, physically constrained trust region around it. A balanced scalarization (λ = 0.5) is then optimized using a GP surrogate and q-Expected Improvement within this restricted region. This second stage is not simply another application of parallel scalarized TuRBO. Its purpose is different: Phase 1 allocates evaluations to discovering competing XRD–ND trade-offs across the broader search space, whereas Phase 2 deliberately concentrates the remaining optimization effort on improving the most promising refinement region. The resulting architecture therefore separates global multi-objective exploration from targeted local refinement, which is particularly appropriate for expensive crystallographic forward models where each function evaluation requires evaluating both XRD and ND diffraction objectives. Importantly, the Pareto archive generated during Phase 1 and Phase 2 is retained. The local refinement stage does not replace the multi-objective formulation with a single-objective optimization problem; rather, it provides an additional exploitation stage for improving promising solutions while preserving the underlying XRD–ND trade-off information. Why this constitutes an advance beyond combining existing techniques: The methodological contribution is consequently the design of a complete optimization strategy around the characteristics of joint XRD–ND refinement: (i) parallel exploration of different modality preferences, (ii) preservation of the resulting Pareto trade-offs in a global archive, and (iii) a subsequent localized Bayesian refinement stage that concentrates expensive evaluations around promising structural solutions. This architecture addresses a practical limitation of simply applying a standard MOBO algorithm: identifying a promising Pareto region and accurately refining a crystallographic solution are related but distinct optimization tasks. Application-level contribution: The contribution is also not limited to algorithmic development. To our knowledge, all the multi-objective Bayesian optimization approaches evaluated in this work have not previously been applied to the specific problem of joint single-crystal XRD–ND crystal-structure refinement. This problem provides a demanding physical-science test case because the XRD and ND objectives arise from complementary scattering mechanisms and can favor different regions of the shared structural parameter space. Our experiments demonstrate that the MOBO approaches substantially improve upon the conventional least-squares baseline for this joint refinement problem.

AuthorsOpenReview ↗

Response to Reviewer pSks (Page 2)

Question 2. The experimental evaluation is limited to a single material system (Ho₂Ti₂O₇). Could the authors provide additional experiments on other material systems or standard multi-objective optimization benchmarks to demonstrate the generalizability of the proposed framework? This may help me better assess my judgement in the submission.

Response to Question 2: We thank the reviewer for raising this important point regarding the generalizability of the proposed framework. We agree that evaluation on additional material systems would provide further empirical evidence of its robustness. However, the experimental setting considered in this work has a practical constraint that makes such an evaluation substantially different from evaluating a conventional optimization algorithm on synthetic or standard benchmark functions. The proposed framework is designed specifically for joint crystal-structure refinement using complementary X-ray and neutron diffraction data. To construct a valid experimental instance of this problem, matched XRD and ND measurements must be obtained from the same crystal specimen so that both objectives correspond to the same underlying crystal structure. Obtaining such paired single-crystal diffraction datasets requires material synthesis, experimental characterization, and access to both X-ray and neutron diffraction facilities. As noted explicitly in the Limitations section of the manuscript, this process is experimentally expensive and time-consuming. We therefore deliberately used a single high-quality Ho₂Ti₂O₇ crystal to provide a controlled and internally consistent evaluation, rather than combining independently measured datasets from different specimens. Importantly, the limitation is not that the optimization formulation is inherently specific to Ho₂Ti₂O₇. The underlying optimization problem is defined by the characteristics of the refinement task: expensive forward-model evaluations, strongly correlated structural parameters, non-convex search landscapes, and competing objectives arising from complementary diffraction modalities. These characteristics are not unique to Ho₂Ti₂O₇. Consequently, applying CMT to another crystal system with corresponding XRD and ND datasets would primarily require adapting the crystallographic parameterization, physical constraints, and GSAS-II evaluation setup to that material system, while the underlying multi-objective optimization framework remains applicable. The manuscript already makes this distinction, stating that the proposed framework is expected to apply to other crystallographic refinement problems. Regarding the suggestion to evaluate CMT on standard synthetic multi-objective optimization benchmarks, we believe such an experiment would not directly address the principal scientific question of this work. CMT was designed around the specific characteristics of joint XRD–ND crystal-structure refinement rather than as a general-purpose optimizer intended to establish superior performance across arbitrary multi-objective functions. In particular, the motivation for CMT is the need to reconcile complementary experimental modalities without collapsing them into a manually weighted scalar objective, while simultaneously dealing with the expensive and highly correlated refinement landscape. The method therefore targets a specific scientific inverse-problem setting rather than claiming universal superiority over existing MOBO methods on arbitrary benchmark functions. Hence, our empirical claims are intentionally restricted to the joint XRD–ND refinement problem studied here, and the method is expected to generalize to any joint XRD-ND refinement task. Standard benchmarks could assess general-purpose optimizer performance, but they would not reproduce the physical structure, parameter correlations, modality-specific sensitivities, and expensive crystallographic forward evaluations that motivate the proposed framework. We have therefore retained the current experimental scope and made this limitation explicit in the manuscript. We have also identified evaluation on additional crystal systems, material conditions, and related scientific inverse problems as an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer pSks (Page 3)

Question 3. The framework uses four trust regions with manually selected scalarization weights, followed by a local refinement stage. Could the authors provide more comprehensive ablation studies to justify these design choices? In particular, how sensitive is the performance to the number of trust regions, the scalarization weights, and the allocation of the evaluation budget between the two phases?

Response to Question 3: We thank the reviewer for this constructive suggestion. We agree that a more extensive ablation study varying the number of trust regions, scalarization weights, and Phase 1/Phase 2 budget allocation would provide additional insight into the sensitivity of these design choices. We considered these factors carefully in the design of CMT; however, a comprehensive factorial ablation is not feasible within the limited rebuttal period because each configuration requires multiple independent runs to obtain reliable performance estimates. For the scalarization weights, the four values were intentionally selected to provide approximately uniform coverage of the scalarization space. The objective was not to manually tune weights toward a particular solution, but rather to distribute the Phase 1 trust regions across different XRD–ND trade-offs. This allows the optimization to explore regions that place greater emphasis on XRD, greater emphasis on ND, and intermediate compromises. The use of multiple scalarizations is therefore a deliberate mechanism for covering different representations of the multi-objective problem rather than a set of hyperparameters tuned to maximize performance on the reported data. For the number of trust regions, four regions were selected as a practical balance between parallel exploration and computational cost. Phase 1 is intended to identify promising regions of the refinement landscape rather than completely solve the refinement problem. Once promising regions have been identified, the proposed Phase 2 is specifically designed to exploit the best region through localized refinement. Thus, increasing the number of Phase 1 trust regions indefinitely would increase exploration cost without necessarily providing a corresponding benefit, particularly when the subsequent local refinement stage is responsible for further improving the solution. The Phase 1/Phase 2 budget allocation follows the same design principle. Phase 1 is allocated sufficient evaluations to discover promising regions, while Phase 2 concentrates evaluations around the most promising solution identified during exploration. This separation reflects the intended global-exploration-to-local-exploitation structure of CMT rather than treating the entire evaluation budget as homogeneous iterations of the same optimizer. The convergence results presented in the manuscript further show that the local refinement stage continues to improve the objective values after the initial exploration stage. We would have welcomed the opportunity to perform a substantially larger ablation study across alternative numbers of trust regions, scalarization schedules, and budget allocations. However, because the reported results themselves are based on repeated independent runs for each method and configuration, conducting a sufficiently rigorous set of additional ablations would require a considerable number of additional expensive optimization runs. We therefore believe that presenting partial or single-run ablations within the rebuttal period would provide less reliable evidence than the repeated-run evaluation already reported in the manuscript.

Reviewer_SiugOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a model called CrystalMO-TuRBO for joint crystal structure refinement from X-ray and neutron diffraction measurements. It treats the two modality-specific discrepancies as separate objectives and combines established Bayesian optimization components in a two-phase procedure. Evaluated objective pairs are retained for Pareto analysis and selection of a single compromise solution.

Critical strengths
  1. The motivation of this work is good. Joint X-ray and neutron refinement is a scientifically motivated setting in which a common structural parameter vector must reconcile measurements with different sensitivities. This gives the work practical relevance beyond a purely synthetic optimization demonstration.

  2. The overall writing is logical and clear in general. Each section is closely connected with each other.

Critical weaknesses
  1. The contribution of this work is unclear. The paper does not yet establish a sufficiently clear methodological advance or transferable insight beyond combining existing BO components. The main methodological addition appears to be this particular exploration-to-refinement schedule. The paper needs to establish what this contributes beyond scalarized trust-region BO followed by local polishing.

  2. The work is more aligned with a journal paper in chemistry, not a top-tier AI conference paper. The method of this work is straightforward without many new insights on the BO. The comparison with the closest published baseline is also not adequately supported. Please refer to the questions below.

Questions for authors
  1. Does Appendix G accurately describe the implementation used for the reported MORBO results? Please explain the deviations from local surrogate modeling and hypervolume-based Thompson selection in the original algorithm.

  2. What is the specific algorithmic advance beyond parallel scalarized TuRBO? It is quite unclear regarding the contribution of this work from the perspective of new AI algorithms or models.

  3. Is the primary goal to approximate the full Pareto frontier or to obtain one preferred compromise? Section 3.5 describes normalized utopia distance as equivalent to a normalized objective sum, while Appendix F.5 specifies normalized Euclidean distance; these criteria are not generally equivalent.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Siug

Thank you for your detailed feedback and questions. Your questions and our responses are as follow"

Question 1. Does Appendix G accurately describe the implementation used for the reported MORBO results? Please explain the deviations from local surrogate modeling and hypervolume-based Thompson selection in the original algorithm.

Response to Question 1: We thank the reviewer for raising this clarification. To the best of our knowledge, MORBO, MOBO, and ParEGO have not previously been applied to the specific problem of joint XRD–ND crystal-structure refinement considered in this work. Therefore, the implementations used in our study were developed specifically for this problem by adapting the core principles of these established methods to the characteristics of crystallographic refinement. The MORBO implementation described in Appendix G and implemented in the anonymized code accompanying this submission follows the central MORBO strategy of conducting multi-objective Bayesian optimization through multiple trust regions operating in parallel. We retain Gaussian-process surrogate modeling, Thompson sampling, random scalarization of the objectives, and hypervolume-based assessment of improvement for trust-region adaptation. In particular, drawing scalarization weights from the simplex allows the search to explore different preferences between the XRD and ND objectives rather than requiring a single weighting to be specified manually before optimization. This is well suited to joint refinement because the relative importance of XRD and ND may depend on the scientific objective or experimental context. Problem-specific adaptations: For our implementation, we use a shared ModelListGP trained on the accumulated observations. The ModelListGP contains independent single-task Gaussian processes for the XRD and ND objectives, with all evaluated points contributing to the surrogate models. This choice reflects the structure of our problem: XRD and ND are evaluated simultaneously at the same parameter vectors and describe the same underlying crystal structure. We therefore allow the accumulated observations to inform the surrogate models across the trust regions. In addition, candidate selection in our implementation is based on the randomly scalarized Thompson-sampled objective. Specifically, a random weight vector is drawn from the simplex, a joint Thompson-sampled objective vector is generated over the candidate pool, and candidates are ranked according to the resulting scalarized score. Hypervolume is used separately to assess whether the newly evaluated point improves the current nondominated set and to guide trust-region adaptation, including expansion, contraction, and re-centering. Thus, the implementation preserves the core MORBO concepts, parallel trust-region search, Gaussian-process modeling, Thompson sampling, random scalarization, and hypervolume-guided trust-region adaptation, while adapting specific components to the joint XRD–ND crystallographic refinement setting.

The responses continue on the next page

AuthorsOpenReview ↗

Responses to Reviewer Siug (Page 2)

Question 2. What is the specific algorithmic advance beyond parallel scalarized TuRBO? It is quite unclear regarding the contribution of this work from the perspective of new AI algorithms or models.

Response to Question 2: The central algorithmic contribution of CrystalMO-TuRBO (CMT) is not simply the use of multiple scalarized trust regions in parallel. Rather, it is the two-phase optimization architecture, in which broad multi-objective exploration is explicitly followed by a problem-specific local refinement stage designed for high-accuracy crystallographic refinement. Phase 1 adapts TuRBO to the joint XRD–ND refinement problem by maintaining parallel trust-region searches under different XRD–ND scalarizations. This allows different trade-offs between the two diffraction objectives to be explored simultaneously while retaining both objective values for subsequent Pareto analysis. In this respect, Phase 1 is intentionally related to existing scalarization-based and multi-region MOBO strategies, including MORBO, rather than being presented as an entirely independent BO primitive. There are, however, implementation differences. CMT Phase 1 uses fixed parallel scalarizations, with an independent TuRBO-1 state and GP for each scalarization, whereas our MORBO implementation uses random scalarization during Thompson-sampling-based candidate selection and a shared ModelListGP. Thus, Phase 1 should be viewed as a problem-specific multi-objective extension of TuRBO that shares several ideas with MORBO, rather than as a reimplementation of MORBO. The contribution of this work is the design of Phase 2 algorithm and integration with Phase 1 algorithm : Once Phase 1 has identified promising regions of the joint refinement landscape, CMT does not continue global multi-objective exploration indefinitely. Instead, it transitions to a dedicated local refinement stage. Phase 2 identifies the current best solution under a balanced XRD–ND scalarization, constructs a localized axis-aligned bounding region around that solution in the physical parameter space, and performs q-Expected Improvement within this restricted region. The optimization therefore changes from broad exploration of competing objective trade-offs to targeted local refinement of the most promising crystallographic solution. This explicit transition from multi-objective global exploration to localized scalarized refinement is the key design element that distinguishes CMT from simply running parallel scalarized TuRBO throughout the entire budget. This distinction is important for the underlying scientific problem. Crystal-structure refinement is not only a search for a Pareto set; the practical goal is to obtain a highly accurate parameterization of the crystal structure after promising regions have been identified. Phase 2 exploits this property by allocating the latter portion of the optimization budget to fine-grained search in a physically constrained neighborhood rather than continuing to spend evaluations exploring distant regions of the parameter space. The paper's ablation results further support the role of this second stage, showing that Phase 2 continues to improve the refinement objectives after the global exploration stage has converged. Novelty beyond algorithm development: We also emphasize that the contribution is not limited to the optimization architecture itself. To our knowledge, the multi-objective Bayesian optimization approaches considered in this study have not previously been applied to the specific problem of joint single-crystal XRD–ND crystal-structure refinement. This is a particularly challenging scientific inverse problem because the two modalities provide complementary information, have different sensitivities to the underlying structural parameters, and can produce competing refinement objectives. Conventional joint refinement addresses this competition through manually specified scalarization weights, whereas the CMT framework preserves the two objectives explicitly during optimization. Our experiments therefore provide a new application of multi-objective Bayesian optimization to an expensive, physics-based materials characterization problem rather than merely evaluating another optimization algorithm on a standard synthetic benchmark. Importantly, the experimental results show that the MOBO approaches substantially outperform the conventional least-squares joint-refinement baseline on this problem, demonstrating the practical value of applying these methods to crystallographic refinement.

Please continue on Page 3...

AuthorsOpenReview ↗

Response to Reviewer Siug (Page 3)

Question 3. Is the primary goal to approximate the full Pareto frontier or to obtain one preferred compromise? Section 3.5 describes normalized utopia distance as equivalent to a normalized objective sum, while Appendix F.5 specifies normalized Euclidean distance; these criteria are not generally equivalent.

Response to Question 3: The primary goal of the multi-objective optimization is to identify a set of non-dominated solutions representing the trade-off between XRD and ND refinement quality, rather than to impose a single preferred compromise during optimization. As described in Appendix F.5, each Pareto-optimal solution represents a different balance between the two complementary diffraction objectives. The complete Pareto set is therefore retained as the relevant output of the multi-objective optimization. Section 3.5 addresses a separate practical requirement of our experimental comparison: because the least-squares baseline produce a single reported parameter vector, we needed a fixed and reproducible rule for selecting one representative solution from the Pareto set so that the different algorithms could be compared using a common single-solution criterion. The criterion described in Section 3.5 is therefore a reporting/comparison convention and should not be interpreted as defining the objective of the multi-objective optimization itself. Appendix F.5 provides the more general procedure for selecting a compromise solution from an empirical Pareto set. Specifically, it first constructs the non-dominated set and defines the ideal (utopia) point from the best observed values of the individual objectives. It then normalizes the objective ranges and uses the normalized Euclidean distance to the ideal point to identify a representative compromise solution. We also acknowledge the reviewer's observation that a normalized objective sum and a normalized Euclidean distance are not generally mathematically equivalent. Our statement in Section 3.5 that the two criteria were equivalent was intended to describe the behavior observed in our experimental comparison, rather than to claim a general mathematical equivalence. We will revise this wording to remove that ambiguity and explicitly distinguish the fixed reporting criterion from the general Pareto-selection procedure. For an actual materials-science application, we do not view the single compromise solution as the only scientifically meaningful output. The Pareto-optimal solutions constitute a set of candidate refinements, each representing a different balance between XRD and ND agreement. A materials scientist can therefore inspect the Pareto set and select the solution that is most appropriate for the scientific question, for example, by prioritizing agreement with one modality, emphasizing a particular structural or magnetic parameter, or selecting a solution that best satisfies the physical interpretation of the experiment. The compromise solution used in our tables is primarily a reproducible way of reducing this set to one representative solution for algorithm-to-algorithm comparison. We will revise Section 3.5 to make this three-level distinction explicit and remove the wording suggesting that the normalized objective sum and normalized Euclidean distance are generally equivalent.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.