Pith. sign in

REVIEW 4 major objections 5 minor 70 references

The paper claims that Gaussian Process surrogates are consistently the most compute-hungry and least effective option in Bayesian optimization across 8 benchmarks and 9 real-world datasets, and that a dataset-feature classifier can predict

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 11:16 UTC pith:RYSHXCEZ

load-bearing objection The compute-scaling measurements are solid, but the optimization-ranking metric is internally contradictory and likely inverted — the paper needs a major revision before its central claim can be trusted. the 4 major comments →

arxiv 2607.29225 v1 pith:RYSHXCEZ submitted 2026-07-31 cs.LG physics.chem-phphysics.comp-ph

Frugal Bayesian Optimization: Scalable Surrogates for Data- and Resource-Limited Discovery

Panagiotis Krokidas , Christoforos Rekatsinas , Vassilis Sioros , Grigorios M. Chatziathanasiou , Efi-Maria Papia , George Giannakopoulos This is my paper
classification cs.LG physics.chem-phphysics.comp-ph
keywords Bayesian optimizationsurrogate modelsGaussian processesRandom ForestsNGBoostBASScompute-aware benchmarkingsurrogate recommendation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that the standard default surrogate in Bayesian optimization, Gaussian Processes, is a poor default when both optimization performance and computational cost are evaluated. Across eight 4D benchmark functions and nine real-world datasets from materials, mechanics, robotics, chemistry, and ML, it measures wall-clock time, GPU memory, best-found value, and top-100 recall for GP, Random Forests, NGBoost, and Bayesian Adaptive Spline Surfaces. Its central finding is that GP training time grows superlinearly and memory keeps accumulating, while the alternatives scale nearly linearly with bounded memory, yet match or beat GP in optimization quality. The paper then trains a classifier that predicts surrogate rankings from cheap dataset features (size, dimensionality, fractal dimension, target variance), achieving high nDCG. If correct, this makes GP-based BO a dominated choice on these problem classes and provides a practical, a-priori surrogate selection tool.

Core claim

The paper's core claim is that, on a compute-aware evaluation, Gaussian Process surrogates are consistently dominated: they incur the highest time and memory overhead without delivering superior optimization or sample efficiency, while Random Forests, NGBoost, and Bayesian Adaptive Spline Surfaces achieve equal or better results at a fraction of the cost. The authors establish this through a two-axis evaluation protocol that multiplies area-under-curve metrics for best-found value and top-100 recall, separately as functions of compute time and of acquired samples, across 8 benchmark functions and 9 real-world datasets. They further claim that a multi-output classifier trained on four inexpen

What carries the argument

The central mechanism is the compute-aware evaluation protocol: a paired measurement of optimization quality (best-found value and top-100 recall) against two resources (wall-clock compute time and GPU memory, alongside the number of acquired samples), summarized by AUC products. On top of this, the paper builds a surrogate-recommendation classifier that takes four predetermined dataset features and outputs a ranking of the four surrogates, trained with nDCG as the ranking quality metric.

Load-bearing premise

The load-bearing premise is that the 17 benchmark tasks—eight 4D multimodal functions and nine fully enumerated discrete candidate sets—are representative of the optimization problems where BO is actually used; if real problems include smooth low-dimensional continuous landscapes or much smaller evaluation budgets, the generalization 'GP is seldom justified' may fail even though each measured result is correct.

What would settle it

Run Bayesian optimization on a smooth, low-dimensional continuous problem (e.g., a 2D or 3D analytical function) with an evaluation budget of 50–100 samples, comparing GP against RF, NGB, and BASS under the authors' metrics; if GP reaches the optimum with fewer samples or with better best-found value at equal compute, the paper's central generalization that GP is dominated would be contradicted. Also, rerun BASS and NGB with default hyperparameters rather than the frugal settings to test whether their equal-or-better performance depends on the deliberately reduced configurations.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, GP-based BO should no longer be the automatic default; practitioners on compute-limited or sample-limited budgets should start with RF (compute-limited) or NGB (sample-limited) based on the classifier.
  • The reported scaling differences mean GP runs can terminate prematurely on standard workstations due to memory pressure, while the alternatives can run 1000-acquisition loops on standard hardware.
  • The recommendation layer enables a-priori surrogate selection from cheap dataset characteristics, avoiding costly pilot comparisons across surrogates.
  • The two-axis evaluation protocol itself could become a standard template for compute-aware benchmarking in BO research.
  • These savings compound: since the alternatives are cheaper per acquisition, they can afford more acquisition steps or broader exploration under a fixed wall-clock budget.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The benchmark suite contains only 4D multimodal functions and discrete candidate sets with known top-100 lists; if real workloads include smooth, low-dimensional continuous landscapes or very small budgets (tens of acquisitions), GP's calibrated uncertainty may regain an advantage—a testable extension beyond the paper's claim.
  • Editorial inference: The frugal hyperparameter choices for BASS and NGB (100 MCMC samples; 100 estimators with a higher learning rate) may understate their full predictive quality; a sensitivity study varying those budgets would clarify whether the equal-or-better result survives at their default settings.
  • Editorial inference: The classifier's high nDCG is achieved on only 17 datasets; a stronger test would be to apply it out-of-the-box to new, unseen problems and check whether the recommended surrogate indeed tops the rankings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents FruBO, a compute-aware benchmarking framework that compares four Bayesian Optimization surrogates — Gaussian Processes, Random Forests, BASS, and NGBoost — on eight synthetic benchmark functions and nine real-world discrete datasets, with N = 1000 acquisitions. Each surrogate is evaluated along two axes: optimization quality (best-found value and top-100 identification, aggregated as the product of two AUCs) and computational frugality (wall-clock training time and GPU memory). The authors report that GP-based BO incurs the highest time and memory overhead without delivering superior optimization or sample efficiency, that RF/NGB/BASS achieve equal or better performance at a fraction of the cost, and they introduce a dataset-feature-based classifier that predicts surrogate rankings with high nDCG.

Significance. If the empirical claims hold, this is a useful compute-aware benchmark and a practical resource-allocation tool for BO practitioners. The compute-scaling measurements are credible and consistent with known O(n^3) GP complexity, the inclusion of memory profiling is valuable, and the open-source framework is a strength. However, the optimization-quality rankings — the second half of the central claim — currently rest on an ambiguous and possibly inverted AUC definition, and the benchmark suite excludes the smooth, low-dimensional, small-budget regimes where GP-based BO is classically strongest. The broad conclusion that 'defaulting to GP is seldom justified' is therefore not yet supported.

major comments (4)
  1. [§1.2 and Tables S1–S8] The direction of the optimization-quality metric is internally inconsistent. The text states that 'smaller AUC values indicate better overall performance' for the product of best-found and top-100 identification AUCs. For the top-100 curve, however, the plotted quantity is 'the number of top-100 global solutions identified' (Figures 2–3), a cumulative count for which larger AUC means faster discovery. Tables S1–S8 report top-100 AUC values of order 1e2–1e5, so the product is dominated by this component; if raw cumulative-count AUC is used, the top-100 component is inverted and the rankings in Table 1 could reverse. If instead the code computes area-over-the-curve or a regret transform, that must be stated explicitly. No significance test or alternative ranking metric is reported. This must be resolved before the optimization-quality conclusions can be assessed.
  2. [§3.1] The non-GP surrogates are deliberately configured for frugality: BASS is run with nmcmc=10000 and nburn=9900, leaving only 100 effective posterior samples, and NGBoost uses 100 estimators instead of the default 500. The paper is transparent about this, but the abstract's claim that these alternatives 'achieve equal or better performance at a fraction of the computational cost' is established for one operating point only. No sensitivity analysis is given for the number of BASS samples or NGB estimators, and no matched-resource or default-hyperparameter comparison is reported. The conclusion that GP is dominated would be considerably strengthened by showing that the ranking is robust to these choices.
  3. [§1.2–1.3] The benchmark composition limits the external validity of the central generalization. All eight synthetic functions are 4D multimodal landscapes (Rastrigin, Ackley, Schwefel, Michalewicz, Schaffer 7, Styblinski–Tang, Weierstrass, Expanded Schaffer F6), and all nine real-world tasks are fully enumerated discrete candidate sets. No smooth, low-dimensional continuous landscape — the regime where GP covariance smoothness and calibrated uncertainty are classically most advantageous — is included. In addition, the fixed budget is N=1000 acquisitions; the sample-limited regime of tens of evaluations, which the paper itself identifies as practically important, is not tested. The Discussion's claim that 'defaulting to GP is seldom justified across diverse problem settings' is stronger than what these 17 tasks support.
  4. [§1.4] The surrogate-recommendation framework uses, as one of its four features, the target variance estimated from a first 200-acquisition run with a GP surrogate. This is not an inexpensive or purely a-priori dataset characteristic, and it somewhat undercuts the claim that the recommendation is obtained from 'inexpensive dataset characteristics'. Moreover, the nDCG evaluation has no baseline comparison (e.g., always predicting the overall best surrogate or ranking by dataset size), and with 17 datasets and four possible ranks, high nDCG values need calibration. The recommendation-layer conclusion is secondary, but the evaluation should be made more rigorous.
minor comments (5)
  1. [Figure 1 caption] The caption says 'across nine benchmark functions' but only eight benchmark functions are used.
  2. [Figure 3 caption] The caption says 'eight real-case data' but nine real-world datasets are used.
  3. [Table 1 and Table S6] The real-case dataset names are inconsistent: 'MOFs' is used where the dataset is COFs, 'proteins' refers to the Ro4 docking dataset, and Table S6 contains 'GM9' instead of 'QM9'.
  4. [Appendix A] Appendix A contains boilerplate placeholder text ('An appendix contains supplementary information...') and should be removed or replaced with actual content.
  5. [Code Availability] The license is stated as 'Apache 2.9'; the current Apache license is version 2.0. This is likely a typo.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical benchmark whose conclusions are direct measurements, not derivations from equivalence-by-construction inputs.

full rationale

This is an empirical benchmarking study, not a derivation chain. The load-bearing claims—GP incurs higher time/memory, and RF/BASS/NGB match or beat GP in optimization quality—are direct measurements over 8 benchmark functions and 9 real-world datasets, with quantitative AUC tables and compute/memory curves. No equation defines the target in terms of the input or vice versa; no fitted parameter is renamed as a prediction. The recommendation classifier (Sec. 1.4) is a standard supervised meta-model: features (dataset size, dimensionality, fractal dimension, target variance estimated from an initial 200-GP-acquisition phase) are cross-validated against rankings derived from the same benchmark. This is fitting and evaluating a predictor, not re-deriving the rankings by construction. The deliberately frugal BASS/NGB hyperparameters (Methods: nburn=9900/10000, n_estimators=100) are disclosed as explicit assumptions; they make the compute comparison configuration-specific but do not make the outcome equivalent to the input. Self-citations ([6], [56], [63], [68–70]) support dataset construction or background and are not load-bearing for the central claim, and no uniqueness theorem is imported. The AUC-direction ambiguity for the top-100 identification curve is a potential metric-correctness concern, not a circularity: even if the direction were wrong, the error would be an empirical/measurement artifact rather than a reduction of the conclusion to its premises.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claims rest on: (1) benchmark representativeness (all eight functions are multimodal 4D; all nine real datasets are fully enumerated discrete candidate sets, excluding smooth continuous GP-favorable regimes); (2) the validity of the product-of-two-AUCs aggregation used to rank surrogates; (3) the sufficiency of the four recommender features, one of which (target variance) is measured by running 200 GP acquisitions and is entangled with the surrogate being recommended, and the adequacy of nDCG without a baseline. No new physical entities are introduced; FruBO is a software/protocol contribution. Free parameters: hand-chosen frugal BASS and NGB hyperparameters, plus the hand-chosen AUC-product aggregation.

free parameters (3)
  • BASS MCMC burn-in schedule = nmcmc=10000, nburn=9900 (100 effective posterior samples)
    Hand-chosen frugal setting that trades BASS predictive quality for speed; the paper explicitly notes this 'prioritizes frugality over extracting the very highest predictive performance' (Methods 3.1). The BO comparison is therefore at one operating point on the cost-quality curve.
  • NGBoost ensemble size / learning rate = n_estimators=100, learning_rate=0.1 (defaults are 500 / 0.01)
    Hand-chosen frugal configuration (Methods 3.1); same operating-point concern, plus the learning-rate justification is borrowed from large-data practice.
  • AUC-product aggregation = product of best-found AUC and top-100 AUC, per axis
    Chosen by hand to define a single 'compute-time performance' and 'sample-acquisition performance' metric (Section 1.2, Fig. 2); no justification for multiplicative rather than additive aggregation, and the units mix objective values with counts.
axioms (5)
  • domain assumption The 4D multi-modal benchmark functions plus the nine enumerated real-world datasets are representative of practical BO workloads.
    Underlies the generalization 'defaulting to GP is seldom justified' (Discussion §2). No smooth/unimodal continuous benchmark is included, and all real datasets are fully enumerated discrete sets, so GP-favorable regimes are absent.
  • ad hoc to paper Multiplying the best-found AUC and top-100 AUC yields a valid ordering of surrogate performance on each axis.
    Section 1.2 defines the single metric this way without justification; rankings in Table 1 and all 'clear winner' statements inherit this choice.
  • domain assumption The four dataset features (size, dimensionality, fractal dimension, target variance from the first 200 GP acquisitions) suffice to predict surrogate ranking.
    Section 1.4; no feature-ablation or alternative-model comparison is given, the fractal-dimension estimator is unspecified, and target variance is entangled with a 200-step GP run.
  • domain assumption nDCG with relevance equal to the true rank of each surrogate is an adequate measure of recommender quality, and high nDCG without a baseline indicates useful prediction.
    Section 1.4 Eqs. (1)-(3); with only 4 ranked items, a random ranking already achieves nDCG ~ 0.78, so a baseline is needed to interpret 'consistently high' values.
  • standard math GP complexity is O(n^3) in time and O(n^2) in memory; RF/NGB/BASS scale near-linearly.
    Standard results (Rasmussen and Williams 2006; cited refs. [7,8]) used to interpret the measured scaling curves (Fig. 1).

pith-pipeline@v1.3.0-daily-deepseek · 25862 in / 31465 out tokens · 311510 ms · 2026-08-03T11:16:13.636361+00:00 · methodology

0 comments
read the original abstract

Bayesian Optimization (BO) is widely adopted for data-efficient optimization in scientific and engineering applications, yet its computational cost is rarely evaluated alongside optimization performance. Here we present a systematic, compute-aware study of BO that evaluates surrogate models along two axes: optimization quality and computational frugality. Across eight benchmark functions and nine real-world datasets spanning materials science, mechanics, robotics, chemistry, and machine learning, we benchmark four surrogate models: Gaussian Processes, Random Forests, NGBoost, and Bayesian Adaptive Spline Surfaces. We show that Gaussian Process-based BO consistently incurs the highest time and memory overhead without delivering superior optimization or sample efficiency. In contrast, scalable alternatives achieve equal or better performance at a fraction of the computational cost. Motivated by these findings, we introduce a surrogate-recommendation framework that predicts the most suitable BO surrogate from inexpensive dataset characteristics. Together, these results establish FruBO as a reproducible, compute-aware baseline for Bayesian Optimization and provide practical guidance for surrogate selection under limited computational and experimental budgets.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

70 extracted references · 16 canonical work pages

  1. [1]

    Technical Report January, University of Wisconsin–Madison, Madison (2009)

    Settles, B.: Active Learning Literature Survey. Technical Report January, University of Wisconsin–Madison, Madison (2009)

    2009

  2. [2]

    Proceedings of the IEEE104(1), 148–175 (2016) https://doi.org/10.1109/JPROC.2015.2494218

    Shahriari, B., Swersky, K., Wang, Z., Adams, R.P., De Freitas, N.: Taking the human out of the loop: A review of Bayesian optimization. Proceedings of the IEEE104(1), 148–175 (2016) https://doi.org/10.1109/JPROC.2015.2494218

    arXiv 2016

  3. [3]

    npj Computational Materials10(1), 1–12 (2024) https://doi.org/10.1038/s41524-024-01326-2 arXiv:2312.16078

    Chitturi, S.R., Ramdas, A., Wu, Y., Rohr, B., Ermon, S., Dionne, J., Jornada, F.H.d., Dunne, M., Tassone, C., Neiswanger, W., Ratner, D.: Targeted materials discovery using Bayesian algorithm execution. npj Computational Materials10(1), 1–12 (2024) https://doi.org/10.1038/s41524-024-01326-2 arXiv:2312.16078

    Pith/arXiv arXiv 2024

  4. [4]

    Digital Discovery3(6), 1086–1100 (2024) https://doi.org/10

    Wu, Y., Walsh, A., Ganose, A.M.: Race to the bottom: Bayesian optimisation for chemical problems. Digital Discovery3(6), 1086–1100 (2024) https://doi.org/10. 1039/d3dd00234a 32

    2024

  5. [5]

    Composite Structures 351, 118597 (2025) https://doi.org/10.1016/j.compstruct.2024.118597

    Loutas, T., Oikonomou, A., Rekatsinas, C.: Bio-inspired discontinuous composite materials with a machine learning optimized architecture. Composite Structures 351, 118597 (2025) https://doi.org/10.1016/j.compstruct.2024.118597

    arXiv 2025

  6. [6]

    Digital Discovery4, 3753–3763 (2025) https://doi.org/10.1039/d5dd00237k

    Krokidas, P., Gkatsis, V., Theocharis, J., Giannakopoulos, G.: Navigating materials design spaces with efficient Bayesian optimization: a case study in functionalized nanoporous materials. Digital Discovery4, 3753–3763 (2025) https://doi.org/10.1039/d5dd00237k

    work page doi:10.1039/d5dd00237k 2025

  7. [7]

    MIT Press, Cambridge, MA (2006)

    Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. MIT Press, Cambridge, MA (2006)

    2006

  8. [8]

    In: 2023 IEEE High Performance Extreme Computing Conference, HPEC 2023 (2023)

    Siemenn, A.E., Buonassisi, T.: Decreasing the Computing Time of Bayesian Opti- mization Using Generalizable Memory Pruning. In: 2023 IEEE High Performance Extreme Computing Conference, HPEC 2023 (2023). https://doi.org/10.1109/ HPEC58863.2023.10363486

    arXiv 2023

  9. [9]

    European Physical Journal C81(2), 1–9 (2021) https://doi.org/10.1140/epjc/ s10052-021-08950-y arXiv:2011.04434

    Tani, L., Rand, D., Veelken, C., Kadastik, M.: Evolutionary algorithms for hyper- parameter optimization in machine learning for application in high energy physics. European Physical Journal C81(2), 1–9 (2021) https://doi.org/10.1140/epjc/ s10052-021-08950-y arXiv:2011.04434

    Pith/arXiv arXiv 2021

  10. [10]

    Computer Physics Communications294, 108955 (2024) https://doi.org/ 10.1016/j.cpc.2023.108955 arXiv:2201.06809

    Tani, L., Veelken, C.: Comparison of Bayesian and particle swarm algorithms for hyperparameter optimisation in machine learning applications in high energy physics. Computer Physics Communications294, 108955 (2024) https://doi.org/ 10.1016/j.cpc.2023.108955 arXiv:2201.06809

    arXiv 2024

  11. [11]

    ACM Transactions on Evolutionary Learning and Optimization4(3) (2024) https://doi.org/10.1145/ 3670683 arXiv:2303.00890

    Santoni, M.L., Raponi, E., De Leone, R., Doerr, C.: Comparison of High- Dimensional Bayesian Optimization Algorithms on BBOB. ACM Transactions on Evolutionary Learning and Optimization4(3) (2024) https://doi.org/10.1145/ 3670683 arXiv:2303.00890

    Pith/arXiv arXiv 2024

  12. [12]

    npj Computational Materials9(1) (2023) https://doi.org/10.1038/ s41524-023-01048-x arXiv:2208.13771

    Siemenn, A.E., Ren, Z., Li, Q., Buonassisi, T.: Fast Bayesian optimization of Needle-in-a-Haystack problems using zooming memory-based initialization (ZoMBI). npj Computational Materials9(1) (2023) https://doi.org/10.1038/ s41524-023-01048-x arXiv:2208.13771

    Pith/arXiv arXiv 2023

  13. [13]

    Advances in Neural Information Processing Systems37(NeurIPS), 1–25 (2024)

    Wei, Y., Zhuang, V., Soedarmadji, S., Sui, Y.: Scalable Bayesian Optimization via Focalized Sparse Gaussian Processes. Advances in Neural Information Processing Systems37(NeurIPS), 1–25 (2024)

    2024

  14. [14]

    Proceedings of Machine Learning Research202, 4035–4054 (2023) arXiv:2306.03566

    Chang, P.E., Verma, P., John, S.T., Solin, A., Khan, M.E.: Memory-Based Dual Gaussian Processes for Sequential Learning. Proceedings of Machine Learning Research202, 4035–4054 (2023) arXiv:2306.03566

    Pith/arXiv arXiv 2023

  15. [15]

    Nature Computational Science 5(January) (2024) https://doi.org/10.1038/s43588-024-00744-y

    Lin, Y., Gao, B., Tang, J., Zhang, Q., Qian, H., Wu, H.: Deep Bayesian active learning using in-memory computing hardware. Nature Computational Science 5(January) (2024) https://doi.org/10.1038/s43588-024-00744-y

    work page doi:10.1038/s43588-024-00744-y 2024

  16. [16]

    IFAC-PapersOnLine59(6), 157–162 (2025) https://doi.org/10.1016/J.IFACOL.2025.07.138 arXiv:2504.10076

    Makrygiorgos, G., Ip, J.H.S., Mesbah, A.: Towards Scalable Bayesian Optimization 33 via Gradient-Informed Bayesian Neural Networks. IFAC-PapersOnLine59(6), 157–162 (2025) https://doi.org/10.1016/J.IFACOL.2025.07.138 arXiv:2504.10076

    Pith/arXiv arXiv 2025

  17. [17]

    In: 12th International Conference on Learning Representations, ICLR 2024 (2024)

    Li, Y.L., Rudner, T.G.J., Wilson, A.G.: a Study of Bayesian Neural Network Surrogates for Bayesian Optimization. In: 12th International Conference on Learning Representations, ICLR 2024 (2024)

    2024

  18. [18]

    Proceedings of the 19th International Con- ference on Artificial Intelligence and Statistics, AISTATS 201651, 1478–1487 (2016) arXiv:1506.03805

    Lakshminarayanan, B., Roy, D.M., Teh, Y.W.: Mondrian forests for large-scale regression when uncertainty matters. Proceedings of the 19th International Con- ference on Artificial Intelligence and Statistics, AISTATS 201651, 1478–1487 (2016) arXiv:1506.03805

    Pith/arXiv arXiv 2016

  19. [19]

    Digital Discovery2(3), 759–774 (2023) https://doi.org/10.1039/d2dd00146b arXiv:2212.01574

    Tom, G., Hickman, R., Zinzuwadia, A., Mohajeri, A., Sanchez-Lengeling, B., Aspuru-Guzik, A.: Calibration and generalizability of probabilistic models on low-data chemical datasets with DIONYSUS. Digital Discovery2(3), 759–774 (2023) https://doi.org/10.1039/d2dd00146b arXiv:2212.01574

    Pith/arXiv arXiv 2023

  20. [20]

    npj Computational Materials7(1), 1–12 (2021) https: //doi.org/10.1038/s41524-021-00662-x

    Lei, B., Kirk, T.Q., Bhattacharya, A., Pati, D., Qian, X., Arroyave, R., Mallick, B.K.: Bayesian optimization with adaptive surrogate models for automated experimental design. npj Computational Materials7(1), 1–12 (2021) https: //doi.org/10.1038/s41524-021-00662-x

    doi:10.1038/s41524-021-00662-x 2021

  21. [21]

    32nd International Conference on Machine Learning, ICML 20153, 2161–2170 (2015) arXiv:1502.05700

    Snoek, J., Ripped, O., Swersky, K., Kiros, R., Satish, N., Sundaram, N., Patwary, M.M.A., Prabhat, Adams, R.P.: Scalable Bayesian optimization using deep neural networks. 32nd International Conference on Machine Learning, ICML 20153, 2161–2170 (2015) arXiv:1502.05700

    Pith/arXiv arXiv 2015

  22. [22]

    Challenges6(1), 117–157 (2015) https://doi.org/10.3390/ challe6010117

    Andrae, A., Edler, T.: On Global Electricity Usage of Communication Technol- ogy: Trends to 2030. Challenges6(1), 117–157 (2015) https://doi.org/10.3390/ challe6010117

    2030

  23. [23]

    Journal of Building Engineering91(May), 109519 (2024) https://doi.org/10.1016/j.jobe.2024.109519

    Yang, D., Wang, X., Shen, R., Li, Y., Gu, L., Zheng, R., Zhao, J., Tian, X.: Global optimization strategy of prosumer data center system operation based on multi- agent deep reinforcement learning. Journal of Building Engineering91(May), 109519 (2024) https://doi.org/10.1016/j.jobe.2024.109519

    arXiv 2024

  24. [24]

    Computer55(7), 18–28 (2022) https://doi

    Patterson,D.,Gonzalez,J.,Holzle,U.,Le,Q.,Liang,C.,Munguia,L.M.,Rothchild, D., So, D.R., Texier, M., Dean, J.: The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. Computer55(7), 18–28 (2022) https://doi. org/10.1109/MC.2022.3148714 arXiv:2204.05149

    arXiv 2022

  25. [25]

    arXiv, 1–25 (2024) arXiv:2403.08151

    Tripp, C.E., Perr-Sauer, J., Gafur, J., Nag, A., Purkayastha, A., Zisman, S., Bensen, E.A.: Measuring the Energy Consumption and Efficiency of Deep Neural Networks: An Empirical Analysis and Design Recommendations. arXiv, 1–25 (2024) arXiv:2403.08151

    Pith/arXiv arXiv 2024

  26. [26]

    Joule7(10), 2191–2194 (2023) https://doi.org/10.1016/j.joule.2023.09.004

    Vries, A.: The growing energy footprint of artificial intelligence. Joule7(10), 2191–2194 (2023) https://doi.org/10.1016/j.joule.2023.09.004

    doi:10.1016/j.joule.2023.09.004 2023

  27. [27]

    Reuters (2023)

    Mehta, C., Cherney, M.A., Nellis, S.: Nvidia adds jet fuel to AI optimism with 34 record results, $25 billion buyback. Reuters (2023)

    2023

  28. [28]

    Steven, R., Klymenko, O.V., Short, M.: Machine learning-accelerated distributed optimisationmethodsforoptimalpowerflow:Areview.RenewableandSustainable EnergyReviews226,116190 (2026)https://doi.org/10.1016/J.RSER.2025.116190

    arXiv 2026

  29. [29]

    (eds.): Advances in Production Management Systems

    Mizuyama, H., Morinaga, E., Nonaka, T., Kaihara, T., Cieminski, G., Romero, D. (eds.): Advances in Production Management Systems. Cyber-Physical-Human Production Systems: Human-AI Collaboration and Beyond. IFIP Advances in Information and Communication Technology, vol. 767. Springer, Cham (2026). https://doi.org/10.1007/978-3-032-03542-4 . Conference proceedings

    work page doi:10.1007/978-3-032-03542-4 2026

  30. [30]

    Sustainability 17(15), 6891 (2025) https://doi.org/10.3390/su17156891

    Rahman, S., Ahsan, A., Pramanik, N.I.: Climate-Regulating Industrial Ecosystems: An AI-Optimised Framework for Green Infrastructure Performance. Sustainability 17(15), 6891 (2025) https://doi.org/10.3390/su17156891

    work page doi:10.3390/su17156891 2025

  31. [31]

    Artificial Intelligence in Geosciences6(2), 100147 (2025) https://doi.org/10.1016/J.AIIG.2025.100147

    Naturinda, E., Kemigyisha, F., Gidudu, A., Kabenge, I., Omia, E., Aboth, J.: Quantification of greenhouse gas emissions from livestock using remote sensing & artificial intelligence. Artificial Intelligence in Geosciences6(2), 100147 (2025) https://doi.org/10.1016/J.AIIG.2025.100147

    arXiv 2025

  32. [32]

    Xin, Y., Zhu, M., Zhang, H., Wang, X.: High-Entropy Materials: A New Paradigm in the Design of Advanced Batteries vol. 18. Springer, ??? (2026). https://doi. org/10.1007/s40820-025-01842-w

    work page doi:10.1007/s40820-025-01842-w 2026

  33. [33]

    Nano-Micro Letters17(1), 1–30 (2025) https://doi.org/10.1007/s40820-024-01634-8

    Bai, X., Zhang, X.: Artificial Intelligence-Powered Materials Science. Nano-Micro Letters17(1), 1–30 (2025) https://doi.org/10.1007/s40820-024-01634-8

    work page doi:10.1007/s40820-024-01634-8 2025

  34. [34]

    Chemical Engineering Journal517, 164419 (2025) https: //doi.org/10.1016/j.cej.2025.164419

    Liu, Z., Shen, D., Chung, Y.G., Li, W., Li, S.: Accelerated discovery of high- performance MOFs for water adsorption chillers through molecular simulation and machine learning. Chemical Engineering Journal517, 164419 (2025) https: //doi.org/10.1016/j.cej.2025.164419

    arXiv 2025

  35. [35]

    Advanced Functional Materials34(43) (2024) https://doi.org/10.1002/adfm.202307478

    Bose, S., Sengupta, D., Rayder, T.M., Wang, X., Kirlikovali, K.O., Sekizkardes, A.K., Islamoglu, T., Farha, O.K.: Challenges and Opportunities: Metal–Organic Frameworks for Direct Air Capture. Advanced Functional Materials34(43) (2024) https://doi.org/10.1002/adfm.202307478

    work page doi:10.1002/adfm.202307478 2024

  36. [36]

    Luers, A., Koomey, J., Masanet, E., Gaffney, O., Creutzig, F., Lavista Ferres, J., Horvitz, E.: Will AI accelerate or delay the race to net-zero emissions? Nature 628(8009), 718–720 (2024) https://doi.org/10.1038/d41586-024-01137-x

    work page doi:10.1038/d41586-024-01137-x 2024

  37. [37]

    Luers,A.:NetzeroneedsAI—fiveactionstorealizeitspromise.Nature644(8078), 871–873 (2025) https://doi.org/10.1038/d41586-025-02641-4

    doi:10.1038/d41586-025-02641-4 2025

  38. [38]

    Journal of Statistical Software94(8), 1–36 (2020) https://doi.org/10.18637/jss.v094.i08

    Francom, D., Sansó, B.: BASS: An R package for fitting and performing sensitivity analysis of bayesian adaptive spline surfaces. Journal of Statistical Software94(8), 1–36 (2020) https://doi.org/10.18637/jss.v094.i08

    work page doi:10.18637/jss.v094.i08 2020

  39. [39]

    Duan, T., Avati, A., Ding, D.Y., Thai, K.K., Basu, S., Ng, A., Schuler, A.: 35 NGBoost:Naturalgradientboostingforprobabilisticprediction.37thInternational Conference on Machine Learning, ICML 2020PartF16814, 2670–2680 (2020) arXiv:1910.03225

    Pith/arXiv arXiv 2020

  40. [40]

    Journal of Machine Learning Research12, 2825–2830 (2011)

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research12, 2825–2830 (2011)

    2011

  41. [41]

    Advances in Neural Information Processing Systems2020-December(MC) (2020) arXiv:1910.06403

    Balandat, M., Karrer, B., Jiang, D.R., Daulton, S., Letham, B., Wilson, A.G., Bakshy, E.: BOTORCH: A framework for efficient Monte-Carlo Bayesian optimiza- tion. Advances in Neural Information Processing Systems2020-December(MC) (2020) arXiv:1910.06403

    Pith/arXiv arXiv 2020

  42. [42]

    In: Coello, C.A.C

    Hutter, F., Hoos, H.H., Leyton-Brown, K.: Sequential Model-Based Optimization for General Algorithm Configuration. In: Coello, C.A.C. (ed.) Learning and Intelligent Optimization, pp. 507–523. Springer, Berlin, Heidelberg (2011)

    2011

  43. [43]

    Proceedings - IEEE International Conference on Robotics and Automation, 16459–16466 (2024) https://doi.org/10.1109/ICRA57147.2024.10611468

    Styrud, J., Mayr, M., Hellsten, E., Krueger, V., Smith, C.: BeBOP-Combining Reactive Planning and Bayesian Optimization to Solve Robotic Manipulation Tasks. Proceedings - IEEE International Conference on Robotics and Automation, 16459–16466 (2024) https://doi.org/10.1109/ICRA57147.2024.10611468

    arXiv 2024

  44. [44]

    Statistics and Computing8(4), 337–346 (1998) https://doi.org/10.1023/A:1008824606259

    Denison, D.G.T., Mallick, B.K., Smith, A.F.M.: Bayesian MARS. Statistics and Computing8(4), 337–346 (1998) https://doi.org/10.1023/A:1008824606259

    work page doi:10.1023/a:1008824606259 1998

  45. [45]

    Buildings12(12) (2022) https://doi.org/10.3390/buildings12122109

    Li, G., Tian, W., Zhang, H., Chen, B.: Building Energy Models at Different Time Scales Based on Multi-Output Machine Learning. Buildings12(12) (2022) https://doi.org/10.3390/buildings12122109

    work page doi:10.3390/buildings12122109 2022

  46. [46]

    Applied Energy 363, 123042 (2024) https://doi.org/10.1016/j.apenergy.2024.123042

    Zhang, H., Tian, W., Tan, J., Yin, J., Fu, X.: Sensitivity analysis of multiple time- scale building energy using Bayesian adaptive spline surfaces. Applied Energy 363, 123042 (2024) https://doi.org/10.1016/j.apenergy.2024.123042

    arXiv 2024

  47. [47]

    Frontiers in Nuclear Engineering1(December), 1–11 (2022) https://doi.org/10.3389/fnuen.2022.1083164

    Ausdemore, M.A., McCombs, A., Ries, D., Zhang, A., Shuler, K., Tucker, J.D., Goode, K., Huerta, J.G.: A probabilistic inverse prediction method for predicting plutonium processing conditions. Frontiers in Nuclear Engineering1(December), 1–11 (2022) https://doi.org/10.3389/fnuen.2022.1083164

    arXiv 2022

  48. [48]

    Chemistry of Materials 36(22), 11109–11118 (2024) https://doi.org/10.1021/acs.chemmater.4c01978

    Davis, J.V., Marrs, F.W., Cawkwell, M.J., Manner, V.W.: Machine Learning Mod- els for High Explosive Crystal Density and Performance. Chemistry of Materials 36(22), 11109–11118 (2024) https://doi.org/10.1021/acs.chemmater.4c01978

    doi:10.1021/acs.chemmater.4c01978 2024

  49. [49]

    Automation in Construction170, 105943 (2025) https://doi.org/10.1016/j.autcon.2024.105943

    Chen, H., Liu, J., Shen, G.Q., Feng, Z.: Control of existing tunnel deformation caused by shield adjacent undercrossing construction using interpretable machine learning and multiobjective optimization. Automation in Construction170, 105943 (2025) https://doi.org/10.1016/j.autcon.2024.105943

    arXiv 2025

  50. [50]

    Chemistry of Materials30(15), 5069–5086 (2018) https://doi.org/ 10.1021/acs.chemmater.8b01425

    Mercado, R., Fu, R.S., Yakutovich, A.V., Talirz, L., Haranczyk, M., Smit, B.: In Silico Design of 2D and 3D Covalent Organic Frameworks for Methane Storage 36 Applications. Chemistry of Materials30(15), 5069–5086 (2018) https://doi.org/ 10.1021/acs.chemmater.8b01425

    work page doi:10.1021/acs.chemmater.8b01425 2018

  51. [51]

    Scientific Data1, 1–7 (2014) https://doi.org/10.1038/sdata.2014.22

    Ramakrishnan, R., Dral, P.O., Rupp, M., Von Lilienfeld, O.A.: Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data1, 1–7 (2014) https://doi.org/10.1038/sdata.2014.22

    doi:10.1038/sdata.2014.22 2014

  52. [52]

    Nature Computational Science5(April) (2025) https://doi.org/10.1038/ s43588-025-00777-x

    Luttens, A., Cabeza de Vaca, I., Sparring, L., Brea, J., Martínez, A.L., Kahlous, N.A., Radchenko, D.S., Moroz, Y.S., Loza, M.I., Norinder, U., Carlsson, J.: Rapid traversal of vast chemical space using machine learning-guided docking screens. Nature Computational Science5(April) (2025) https://doi.org/10.1038/ s43588-025-00777-x

    2025

  53. [53]

    Nature Computational Science5(September) (2024) https://doi.org/10.1038/s43588-025-00858-x

    Wei, Y., Peng, B., Xie, R., Chen, Y., Qin, Y., Wen, P., Bauer, S., Tung, P.-Y., Raabe, D.: Deep active learning for complex systems. Nature Computational Science5(September) (2024) https://doi.org/10.1038/s43588-025-00858-x

    work page doi:10.1038/s43588-025-00858-x 2024

  54. [54]

    Elmi, M., Pawar, V.M., Shaw, M., Wong, D., Zhan, H., Srinivasan, M.A.: Deter- mining the biomechanics of touch sensation in C. elegans. Scientific Reports7(1), 12329 (2017) https://doi.org/10.1038/s41598-017-12190-0

    work page doi:10.1038/s41598-017-12190-0 2017

  55. [55]

    elegans biomechanics: an empirical and multi-compartmental in silico modelling study

    Essmann, C.L., Elmi, M., Rekatsinas, C., Chrysochoidis, N., Shaw, M., Pawar, V., Srinivasan, M.A., Vavourakis, V.: The influence of internal pressure and neuromus- cular agents on C. elegans biomechanics: an empirical and multi-compartmental in silico modelling study. Frontiers in Bioengineering and Biotechnology12(March), 1–14 (2024) https://doi.org/10.3...

    arXiv 2024

  56. [56]

    Advanced Intelligence Discovery (2025)

    Rekatsinas, C., Krokidas, P., Vavourakis, V., Essmann, C., Giannakopoulos, G.: A physics-constrained machine learning pipeline for young’s modulus prediction in multi-material hyperelastic cylinders guided by contact mechanics. Advanced Intelligence Discovery (2025). In press

    2025

  57. [57]

    Journal of Applied Mechanics47(2), 329–334 (1980) https://doi.org/10.1115/1.3153664

    Hashin, Z.: Failure Criteria for Unidirectional Fiber Composites. Journal of Applied Mechanics47(2), 329–334 (1980) https://doi.org/10.1115/1.3153664

    doi:10.1115/1.3153664 1980

  58. [58]

    Journal of Composite Materials58(27), 2897–2914 (2024) https://doi.org/10

    Rekatsinas, C., Theodosiou, T., Siorikis, D., Tsiaktanis, K., Chrysochoidis, N., Nastos, C., Saravanos, D.: Micromechanics-based multi-scale framework with strain-rate effects for the simulation of ballistic impact on composite laminates. Journal of Composite Materials58(27), 2897–2914 (2024) https://doi.org/10. 1177/00219983241283618

    2024

  59. [59]

    In: Computer Aided Chemical Engineering vol

    Lee, Y.S., Graham, E., Jackson, G., Galindo, A., Adjiman, C.S.: A comparison of the performance of multi-objective optimization methodologies for solvent design. In: Computer Aided Chemical Engineering vol. 46, pp. 37–42. Elsevier, ??? (2019). https://doi.org/10.1016/B978-0-12-818634-3.50007-2

    work page doi:10.1016/b978-0-12-818634-3.50007-2 2019

  60. [60]

    Nature Reviews Materials6, 201–206 (2021) https://doi.org/10.1038/ s41578-021-00284-1 37

    Andrei, E.Y., Efetov, D.K., Jarillo-Herrero, P.,et al.: The marvels of moiré materials. Nature Reviews Materials6, 201–206 (2021) https://doi.org/10.1038/ s41578-021-00284-1 37

    2021

  61. [61]

    Nature Physics16(6), 631–635 (2020)

    Chen, X., Fan, X., Li, L., Zhang, N., Niu, Z., Guo, T.,et al.: Moiré engineering of electronic phenomena in correlated oxides. Nature Physics16(6), 631–635 (2020)

    2020

  62. [62]

    Materials Advances 2(17), 5542–5559 (2021)

    Shi, B., Qi, P., Jiang, M., Dai, Y., Lin, F., Zhang, H., Fang, Z.: Exotic physical properties of 2d materials modulated by moiré superlattices. Materials Advances 2(17), 5542–5559 (2021)

    2021

  63. [63]

    Computa- tional Materials Science261, 114270 (2026) https://doi.org/10.1016/j.commatsci

    Papia, E.-M., Kondi, A., Nioras, D., Sofos, F., Constantoudis, V.: Data-driven investigation of analytical relationships in moiré structured materials. Computa- tional Materials Science261, 114270 (2026) https://doi.org/10.1016/j.commatsci. 2025.114270

    arXiv 2026

  64. [64]

    Magnetism3(3), 245–258 (2023)

    Jadaun, P., Soreé, B.: Review of orbital magnetism in graphene-based moire materials. Magnetism3(3), 245–258 (2023)

    2023

  65. [65]

    APL Materials12(1) (2024)

    Pryds, N., Park, D.S., Jespersen, T.S., Yun, S.: Twisted oxide membranes: A perspective. APL Materials12(1) (2024)

    2024

  66. [66]

    Trends in Chemistry3(5), 342–358 (2021)

    Hu, G., Wang, M., Mazor, Y., Qiu, C.W., Alù, A.: Tailoring light with layered and moiré metasurfaces. Trends in Chemistry3(5), 342–358 (2021)

    2021

  67. [67]

    Science374(6571), 1140–1144 (2021)

    Song, T., Sun, Q.C., Anderson, E., Wang, C., Qian, J., Taniguchi, T.,et al.: Direct visualization of magnetic domains and moiré magnetism in twisted 2d magnets. Science374(6571), 1140–1144 (2021)

    2021

  68. [68]

    Journal of Sound and Vibration537(July), 117222 (2022) https://doi.org/10.1016/j.jsv.2022.117222

    Chatziathanasiou, G.M., Chrysochoidis, N.A., Rekatsinas, C.S., Saravanos, D.A.: A semi-active shunted piezoelectric tuned-mass-damper for multi-modal vibration control of large flexible structures. Journal of Sound and Vibration537(July), 117222 (2022) https://doi.org/10.1016/j.jsv.2022.117222

    arXiv 2022

  69. [69]

    Journal of Vibration and Control28(21-22), 2969–2983 (2022) https://doi.org/10.1177/ 10775463211026487

    Chatziathanasiou, G.M., Chrysochoidis, N.A., Saravanos, D.A.: A semi-active shunted piezoelectric tuned mass damper for robust vibration control. Journal of Vibration and Control28(21-22), 2969–2983 (2022) https://doi.org/10.1177/ 10775463211026487

    2022

  70. [70]

    Demokritos

    Chatziathanasiou, G.M., Chrysochoidis, N.A., Saravanos, D.A.: Multi-tonal vibration suppression in flexible airframe structures via a semi-active shunted piezoelectric tuned mass damper. Journal of Vibration and Control0(0) (2025) https://doi.org/10.1177/10775463251341365 38 Frugal Bayesian Optimization: Scalable Surrogates for Data- and Resource-Limited ...

    work page doi:10.1177/10775463251341365 2025