REVIEW 4 major objections 4 minor 34 references
Active Learning Guided Design Space Refinement for Scalable Multi-Objective Bayesian Optimization in Materials Discovery
T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read An active-learning pre-filter halves a materials search space while keeping over 99% of the useful trade-offs, and warm-started Bayesian optimization on the reduced space finds Pareto-optimal candidates faster under the same budget.
desk verdict The core idea is plausible and transparently presented, but the headline improvement is confounded by unequal initialization, and the paper needs a random-pruning matched-budget baseline to support its claims. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the refinement-to-warm-start loop: (1) a Pareto-aware proxy score $s(x) = 0.35s_{\text{mean}}(x) + 0.25s_{\text{worst}}(x) + 0.25s_{\text{ideal}}(x) + 0.15s_{\text{geo}}(x)$ that scalarizes the multi-objective landscape into one number; (2) a 90th-percentile threshold $\tau = Q_{90}(\{s(x)\})$ that turns that score into binary promising/non-promising labels; (3) a single XGBoost classifier producing positive-class probability $p(x)$ for every candidate; (4) the density-aware greedy sampling (DAGS) acquisition rule, which picks the next candidate to label using model-boundary uncertainty, feature-space novelty, and local-density information; and (5) the retention union $X_{k+1} = X^{\text{pred}} \cup X^{\text{model}} \cup X^{\text{unc}} \cup X^{\text{conf}} \cup X^{\text{query}}$ that defines the reduced space, with the uncertainty subcategory capped at the 65th percentile so that boundary regions do not swallow the reduction. The same classifier probability drives pseudo-labeling at threshold $\lambda_h = 0.90$, but pseudo-labeled candidates count only as classifier training data, never as oracle observations or warm-start points. The claim is carried by these parts working together: the proxy defines what 'promising' means, the classifier propagates that definition across the pool, DAGS decides which labels are worth paying for, and the retention union guards against discarding Pareto-relevant candidates.
What would settle it
Run the refinement at the same 45–50% reduction rate with candidates pruned uniformly at random, and compare hypervolume retention: if random pruning also keeps more than 99% of the hypervolume, the classifier is not carrying the result. As a second check, give full-space BO the same total oracle budget as the warm-start pipeline — roughly 170 evaluations once the refinement stage's 19–21 queries are counted — and ask whether its Pareto-AUC still trails the reduced-space curve.
Extended reading notes
Core claim
The paper's central claim is that multi-objective optimization of large discrete materials spaces can be accelerated by an active-learning refinement stage that precedes Bayesian optimization. For each objective the method computes a normalized benefit value, combines four components — mean performance, worst-objective, distance to the ideal point, and geometric mean — into a single proxy $s(x)$ with fixed weights $0.35, 0.25, 0.25, 0.15$, and treats candidates below the 90th percentile of $s(x)$ as non-promising. A single XGBoost classifier, trained on a balanced set of 16 labeled candidates and expanded by density-aware greedy sampling queries, assigns every candidate a positive-class probability $p(x)$; the reduced space retains the union of predicted-promising, high-confidence, uncertainty-retained (above the 65th percentile), high-confidence pseudo-labeled, and oracle-queried candidates, and the queried samples are transferred to the BO stage as warm-start observations. Reported results: the pressure-vessel space shrinks 49.6% with 99.6% hypervolume retention, the COF space shrinks 44.2% with 99.5% retention, and Pareto discovery area-under-the-curve rises from 1562 to 2123 and from 2547 to 3199 respectively, with near-plateau hypervolume reached in roughly 20–30 BO iterations versus 40–50 for full-space BO. The paper explicitly frames the benefit as improved BO-stage efficiency rather than fewer total oracle evaluations, because the refinement stage consumes its own oracle queries.
Load-bearing premise
The refinement's value rests on the hand-picked weighted score and its 90th-percentile cutoff correctly telling promising materials from discardable ones, and on that cutoff being computable at all, which requires objective values for the whole candidate pool up front; the paper does not validate the score against random pruning or alternative scalarizations.
Editorial extensions
If this is right
- For any large precomputed materials screen, the refinement stage can halve the candidate pool before expensive evaluation begins, with hypervolume retention above 99% as the reported guardrail.
- Multi-objective BO on the reduced space reaches near-plateau hypervolume in roughly half the iterations — about 20–30 instead of 40–50 on the pressure-vessel benchmark — so short-budget runs get usable Pareto fronts sooner.
- Cumulative Pareto-front discovery improves throughout the run, not only at convergence: Pareto-AUC rises from 1562 to 2123 on the pressure-vessel benchmark and from 2547 to 3199 on the COF benchmark.
- Because the recipe is benchmark-agnostic — one three-objective engineering problem and one two-objective adsorption problem — the same refinement-plus-warm-start pipeline transfers to other discrete materials spaces that come with descriptors and a small labeled seed.
- The convergence gain is a BO-stage effect: the warm-start configuration consumes roughly 19–21 oracle evaluations during refinement, so the comparison is 150 BO evaluations with warm-start reuse rather than an equal total-oracle-evaluation comparison.
Reading between the lines
- My reading: the headline retention numbers are not yet evidence that the classifier does the work, since the paper never compares the proxy filter against random pruning at the same reduction rate; a testable open question is whether a random half of these datasets would also retain roughly 99% hypervolume because of near-duplicate candidates.
- My reading: the pipeline as described is offline — the percentile cutoff in Eq. (20) needs objective values for the whole pool — so a live discovery loop would need an online stand-in for the threshold, a step the paper explicitly leaves to future work.
- My reading: a total-budget comparison (150 full-space evaluations versus roughly 19–21 refinement queries plus 150 reduced-space evaluations) would settle whether the Pareto-AUC advantage reflects search focus or simply more evaluated points; the paper reports the extra queries but does not renormalize the budget.
- My reading: the most defensible deployment is offline pool curation, where cheap classifier labels can remove half of the enumerated candidates before expensive simulations run, making the framework's real economy the avoided simulations rather than the BO-stage speedup.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an active-learning-guided search-space refinement framework for multi-objective Bayesian optimization in materials discovery. The DAGS active-learning method is used to iteratively train an XGBoost classifier on a proxy score derived from a hand-weighted scalarization of the objectives, then to prune roughly half of the candidate space while retaining candidates deemed promising, uncertain, or pseudo-labeled. The reduced space is passed to a qNEHVI Bayesian optimizer, warm-started with the oracle-evaluated samples collected during refinement. The authors evaluate the method on a CFRP pressure-vessel design benchmark (52,272 candidates, three objectives) and a COF methane-adsorption benchmark (69,839 candidates, two objectives), reporting that the refinement preserves more than 99% of the original hypervolume, accelerates early BO convergence, and increases cumulative Pareto discovery (Pareto-AUC from 1562 to 2123 and from 2547 to 3199).
Significance. If the claimed improvements were attributable to the DAGS-guided refinement itself, the framework would be a useful contribution to large-scale multi-objective materials optimization, where candidate pools are large and evaluations are expensive. The paper is clearly written, the two benchmarks are complementary, and the authors are transparent about several limitations, including the extra oracle samples used by the warm start and the absence of formal significance claims. The code and data availability statements are also a strength. However, the central empirical claim is currently confounded by unequal initialization between the compared BO configurations, and the hypervolume-retention metric lacks a random-pruning control; these issues must be addressed before the headline conclusions can be accepted.
major comments (4)
- [§2.6, Eqs. (32)-(33); §3.2; §3.3] The central empirical comparison is confounded by unequal initialization. Equation (32) sets D_0^ref = D_DAGS, so the DAGS-guided BO configuration warm-starts with approximately 19 oracle-evaluated samples on the pressure-vessel benchmark and approximately 21 on the COF benchmark, whereas Full BO starts with 5 samples; both then run 150 new BO evaluations. The DAGS configuration therefore receives roughly 14-16 additional oracle labels before optimization begins. Any warm-started optimizer would be expected to show faster early hypervolume convergence and higher cumulative Pareto discovery for this reason alone. The reported gains (Pareto-AUC 1562→2123 and 2547→3199 in Table 2 and Figures 3-5) cannot be attributed to AL-guided refinement without additional controls. The authors should add an equal-total-budget Full BO baseline (e.g., Full BO with 155-171 total evaluations, or with 19/21 random initial points) and a reduced-space baseline with random pruning at the same retention rate and matched warm-start size. Without these, the headline claim that the reduced-space strategy improves convergence and Pareto discovery is unsupported.
- [§2.4, Eqs. (19)-(21); §3.1, Table 2] The hypervolume-retention result is not probative without a random-pruning control. The paper reports that removing roughly half of each candidate space preserves 99.6% and 99.5% of the original hypervolume, and interprets this as evidence that discarded solutions were predominantly low-quality. But the reduced space is constructed by retaining candidates with high proxy score, high classifier confidence, high uncertainty, and pseudo-labels, so a high retention rate is partly by construction: the proxy score directly ranks candidates by a scalarization of the objectives, and the retained subsets are enriched for high proxy values. A random subset of equal size would provide the natural baseline; if random pruning also preserves most of the hypervolume, the claim in §3.1 that the refinement removes 'objective-space redundancy' is not established. The authors should report retention metrics for random pruning at the same reduction ratio, and ideally also for an alternative scalarization or a non-scalarized retention rule.
- [§3.3, Figure 5; §3, paragraph on statistics] The Pareto-AUC improvements are reported as point means without error bars or significance tests, despite the paper's own statement that no formal significance claims are made unless a paired test is reported. The increases from 1562 to 2123 and from 2547 to 3199 are presented as 'confirming' improved Pareto discovery, but with only ten seeds and no per-run distributions, these differences could be within the run-to-run variability, especially given the initialization confound in the first major comment. The authors should report per-run Pareto-AUC values and paired tests (e.g., Wilcoxon signed-rank) comparing DAGS-BO against the controlled baselines, and should rephrase the conclusions in §3.3 and §4 to match the resulting evidence.
- [§2.4, paragraph beginning 'Both the global percentile threshold...'] The method as evaluated relies on offline access to objective values for the full candidate set. The paper explicitly acknowledges that the global 90th-percentile threshold in Eq. (20) and the balanced initial labeled set are feasible only because the benchmarks are offline and the objectives are known for all candidates, and that a genuinely online application would require an online approximation. This is a load-bearing limitation for the stated goal of 'autonomous materials discovery settings': the refinement stage, as implemented, requires exactly the expensive oracle evaluations that the framework is intended to save. The authors should either develop and evaluate an online variant with an adaptive threshold and an initial set obtained without full-space oracle access, or reframe the contribution as offline search-space pruning rather than autonomous active-learning-guided refinement.
minor comments (4)
- [Figure 5] The caption states that error bars indicate one standard deviation, but the bars in Figure 5 appear to lack visible error bars; please verify the figure or adjust the caption.
- [Table 2] The 'Pareto-AUC Imp.' column reports values as '1562→2123' and '2547→3199' without units or a description of what the arrows denote; please clarify that these are mean values and define the AUC unit.
- [Eq. (36)] The top-performing candidate retention metric T_ret is reported as 97.3% and 74.4% in §3.1, but the number of 'top-ranked candidates' used in the definition is never specified; please state the top-k count.
- [References] Reference [33] contains a formatting error in the author list ('Wei, Y., Zhuang, V. Soedarmadji, S. & Sui, Y.'); a missing '&' or comma between authors.
Circularity Check
Hypervolume-retention claim is partly built into the proxy-based label construction; BO gain is an unequal-budget confound rather than a definitional tautology.
-
self definitional
[Section 2.4 Eqs. (19)-(21) and Section 3.1 Eq. (34)]
"The final proxy score is s(x) = 0.35smean(x) + 0.25sworst(x) + 0.25sideal(x) + 0.15sgeo(x). (19) ... A percentile threshold is then computed as τ = Q_p({s(x) : x ∈ X0}), (20) ... Binary labels are assigned as y(x) = 1, s(x) ≥ τ; 0, s(x) < τ. (21) ... HVret = HV(Xk)/HV(X0), (34)"
The 'promising' class that determines which candidates are retained is defined by thresholding a weighted scalarization of the same normalized objectives z_j(x) used in the HV evaluation. Each term in Eq. (19) is nondecreasing in each objective, so the 90th-percentile rule (Eqs. 20-21) preferentially selects candidates with high objective values; the reported >99% HV retention is therefore substantially built into the selection rule rather than being an independent test of whether AL located Pareto-relevant structure. The paper's statement that 'true Pareto dominance is used exclusively for evaluation' does not remove the circularity, because the proxy itself is a function of the objective values entering HV.
full rationale
The partial circularity is confined to the retention result. The reduced space is produced by a classifier whose training labels are obtained by thresholding a hand-crafted scalarization of the same objective values that define the hypervolume metric; hence >99% HV retention is partly a property of the labeling construction rather than an independent finding. The paper concedes at Section 2.4 that 'Both the global percentile threshold and balanced initialization are feasible here because the experiments use offline benchmark datasets with objective values already available for all candidates,' confirming that the labeler sees the evaluation targets. The BO-stage comparison is a real confound, because Eq. (32) sets D_0^ref = D_DAGS, giving the warm-start roughly 19-21 oracle samples versus 5 for Full BO; the Discussion explicitly frames the gain as 'BO-stage efficiency' rather than total-oracle reduction. This is an unequal-budget comparison, not a definitional circularity, so it does not by itself raise the circularity score. The citation to the authors' DAGS [31] is attribution of a component rather than an imported uniqueness theorem or ansatz, and is therefore not load-bearing circularity. On balance, one central retention claim is partly circular, while the rest of the pipeline is not equivalent to its inputs.
Assumptions & free parameters
free parameters (5)
- Proxy score weights =
0.35, 0.25, 0.25, 0.15
- Percentile threshold p =
90
- Pseudo-label confidence threshold lambda_h =
0.90
- Retention percentiles =
50 / 65
- Initial labeled set size and DAGS iterations =
16 points; 3 and 5 iterations
assumptions (4)
- domain assumption The proxy score s(x) (Eq. 19) is a valid surrogate for Pareto-optimality in the objective space.
- domain assumption The XGBoost classifier trained on proxy-derived labels generalizes across the full candidate space.
- domain assumption The five retention categories in Eq. (25) (classifier-positive, high confidence, uncertain, pseudo-labeled, queried) collectively preserve the Pareto-relevant manifold.
- domain assumption All objective values are available offline for thresholding and oracle queries (Section 2.4).
Cite this review
Pith. "Pith review of Active Learning Guided Design Space Refinement for Scalable Multi-Objective Bayesian Optimization in Materials Discovery." pith.science (2026). https://pith.science/paper/NWKKIMJS
@misc{pith2026260804651,
author = {Pith},
title = {Pith review of: Active Learning Guided Design Space Refinement for Scalable Multi-Objective Bayesian Optimization in Materials Discovery},
year = {2026},
howpublished = {\url{https://pith.science/paper/NWKKIMJS}},
note = {Machine review of arXiv:2608.04651}
}
read the original abstract
Advanced materials discovery increasingly relies on machine learning and Bayesian optimization to explore large discrete design spaces under limited evaluation budgets. However, conventional Bayesian optimization (BO) can become inefficient as candidate spaces grow, often evaluating low-value regions before reaching informative areas. We propose an active-learning (AL)-guided adaptive search-space refinement framework combined with multi-objective BO to accelerate materials optimization while preserving Pareto-relevant regions. We evaluate the approach on CH4/N2 separation in covalent-organic frameworks and pressure-vessel design with material-direction stress components and thickness objectives. Results show that the AL-guided refinement reduces the candidate space by approximately half while preserving more than 99 percent of the original hypervolume. The reduced-space strategy improves early convergence and cumulative Pareto-front discovery from the BO, demonstrating efficient large-scale materials optimization across constrained autonomous materials discovery settings.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
Frazier, P. I. (2018). A tutorial on Bayesian optimization. arXiv preprint arXiv:1807.02811
arXiv 2018
-
[2]
Park, K., Song, C., Park, J., & Ryu, S. (2023). Multi-objective Bayesian optimization for the design of nacre-inspired composites: optimizing and understanding biomimetics through AI. Materials Horizons, 10(10), 4329–4343
work page 2023
-
[3]
Di Fiore, F., Nardelli, M., & Mainini, L. (2023). Active learning and Bayesian optimization: A unified perspective to learn with a goal. arXiv preprint arXiv:2303.01560
arXiv 2023
-
[4]
Hvarfner, C., Hellsten, E., Hutter, F., & Nardi, L. (2023). Self-correcting Bayesian optimization through Bayesian active learning.Advances in Neural Information Processing Systems, 36, 79173–79199
work page 2023
-
[5]
Snoek, J., Larochelle, H., & Adams, R. P. (2012). Practical Bayesian optimization of machine learning algorithms.Advances in Neural Information Processing Systems, 25
work page 2012
-
[6]
D., Allaire, D., & Arr´ oyave, R
Khatamsaz, D., Vela, B., Singh, P., Johnson, D. D., Allaire, D., & Arr´ oyave, R. (2023). Bayesian optimization with active learning of design constraints using an entropy-based approach.npj Computational Materials, 9(1), 49. 17
work page 2023
-
[7]
D., Allaire, D., & Arr´ oyave, R
Khatamsaz, D., Vela, B., Singh, P., Johnson, D. D., Allaire, D., & Arr´ oyave, R. (2022). Multi- objective materials Bayesian optimization with active learning of design constraints: Design of ductile refractory multi-principal-element alloys.Acta Materialia, 236, 118133
work page 2022
-
[8]
Ozaki, R., Ishikawa, K., Kanzaki, Y., Takeno, S., Takeuchi, I., & Karasuyama, M. (2024). Multi-objective Bayesian optimization with active preference learning. InProceedings of the AAAI Conference on Artificial Intelligence(Vol. 38, No. 13, pp. 14490–14498)
work page 2024
Show all 34 references
-
[9]
A., & Garrido-Merch´ an, E
Mart ´ ın, L. A., & Garrido-Merch´ an, E. C. (2021). Many objective Bayesian optimization. arXiv preprint arXiv:2107.04126
2021 arXiv
-
[10]
W., & Chen, W
Zhang, Y., Apley, D. W., & Chen, W. (2020). Bayesian optimization for materials design with mixed quantitative and qualitative variables.Scientific Reports, 10(1), 4924
2020
-
[11]
Zuo, Y., Qin, M., Chen, C., Ye, W., Li, X., Luo, J., & Ong, S. P. (2021). Accelerating materials discovery with Bayesian optimization and graph deep learning.Materials Today, 51, 126–135
2021
-
[12]
Kotthoff, L., Wahab, H., & Johnson, P. (2021). Bayesian optimization in materials science: a survey. arXiv preprint arXiv:2108.00002
2021 arXiv
-
[13]
P., Benin, A
Cˆ ot´ e, A. P., Benin, A. I., Ockwig, N. W., O’Keeffe, M., Matzger, A. J., & Yaghi, O. M. (2005). Porous, crystalline, covalent organic frameworks.Science, 310(5751), 1166–1170
2005
-
[14]
R., Ramdas, A., Wu, Y., Rohr, B., Ermon, S., Dionne, J., et al
Chitturi, S. R., Ramdas, A., Wu, Y., Rohr, B., Ermon, S., Dionne, J., et al. (2024). Targeted materials discovery using Bayesian algorithm execution.npj Computational Materials, 10(1), 156
2024
-
[15]
E., Ren, Z., Tiihonen, A., Liu, Z., Sun, S., et al
Liang, Q., Gongora, A. E., Ren, Z., Tiihonen, A., Liu, Z., Sun, S., et al. (2021). Benchmarking the performance of Bayesian optimization across multiple experimental materials science domains.npj Computational Materials, 7(1), 188
2021
-
[16]
Krokidas, P., Gkatsis, V., Theocharis, J., & Giannakopoulos, G. (2025). Navigating materials design spaces with efficient Bayesian optimization: a case study in functionalized nanoporous materials.Digital Discovery, 4(12), 3753–3763
2025
-
[17]
Jin, Y., & Kumar, P. V. (2023). Bayesian optimisation for efficient material discovery: a mini review.Nanoscale, 15(26), 10975–10984
2023
-
[18]
J., Arr´ oyave, R., Qian, X., & Dougherty, E
Qian, X., Yoon, B. J., Arr´ oyave, R., Qian, X., & Dougherty, E. R. (2023). Knowledge-driven learning, optimization, and experimental design under uncertainty for materials discovery. Patterns, 4(11)
2023
-
[19]
Arr´ oyave, R., Khatamsaz, D., Vela, B., Couperthwaite, R., Molkeri, A., Singh, P., et al. (2022). A perspective on Bayesian methods applied to materials discovery and design.MRS Communications, 12(6), 1037–1049
2022
-
[20]
Q., Bhattacharya, A., Pati, D., Qian, X., Arroyave, R., & Mallick, B
Lei, B., Kirk, T. Q., Bhattacharya, A., Pati, D., Qian, X., Arroyave, R., & Mallick, B. K. (2021). Bayesian optimization with adaptive surrogate models for automated experimental design.npj Computational Materials, 7(1), 194
2021
-
[21]
Wang, Y., Iyer, A., Chen, W., & Rondinelli, J. M. (2020). Featureless adaptive optimization accelerates functional electronic materials design.Applied Physics Reviews, 7(4). 18
2020
-
[22]
Y., Sargent, E
Chen, J., Ou, P., Chang, Y., Zhang, H., Li, X. Y., Sargent, E. H., & Chen, W. (2025). Adaptive uncertainty-aware deep learning for materials discovery with high-dimensional design inputs. InInternational Design Engineering Technical Conferences and Computers and Information in...
2025
-
[23]
Y., Sargent, E
Chen, J., Ou, P., Chang, Y., Zhang, H., Li, X. Y., Sargent, E. H., & Chen, W. (2026). Materials discovery using uncertainty-aware constrained Bayesian optimization with representation learning of high-dimensional inputs.Journal of Mechanical Design, 148(2), 021707
2026
-
[24]
Chowdhury, C. (2024). Bayesian optimization for efficient prediction of gas uptake in nanoporous materials.ChemPhysChem, 25(16), e202300850
2024
-
[25]
K., Qian, X., Karaman, I., et al
Solomou, A., Zhao, G., Boluki, S., Joy, J. K., Qian, X., Karaman, I., et al. (2018). Multi-objective Bayesian materials discovery: Application on the discovery of precipitation strengthened NiTi shape memory alloys through micromechanical modeling.Materials & Design, 160, 810–827
2018
-
[26]
Wei, Q., Wang, Y., Yang, G., Li, T., Yu, S., Dong, Z., & Zhang, T. Y. (2025). Discovering novel lead-free solder alloy by multi-objective Bayesian active learning with experimental uncertainty. npj Computational Materials, 11(1), 10
2025
-
[27]
Wang, K., & Dowling, A. W. (2022). Bayesian optimization for chemical products and functional materials.Current Opinion in Chemical Engineering, 36, 100728
2022
-
[28]
S., Yakutovich, A
Mercado, R., Fu, R. S., Yakutovich, A. V., Talirz, L., Haranczyk, M., & Smit, B. (2018). In silico design of 2D and 3D covalent organic frameworks for methane storage applications. Chemistry of Materials, 30(15), 5069–5086
2018
-
[29]
Hashin, Z. (1980). Failure criteria for unidirectional fiber composites.Journal of Applied Mechanics, 47(2), 329–334
1980
-
[30]
Loutas, T., Oikonomou, A., & Rekatsinas, C. (2025). Bio-inspired discontinuous composite materials with a machine learning optimized architecture.Composite Structures, 351, 118597
2025
-
[31]
Gkatsis, V., Maratos, P., Rekatsinas, C., Giannakopoulos, G., & Krokidas, P. (2025). Density- aware active learning for materials discovery: a case study on functionalized nanoporous materials.Physical Chemistry Chemical Physics, 27(43), 23152–23165
2025
-
[32]
E., Ren, Z., Li, Q
Siemenn, A. E., Ren, Z., Li, Q. & Buonassisi, T. (2023). Fast Bayesian optimization of Needle-in- a-Haystack problems using zooming memory-based initialization (ZoMBI).npj Computational Materials, 9 (1)
2023
-
[33]
Soedarmadji, S
Wei, Y., Zhuang, V. Soedarmadji, S. & Sui, Y. (2024). Scalable Bayesian Optimization via Focalized Sparse Gaussian Processes.Advances in Neural Information Processing Systems, 37 1–25
2024
-
[34]
M., Papia, E.-M., & Giannakopou- los, G
Krokidas, P., Rekatsinas, C., Sioros, V., Chatziathanasiou, G. M., Papia, E.-M., & Giannakopou- los, G. (2026). Frugal Bayesian optimization: Scalable surrogates for data- and resource-limited discovery. arXiv preprint arXiv:2607.29225. 19
2026 arXiv
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.