Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Benchmarking Self-Driving Labs

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Across 42 benchmarking studies, self-driving laboratories reach a target performance in a median of six times fewer experiments than reference strategies, and the advantage grows with dimensionality.

desk verdict Useful systematic compilation of AF/EF values for the SDL field, but the AF–dimensionality trend is over-interpreted given the uncontrolled baselines; the EF peak is more solid. read the letter →

arxiv 2508.06642 v1 pith:55HJOVF4 submitted 2025-08-08 physics.comp-ph cond-mat.mtrl-scics.LGphysics.data-an

classification physics.comp-phcond-mat.mtrl-scics.LGphysics.data-an
keywords self-drivinglabsBayesianoptimizationaccelerationfactorenhancementactivelearningbenchmarkingdimensionalitymaterialsdiscoveryautonomousexperimentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the value of self-driving laboratories is real, measurable, and predictable. Reviewing 42 benchmarking studies, it finds that active-learning systems reach a target performance in a median of 6 times fewer experiments than a reference strategy, and that this acceleration tends to grow with the dimensionality $d$ of the search space—what the authors call a "blessing of dimensionality." It also finds that enhancement factors, which measure how much better the algorithm's best result is after a fixed number of experiments, vary by more than two orders of magnitude but consistently peak at about 10–20 experiments per dimension. Simulated Bayesian optimization campaigns on Gaussian test functions explain the spread: the maximum enhancement is set by the contrast of the property space, while the speed of convergence is set by its Lipschitz complexity and is degraded by noise. If correct, these results give the field shared metrics and concrete expectations for when self-driving labs pay off.

What carries the argument

The load-bearing objects are the two metrics $AF$ and $EF$, defined relative to a reference campaign; the random-sampling convergence law $1/2 = F_y(y_n)^n$, which gives the expected best result of uniform sampling; the contrast $C = y^*/\mathrm{median}(y)$, which caps $EF$; and the Lipschitz complexity $L=\max|\nabla f|$, which sets how many experiments are needed to converge. The simulations connect these measurable statistics of the objective function to the observed benchmark scatter, explaining why $EF$ varies so widely across materials spaces even when algorithms are similar.

What would settle it

Re-run the meta-analysis restricted to benchmarks that used uniform random sampling as the reference and experimental data as ground truth, and test whether AF still rises with dimensionality; a Spearman rank correlation indistinguishable from zero in that matched subset would falsify the 'blessing of dimensionality' claim. An even sharper test: for each study, estimate the contrast $C$ and check whether the reported maximum $EF$ exceeds $C$; any violation would indicate a misreported metric.

Watch

Extended reading notes

Core claim

The paper's central claim is that two dimensionless metrics—acceleration factor $AF$, the ratio of reference experiments to active-learning experiments needed to hit a performance target, and enhancement factor $EF$, the ratio of best performances at a fixed experiment count—capture the benefit of self-driving labs. Across 42 studies and 63 benchmarks, the median $AF$ is 6, and $AF$ increases with dimensionality, meaning active learning manages high-dimensional search better than random sampling does. Reported $EF$ values vary from roughly 1.1 to 23, yet the peak of $EF$ consistently occurs around 10–20 experiments per dimension, after which random sampling catches up. Simulated campaigns wi

Load-bearing premise

The pooled analysis treats AF and EF values from experimental, retrospective, and computational benchmarks as comparable even though they use different reference strategies (random, grid, Latin hypercube, human, and algorithmic) and different objective spaces; if reference choice or benchmark type is correlated with dimensionality, the reported blessing of dimensionality could be an artifact of that pooling.

Editorial extensions

If this is right

  • A campaign designer can expect an active-learning SDL to reach a given target in roughly one-sixth the experiments of random sampling, with larger savings as dimensionality grows.
  • The 10–20 experiments-per-dimension peak means the largest benefit of algorithm-guided search arrives early; beyond roughly $20d$ experiments, random sampling narrows the gap.
  • Reported enhancement factors cannot be compared across studies without accounting for the contrast of the property space; a high $EF$ may indicate a high-contrast objective rather than a superior algorithm.
  • Reducing experimental noise matters most in complex, high-$L$ spaces, where noise inflates the experiment count required for convergence.
  • Adopting $AF$ and $EF$ relative to a standard random-sampling baseline would make future SDL benchmarks directly comparable and would test the generality of the median-6 result.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the authors leave implicit: for any new optimization campaign, estimate $C$ and $L$ from a small retrospective dataset before launching, and predict the campaign's $AF$/$EF$ envelope; if predictions hold across domains, the framework becomes a planning tool rather than a post-hoc metric.
  • The dimensionality trend could be confounded by the choice of reference: low-dimensional studies more often compare against human or grid baselines, while high-dimensional studies use random sampling. Reanalyzing with matched baselines would separate algorithmic benefit from benchmark artifacts.
  • If the blessing-of-dimensionality result generalizes, it implies self-driving labs should be aimed at high-dimensional formulation and process spaces, where the ratio of acceleration is largest and enumeration by humans is least feasible.
  • The 10–20 experiments-per-dimension peak suggests a budgeting heuristic: allocate roughly an order of magnitude per dimension for the active-learning advantage, and expect diminishing returns afterward; this could be tested prospectively by running multiple SDL campaigns with different budgets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This manuscript reviews quantitative benchmarking of self-driving laboratories. It defines the acceleration factor AF (Eq. 2) and enhancement factor EF (Eq. 3), surveys 42 SDL benchmarking studies reporting 63 benchmarks, extracts 33 AF values (Table 1) and EF trajectories from experimental/retrospective studies with random baselines (Fig. 4), and reports a median AF of 6, a positive dependence of AF on parameter-space dimensionality ('blessing of dimensionality'), and a consistent EF peak at 10–20 experiments per dimension. To interpret these observations, the authors simulate Bayesian optimization campaigns in 2D Gaussian-peak landscapes, showing that max(EF) increases with the contrast C of the property space, that the experiment number at max(EF) increases with Lipschitz complexity L, and that noise changes this optimum experiment number. Extracted data and simulation code are provided.

Significance. If the quantitative claims survive scrutiny, the paper would provide the SDL community with simple, useful guidelines (median AF ≈ 6; EF peaks near 10–20 n/d) and a common vocabulary for reporting acceleration. Strengths include a carefully curated and openly available dataset, clear definitions of AF and EF, reproducible simulation code, and falsifiable predictions about how EF depends on contrast and complexity. The main risk is that the literature-based AF/EF values are pooled across heterogeneous reference strategies, benchmark types, and threshold choices without statistical controls, so the headline trends are not yet established at the level claimed.

major comments (4)
  1. [§3.3, Fig. 3, Table 1] The central AF-vs-dimensionality claim pools AF values computed against different reference campaigns. Since Eq. (2) normalizes by the reference campaign's sample count, AF magnitude depends directly on the reference strategy's quality and budget. Table 1 mixes random (cases 8, 11, 14–16, 18, 21–30), LHS (1, 17), grid (2, 4, 5, 13), human (10, 33), and algorithmic (6, 12, 19, 20, 31, 32) references; case 5 even uses 'best BO performance within a time budget' as the reference. The high-AF cases 1 (37.5), 5 (56.25), and 17 (61) use non-random baselines, while random-baseline AF values span 1.3–100 (e.g., cases 8, 21, 26). Without stratifying by reference type or including it as a covariate, the reported 'blessing of dimensionality' could be an artifact of reference choice. This is load-bearing because the abstract and conclusions present the AF-d trend as a key finding.
  2. [§3.3, Fig. 3] The positive AF-d trend is asserted from 33 points without any correlation coefficient, confidence interval, or significance test, and the individual AF estimates have no error bars despite being derived from single or few campaigns. Please add at least a rank correlation with uncertainty, a sensitivity analysis excluding non-random and algorithmic references, and a clear description of how each AF was extracted from the original papers (which target y_AF, which n_ref/n_AL points, how horse-race plots were digitized). Without this, the statement that AF 'tends to increase with d' is not quantitatively supported.
  3. [§3.3, Fig. 4] The claim that EF 'consistently peaks at 10–20 experiments per dimension' is not backed by a quantitative analysis of Fig. 4. The figure shows many trajectories, but the paper does not state the rule used to identify a peak for each study, the distribution of peak n/d values, or how many trajectories actually exhibit a peak within the plotted range. Please provide per-study peak locations and a measure of central tendency and spread. This is needed because the EF peak is one of the two headline conclusions.
  4. [§4, Figs. 5–6; Conclusions] The abstract states that the simulations reveal 'how EF depends upon the statistical properties of the parameter space while AF depends on its complexity,' and the conclusions state that 'noise affects AF more than EF.' However, no AF is computed in the simulations; only EF and n* (the experiment number at max EF) are reported. Since AF = n_ref/n_AL for a chosen target, n* is only a proxy, and a direct AF computation (or an explicit mapping) is needed before making claims about AF dependence on complexity and noise. Please either compute AF in the simulated campaigns or soften the claims to be explicitly about n*.
minor comments (5)
  1. [§2, Eq. (1)] Eq. (1) is described as the 'average progress' or 'expected best observed response' of random sampling, but it actually defines the median of the best-of-n sample. These coincide only in special cases. Please state explicitly that the theoretical curves are median curves (or derive the expectation version), since Eq. (4) also relies on median(y).
  2. [Table 1] The 'Comparison' column mixes fundamentally different references. Please add a column or caption note categorizing the reference strategy as random/LHS/grid/human/algorithmic, and clarify unusual cases such as case 5 ('best BO performance within a time budget') and case 10 ('Random search vs. human'), where the direction of AF is ambiguous.
  3. [§3.3] The paper reports 42 unique studies and 63 benchmarks, but Table 1 lists only 33 AF cases. Please clarify the selection criteria that reduce 63 benchmarks to 33 AF values and how the remaining cases are handled (e.g., multiple benchmarks per study).
  4. [References] Reference 47 is incomplete (no title or journal), and several equations/text fragments contain rendering artifacts (e.g., '𝑥⃗∗ =a r g m a x', missing subscripts in Eqs. 2–3). Please run a final formatting pass.
  5. [§4] The simulations use only two-dimensional Gaussian-peak families. The authors appropriately describe these as exploratory, but the generalization to higher-dimensional materials spaces should be explicitly caveated in the conclusions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claims are literature-based empirical summaries and its simulations are illustrative, not fitted predictions disguised as discoveries.

full rationale

The central claims (median AF = 6, AF tending to increase with dimensionality, EF peaking at 10–20 experiments per dimension) are extracted from a corpus of 42 externally reported benchmarking studies, not derived from fitted parameters in this paper. The definitions of AF and EF in Eqs. (2) and (3) are explicit and standard, and Eq. (1) is a textbook order-statistics result for random sampling. The simulations in Sec. 4 are exploratory: they show how max(EF) varies with contrast C and Lipschitz complexity L, but these are not used to 'predict' the literature values; rather, the literature values motivate the simulations. Some prior papers by the same group are included in the surveyed data, but the aggregate statistics do not reduce to those papers alone and the qualitative conclusions would not be forced by their inclusion. The legitimate concern that pooling AF values across different reference baselines (random, LHS, grid, human, algorithmic) may confound the dimensionality trend is a validity/commensurability limitation, not circularity: it does not make any reported claim equivalent to its input by construction. Therefore no circular step meets the evidentiary standard required by the review rules.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The main claims are empirical aggregations of published benchmark values supported by simple simulations. The fits in Section 4 are exploratory; the more fragile assumptions are the cross-study comparability of AF/EF and the representativeness of Gaussian landscapes.

free parameters (4)
  • a (log-C fit coefficient)
    Fitted in Fig. 5D to max(EF) = a*log(C)+b across five simulated Gaussian-peak landscapes to illustrate the dependence of maximum enhancement on contrast.
  • b (log-C fit intercept)
    Intercept of the same Fig. 5D fit.
  • slope of n* vs L fit
    Fitted in Fig. 6C to quantify how the experiment number at peak EF grows with Lipschitz complexity.
  • intercept of n* vs L fit
    Intercept of the same Fig. 6C fit.
assumptions (4)
  • standard math The expected best observed value after n uniform random samples satisfies (F_y(y_n))^n = 1/2 (Eq. 1).
    This closed-form order-statistics result is used in Section 2 to define random-sampling reference convergence and in simulations as theoretical curves.
  • domain assumption Reported AF and EF values across the surveyed literature are directly comparable despite differences in benchmark type, reference strategy, objective function, and noise.
    Pooling in Section 3.3 (Table 1, Fig. 3) assumes cross-study comparability; the pooled median AF and the AF-vs-dimensionality trend depend on it.
  • domain assumption Two-dimensional Gaussian-peak functions with controlled contrast and Lipschitz complexity are representative proxies for real materials optimization landscapes.
    Section 4 uses these landscapes to draw conclusions about EF dependence on C and L and to compare n* with the literature peak at 10 to 20 experiments per dimension.
  • domain assumption The BoTorch package's default Gaussian process model and acquisition settings are typical of modern BO implementations.
    All simulated campaigns in Section 4 rely on BoTorch defaults, with no sensitivity analysis to alternative kernels or acquisition functions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Self-Driving Labs." pith.science (2026). https://pith.science/paper/55HJOVF4

@misc{pith2026250806642,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Self-Driving Labs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55HJOVF4}},
  note         = {Machine review of arXiv:2508.06642}
}
read the original abstract

A key goal of modern materials science is accelerating the pace of materials discovery. Self-driving labs, or systems that select experiments using machine learning and then execute them using automation, are designed to fulfil this promise by performing experiments faster, more intelligently, more reliably, and with richer metadata than conventional means. This review summarizes progress in understanding the degree to which SDLs accelerate learning by quantifying how much they reduce the number of experiments required for a given goal. The review begins by summarizing the theory underlying two key metrics, namely acceleration factor AF and enhancement factor EF, which quantify how much faster and better an algorithm is relative to a reference strategy. Next, we provide a comprehensive review of the literature, which reveals a wide range of AFs with a median of 6, and that tends to increase with the dimensionality of the space, reflecting an interesting blessing of dimensionality. In contrast, reported EF values vary by over two orders of magnitude, although they consistently peak at 10-20 experiments per dimension. To understand these results, we perform a series of simulated Bayesian optimization campaigns that reveal how EF depends upon the statistical properties of the parameter space while AF depends on its complexity. Collectively, these results reinforce the motivation for using SDLs by revealing their value across a wide range of material parameter spaces and provide a common language for quantifying and understanding this acceleration.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 61 canonical work pages

  1. [1]

    J. J. de Pablo, N. E. Jackson, M. A. Webb, L.-Q. Chen, J. E. Moore, D. Morgan, R. Jacobs, T. Pollock, D. G. Schlom and E. S. Toberer, Npj Comput Mater, 2019, 5, 1-23

  2. [2]

    R. B. Canty, J. A. Bennett, K. A. Brown, T. B uonassisi, S. V . Kalinin, J. R. Kitchin, B. Maruyama, R. G. Moore, J. Schrier, M. Seifrid, S. Sun, T. Vegge and M. Abolhasani, Nat Commun, 2025, 16, 3856

  3. [3]

    Vescovi, T

    R. Vescovi, T. Ginsburg, K. Hippe, D. Ozgulbas, C. Stone, A. Stroka, R. Butler, B. Blaiszik, T. Brettin, K. Chard, M. Hereld, A. Ramanathan, R. Stevens, A. Vriza, J. Xu, Q. Zhang and I. Foster, Digit Discov, 2023, 2, 1980-1998

  4. [4]

    Stach, B

    E. Stach, B. DeCost, A. G. Kusne, J. Hattrick-Simp ers, K. A. Brown, K. G. Reyes, J. Schrier, S. Billinge, T. Buonassisi, I. Foster, C. P. Gomes, J. M. Gregoire, A. Mehta, J. Montoya, E. Olivetti, C. Park, E. Rotenberg, S. K. Saikin, S. Smullin, V . Stanev and B. Maruyama, Matter, 2021, 4, 2702–2726

  5. [5]

    G. Tom, S. P. Schmid, S. G. Baird, Y . Cao, K. Darvish, H. Hao, S. Lo, S. Pablo-García, E. M. Rajaonson, M. Skreta, N. Yoshikawa, S. Corapi, G. D. Akkoc, F. Strieth-Kalthoff, M. Seifrid and A. Aspuru-Guzik, Chemical Reviews, 2024, 124, 9633-9732

  6. [6]

    E. J. Kluender, J. L. Hedrick, K. A. Brown, R. Rao, B. Meckes, J. Du, L. Moreau, B. Maruyama and C. A. Mirkin, Proceedings of the National Academy of Sciences of the United States of America, 2019, 116 40–45

  7. [7]

    J. Zhou, M. Luo, L. Chen, Q. Zhu, S. Jiang, F. Zhang, W. Shang and J. Jiang, Digit Discov, 2025, 4, 636-652

  8. [8]

    M. Zaki, C. Prinz and B. Ruehle, Acs Nano, 2025, DOI: 10.1021/acsnano.4c17504

Show all 63 references
  1. [9]

    Wang, Y .-J

    C. Wang, Y .-J. Kim, A. Vriza, R. Batra, A. Bask aran, N. Shan, N. Li, P. Darancet, L. Ward, Y . Liu, M. K. Y . Chan, S. K. R. S. Sankaranarayanan, H. C. Fry, C. S. Miller, H. Chan and J. Xu, Nat Commun, 2025, 16, 1498

  2. [10]

    T. Song, M. Luo, X. Zhang, L. Chen, Y . Huang, J. Cao, Q. Zhu, D. Liu, B. Zhang, G. Zou, G. Zhang, F. Zhang, W. Shang, Y . Fu, J. Jiang and Y . Luo, Journal of the American Chemical Society, 2025, 147, 12534-12545

  3. [11]

    Sanin, J

    A. Sanin, J. K. Flowers, T. H. Piotrowiak, F. Felsen, L. Merker, A. Ludwig, D. Bresser and H. S. Stein, Adv Energy Mater, 2025, 15, 2404961

  4. [12]

    S. Putz, J. Döttling, T. Ballweg, A. Tschöpe, V . Biniyaminov and M. Franzreb, Advanced Intelligent Systems, 2025, 7, 2400564

  5. [13]

    Nishio, A

    K. Nishio, A. Aiba, K. Takihara, Y . Suzuki, R. Nakayama, S. Kobayashi, A. Abe, H. Baba, S. Katagiri, K. Omoto, K. Ito, R. Shimizu and T. Hitosugi, Digit Discov, 2025, 4, 1734-1742

  6. [14]

    Strieth-Kalthoff, H

    F. Strieth-Kalthoff, H. Hao, V . Rathore, J. Derasp, T. Gaudin, N. H. Angello, M. Seifrid, E. Trushina, M. Guy, J. Liu, X. Tang, M. Mamada, W. Wang, T. Tsagaantsooj, C. Lavigne, R. Pollice, T. C. Wu, K. Hotta, L. Bodo, S. Li, M. Haddadnia, A. Wołos, R. Roszak, C. T. Ser, C. Bo...

  7. [15]

    K. L. Snapp, B. Verdier, A. E. Gongora, S. Silv erman, A. D. Adesiji, E. F. Morgan, T. J. Lawton, E. Whiting and K. A. Brown, Nat Commun, 2024, 15, 4290

  8. [16]

    Matsuda, G

    S. Matsuda, G. Lambard and K. Sodeyama, Cell Reports Physical Science, 2022, 3, 100832

  9. [17]

    Delgado-Licona and M

    F. Delgado-Licona and M. Abolhasani, Advanced Intelligent Systems, 2023, 5, 2200331

  10. [18]

    L. Hung, J. A. Yager, D. Monteverde, D. Baiocchi, H.-K. Kwon, S. Sun and S. Suram, Digit Discov, 2024, 3, 1273-1279

  11. [19]

    B. Rohr, H. S. Stein, D. Guevarra, Y . Wang, J. A. Haber, M. Aykol, S. K. Suram and J. M. Gregoire, Chem. Sci., 2020, 11, 2696-2706. 18

  12. [20]

    Q. H. Liang, A. E. Gongora, Z. K. Ren, A. Tiihonen, Z. Liu, S. J. Sun, J. R. Deneault, D. Bash, F. Mekki-Berrada, S. A. Khan, K. Hippalgaonkar, B. Maruyama, K. A. Brown, J. I. I. I. Fisher and T. Buonassisi, Npj Comput Mater, 2021, 7

  13. [21]

    Ament, M

    S. Ament, M. Amsler, D. R. Sutherland, M. C. Chang, D. Guevarra, A. B. Connolly, J. M. Gregoire, M. O. Thompson, C. P. Gomes and R. B. van Dover, Sci Adv, 2021, 7

  14. [22]

    Annevelink, R

    E. Annevelink, R. Kurchin, E. Muckley, L. Kavalsky, V . I. Hegde, V . Sulzer, S. Zhu, J. K. Pu, D. Farina, M. Johnson, D. Gandhi, A. Dave, H. Y . Lin, A. Edelman, B. Ramsundar, J. Saal, C. Rackauckas, V . R. Shah, B. Meredig and V . Viswanathan, Mrs Bull, 2022, 47, 1036-1044

  15. [23]

    A. Dave, J. Mitchell, S. Burke, H. Y . Lin, J. Whitacre and V . Viswanathan, Nat Commun, 2022, 13

  16. [24]

    Thelen, M

    A. Thelen, M. Zohair, J. Ramamurthy, A. Harkaway, W. M. Jiao, M. Ojha, M. Ul Ishtiaque, T. A. Kingston, C. L. Pint and C. Hu, J Power Sources, 2023, 581

  17. [25]

    Fatehi, M

    E. Fatehi, M. Thadani, G. Birsan and R. W. Black, arXiv preprint arXiv:2305.12541, 2023

  18. [26]

    C. K. H. Borg, E. S. Muckley, C. Nyby, J. E. Saal, L. Ward, A. Mehta and B. Meredig, Digit Discov, 2023, 2, 327-338

  19. [27]

    Honarmandi, V

    P. Honarmandi, V . Attari and R. Arroyave, Comp Mater Sci, 2022, 210

  20. [28]

    Lookman, P

    T. Lookman, P. V . Balachandran, D. Z. Xue and R. H. Yuan, Npj Comput Mater, 2019, 5

  21. [29]

    A. K. Y . Low, E. Vissol-Gaudin, Y . F. Lim and K. Hippalgaonkar, J Mater Inform, 2023, 3

  22. [30]

    D. N. Cakan, E. Oberholtz, K. Kausha l, S. P. Dunfield and D. P. Fenning, Mater Adv, 2025, 6, 598-606

  23. [31]

    Langner, F

    S. Langner, F. Häse, J. D. Perea, T. Stubhan, J. Hauch, L. M. Roch, T. Heumueller, A. Aspuru- Guzik and C. J. Brabec, Adv Mater, 2020, 32

  24. [32]

    Z. Liu, N. Rolston, A. C. Flick, T. W. Col burn, Z. K. Ren, R. H. Dauskardt and T. Buonassisi, Joule, 2022, 6, 834-849

  25. [33]

    H. Ros, M. Cook and D. Shorthouse, 2024

  26. [34]

    R. K. Vasudevan, K. P. Kelley, J. Hinkle, H. Funakubo, S. Jesse, S. V . Kalinin and M. Ziatdinov, Acs Nano, 2021, 15, 11253-11262

  27. [35]

    A. E. Gongora, B. W. Xu, W. Perry, C. Okoye, P. Riley, K. G. Reyes, E. F. Morgan and K. A. Brown, Sci Adv, 2020, 6

  28. [36]

    A. E. Gongora, K. L. Snapp, E. Whiting, P. Riley, K. G. Reyes, E. F. Morgan and K. A. Brown, Iscience, 2021, 24

  29. [37]

    G. Y . Li and X. N. Jin, Ieee Int Con Auto Sc, 2024, DOI: 10.1109/Case59546.2024.10711759, 2955-2960

  30. [38]

    M. J. Zhu, A. Mroz, L. F. Gui, K. E. Jelfs, A. Bemporad, E. A. D. Chanona and Y . S. Lee, Digit Discov, 2024, 3 , 2589-2606

  31. [39]

    Ziomek, M

    J. Ziomek, M. Adachi and M. A. Osborne, arXiv preprint arXiv:2410.10384, 2024

  32. [40]

    Bateni, S

    F. Bateni, S. Sadeghi, N. Orouji, J. A. Bennett, V . S. Punati, C. Stark, J. Y . Wang, M. C. Rosko, O. Chen, F. N. Castellano, K. G. Reyes and M. Abolhasani, Adv Energy Mater, 2024, 14

  33. [41]

    R. W. Epps, M. S. Bowen, A. A. V olk, K. Abde l-Latif, S. Y . Han, K. G. Reyes, A. Amassian and M. Abolhasani, Adv Mater, 2020, 32

  34. [42]

    Y . B. Jiang, D. Salley, A. Sharma , G. Keenan, M. Mullin and L. Cronin, Sci Adv, 2022, 8

  35. [43]

    A. A. V olk, R. W. Epps, D. T. Yonemoto, B. S. Masters, F. N. Castellano, K. G. Reyes and M. Abolhasani, Nat Commun, 2023, 14

  36. [44]

    T. Y . Wu, S. Kheiri, R. J. Hickman, H. C. Tao, T. C. Wu, Z. B. Yang, X. Ge, W. Zhang, M. Abolhasani, K. Liu, A. Aspuru-Guzik and E. Kumacheva, Nat Commun, 2025, 16

  37. [45]

    Sadeghi, F

    S. Sadeghi, F. Bateni, T. Kim, D. Y . Son, J. A. Bennett, N. Orouji, V . S. Punati, C. Stark, T. D. Cerra, R. Awad, F. Delgado-Licona, J. E. Xu, N. Mukhin, H. Dickerson, K. G. Reyes and M. Abolhasani, Nanoscale, 2024, 16, 580-591

  38. [46]

    B. P. MacLeod, F. G. L. Parlane, C. C. R upnow, K. E. Dettelbach, M. S. Elliott, T. D. Morrissey, T. H. Haley, O. Proskurin, M. B. Rooney, N. Taherimakhsousi, D. J. Dvorak, H. N. Chiu, C. E. B. Waizenegger, K. Ocean, M. Mokhtari and C. P. Berlinguette, Nat Commun, 2022, 13. 19

  39. [47]

    Tamura, H

    C. Tamura, H. Job, H. Chang, W. Wang, Y . Liang and S. Sun, 2025

  40. [48]

    Nakayama, R

    R. Nakayama, R. Shimizu, T. Haga, T. Kimura , Y . Ando, S. Kobayashi, N. Yasuo, M. Sekijima and T. Hitosugi, Sci Tech Adv Mat-Met, 2022, 2, 119-128

  41. [49]

    Guay-Hottin, L

    R. Guay-Hottin, L. Kardassevitch, H. Pham, G. Lajoie and M. Bonizzato, Knowl-Based Syst, 2025, 311

  42. [50]

    Duros, J

    V . Duros, J. Grizou, W. M. Xuan, Z. Hosni, D. L. Long, H. N. Miras and L. Cronin, Angew Chem Int Edit, 2017, 56, 10815-10820

  43. [51]

    Grizou, L

    J. Grizou, L. J. Points, A. Sharma and L. Cronin, Sci Adv, 2020, 6

  44. [52]

    Y . Bai, Z. H. J. Khoo, R. Made, H. Q. Xie, C. Y . J. Lim, A. D. Handoko, V . Chellappan, J. J. Cheng, F. X. Wei, Y . F. Lim and K. Hippalgaonkar, Adv Mater, 2024, 36

  45. [53]

    Kavalsky, V

    L. Kavalsky, V . I. Hegde, E. Muckley, M. S. Johnson, B. Meredig and V . Viswanathan, Digit Discov, 2023, 2, 1112-1125

  46. [54]

    Liang, A

    Q. Liang, A. E. Gongora, Z. Ren, A. Tiihonen, Z. Liu, S. Sun, J. R. Deneault, D. Bash, F. Mekki- Berrada, S. A. Khan, K. Hippalgaonkar, B. Maruyama, K. A. Brown, J. Fisher Iii and T. Buonassisi, Npj Comput Mater, 2021, 7, 188

  47. [55]

    Conrad, M

    F. Conrad, M. Mälzer, M. Schwarzenberger, H. Wiemer and S. Ihlenfeldt, Sci Rep-Uk, 2022, 12

  48. [56]

    B. J. Shields, J. Stevens, J. Li, M. Parasram, F. Damani, J. I. M. Alvarado, J. M. Janey, R. P. Adams and A. G. Doyle, Nature, 2021, 590, 89-96

  49. [57]

    B. W. Lei, T. Q. Kirk, A. Bhattacharya, D. Pati, X. N. Qian, R. Arroyave and B. K. Mallick, Npj Comput Mater, 2021, 7

  50. [58]

    D. A. Cohn, Z. Ghahramani and M. I. Jordan, J Artif Intell Res, 1996, 4, 129-145

  51. [59]

    Berkenkamp, A

    F. Berkenkamp, A. P. Schoellig and A. Krause, Journal of Machine Learning Research, 2019, 20, 1-24

  52. [60]

    D. C. Liu and J. Nocedal, Mathematical programming, 1989, 45, 503-528

  53. [61]

    M. D. Hoffman and A. Gelman, J. Mach. Learn. Res., 2014, 15, 1593-1623

  54. [62]

    Mekki-Berrada, Z

    F. Mekki-Berrada, Z. Ren, T. Huang, W. K. Wong, F. Zheng, J. Xie, I. P. S. Tian, S. Jayavelu, Z. Mahfoud and D. Bash, Npj Comput Mater, 2021, 7, 55

  55. [63]

    G. R. Wood and B. Zhang, Journal of Global Optimization, 1996, 8, 91-103

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.