Pith. sign in

REVIEW 4 major objections 6 minor 46 references

Extended Factorization Machine Annealing for Rapid Discovery of Transparent Conducting Materials

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an extended factorization-machine annealer, using bit-string encoding, Hopfield memory, random flips, and local search, finds better (Al,Ga,In)2O3 candidates faster than Bayesian optimization and genetic algorithms.

desk verdict The core empirical claim is unverifiable because the paper never defines its 'score,' but the method write-up is coherent and the heuristics merit a proper look. read the letter →

arxiv 2507.23160 v1 pith:SLERG7VV submitted 2025-07-30 cond-mat.mtrl-sci cs.LG

Daisuke Makino , Tatsuya Goto , Yoshinori Suga This is my paper
classification cond-mat.mtrl-scics.LG
keywords transparentconductingoxides(AlGaIn)2O3factorizationmachineannealingcombinatorialoptimizationQUBOmaterialsinformaticsmulti-objectiveNomad2018
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

The paper extends a Factorization Machine Annealing (FMA) optimizer to search continuous composition and lattice parameters of the transparent-conductor oxide (Al,Ga,In)2O3, encoding the continuous variables as bit strings and annealing over the resulting binary space. Its central claim is that this pipeline finds higher-value band-gap candidates faster than Bayesian optimization (TPE) and a genetic algorithm (NSGA-II): in a 60-second comparison the extended FMA reached 123.5, while the two baselines reached 100.5 and 76.2. The practical stakes are cost and speed in inverse materials design; with any fast learned property model, the optimizer is intended to propose promising compositions and structures in minutes rather than through exhaustive first-principles screening. The paper also demonstrates a two-objective search that drives predicted band gaps toward 4.0 eV while keeping formation energy low, and shows the same search components improve the MaxCut benchmark.

What carries the argument

The load-bearing object is the binary energy function $$E(s) = \$\alpha$ \hat{y}(x(s)) + \$\beta$ \Psi_{\mathrm{lagrange}}(x(s)) + \gamma \Psi_{\mathrm{hopfield}}(s),$$ minimized over a bit string $s$ that encodes composition ratios and lattice parameters. The Factorization Machine supplies a compact surrogate for the band gap or formation energy of the decoded continuous variables; simulated annealing minimizes the energy; a Hopfield network stores previously found good bit patterns and penalizes distance from them; adaptive random flips increase exploration when progress stalls; and a final local bit-flip search refines the QUBO optimum against the true surrogate. The extended-Lagrange term enforces the composition constraint $x+y+z=1$.

What would settle it

Recalculate the top optimized compositions with density-functional theory: if the predicted band gaps and formation energies deviate substantially, or if repeating the 60-second comparison with DFT as the objective no longer puts extended FMA ahead of the baselines, the claimed advantage is an artifact of the surrogate rather than of the materials.

Watch

Extended reading notes

Core claim

The central claim is that continuous-parameter materials search can be cast as a binary optimization problem and solved more efficiently with an extended factorization-machine annealer than with standard black-box optimizers. The authors binarize composition ratios and lattice parameters, fit a Factorization Machine to predict the property of interest, and anneal over the bit string with an energy that combines the predicted property, a Lagrange penalty for the $x+y+z=1$ constraint, and a Hopfield memory term that pulls the search toward previously good solutions. Adaptive random flips supply global exploration, and a bit-string local search refines the final candidate. On a band-gap maximization task over the Nomad2018 (Al,Ga,In)2O3 dataset, the method reached 123.5 in 60 seconds, ahead of NSGA-II at 100.5 and TPE at 76.2; the authors further find that the Hopfield penalty strength near $\lambda=0.5$, an 8-bit representation, and binary encoding over Gray code each give the best performance in their tests.

Load-bearing premise

The search is only as trustworthy as the machine-learning property predictor used as its oracle, and the paper relies on a model trained on the Nomad2018 data without testing any optimized candidate against density-functional theory or experiment.

Editorial extensions

If this is right

  • In time-limited screening, extended FMA should reach a given best-value target in roughly 5 to 10 seconds where TPE and NSGA-II need 60 seconds, cutting the cost of surrogate-based materials search.
  • Multi-objective targets such as a band gap near 4.0 eV with low formation energy can be produced even when no such combination is present in the training data, extending inverse design outside the observed distribution.
  • The Hopfield-penalty, random-flip, and local-search components transfer to non-materials combinatorial problems such as MaxCut, so the framework is a general search tool rather than a transparent-conductor-specific fix.
  • Search quality is sensitive to discretization choices: 8-bit binary encoding outperforms 4-, 12-, and 16-bit variants and Gray code, so practical deployments should treat bit width and penalty strength as first-order tuning parameters.

Reading between the lines

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

  • Because no optimized candidate is checked against density-functional theory or experiment, the paper's 'discovery' of $E_g\approx 4.0$ eV compositions is strictly a claim about a machine-learned surrogate; its material-level significance will only be settled by recalculating those compositions with first-principles methods.
  • The fixed 60-second wall-clock comparison favors methods that warm-start quickly; on larger search spaces or with more expensive surrogates, the Factorization Machine's $O(kn)$ evaluations and bit-flip local search may scale differently, so the ranking needs re-testing rather than automatic extrapolation.
  • The Hopfield memory could be made more powerful by storing a diverse archive of high-scoring bit patterns instead of a single best solution, which would test whether the $\lambda\approx 0.5$ optimum comes from memory diversity rather than anchoring strength alone.
  • A direct testable extension is to replace the trained surrogate with a density-functional-theory objective or an active-learning loop and rerun the 60-second comparison, separating algorithmic gains from oracle error.
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 / 6 minor

Summary. The manuscript proposes an extension of Factorization Machine Annealing (FMA) for inverse design of (Al_x Ga_y In_z)_2O_3 transparent conductors, using an ElemNet model trained on the Kaggle Nomad2018 data as a black-box objective. The proposed extensions are binarization of continuous variables, a Hopfield penalty for exploitation, adaptive random flips for exploration, and a bit-string local search. The central empirical claim is that extended FMA outperforms TPE and NSGA-II within a 60-second budget (Section 4.2), with reported scores of 123.5, 100.5, and 76.2, respectively. Additional experiments vary the Hopfield penalty coefficient, bit length, and encoding scheme, and apply the method to multi-objective bandgap/formation-energy search and to the MaxCut problem.

Significance. If the central claim were properly supported, the method could be a useful addition to the optimization toolbox for materials discovery, especially because FMA-based approaches are still uncommon in materials informatics. The paper also describes a clearly structured algorithmic extension and includes ablation-style comparisons for the Hopfield penalty, random flips, and local search. However, the current manuscript does not define the score, does not precisely specify the algorithm's implementation inside the time budget, and never validates any optimized candidate against DFT or experiment, so the significance of the reported results remains conditional and cannot yet be assessed as a discovery claim.

major comments (4)
  1. [Section 4.2, Fig. 1] The score reported as 123.5, 100.5, and 76.2 is never defined. Because the stated task is maximizing the bandgap of (Al,Ga,In)2O3, whose physical values are on the order of a few eV, 123.5 cannot be a raw bandgap; the transformation or aggregated objective must be stated explicitly. The comparison also omits the number of ElemNet evaluations per method, the hardware specification, and whether FM training or retraining is included in the 60-second wall-clock budget. Without this information, the central speed/accuracy comparison cannot be interpreted or reproduced.
  2. [Sections 3.4 and 4.1] The extended FMA algorithm is underspecified. Section 3.4 lists high-level steps, but it does not state how the Factorization Machine is trained, on what data, how often it is retrained, or how it interacts with the ElemNet black-box function during the 60-second optimization loop. The energy function in Section 3.3.4 includes alpha times y-hat(x(s)), but the role of the FM surrogate as opposed to direct black-box evaluation is unclear. As written, the method being benchmarked against TPE and NSGA-II is not precisely instantiated.
  3. [Section 4.6, Fig. 5] The green stars presented as newly discovered candidates are predictions of an ElemNet model and are never checked against DFT or experiment. Because the optimizer explicitly extrapolates to property values not present in the training data (e.g., Eg = 4.0 eV), the claim of discovering materials with low formation energy and a bandgap near 4.0 eV is unsupported. At minimum, a first-principles validation of several optimized candidates is needed to substantiate the 'discovery' framing used in the title and abstract.
  4. [Sections 4.2-4.5 and Table 1] The performance comparisons report 'average scores' without stating the number of independent runs, the variance, or any statistical test. For example, Table 1 lists single numbers (120.78, 125.61, 114.97) at iteration 50, and Section 4.2 reports a 60-second comparison with no error bars. The claimed ranking of extended FMA over NSGA-II and TPE cannot be distinguished from noise without repeated trials and a measure of spread.
minor comments (6)
  1. [Figures 1 and 4] These figures lack axis labels and legend details, making it difficult for the reader to determine what is being plotted and how the curves were obtained.
  2. [Section 3.3.1] The term 'extended Lagrangian method' is used for a simple quadratic penalty of the form (x+y+z-1)^2; this is nonstandard terminology and should be either defined precisely or replaced with a more conventional description.
  3. [Section 4.1] The manuscript uses ElemNet without citing its original source; a reference for ElemNet should be added at first use.
  4. [Abstract and Section 5.1] The phrase 'faster and more accurate searches' appears in the abstract, but no definition of 'accurate' is given anywhere in the paper; the authors should state whether accuracy refers to the surrogate score, to agreement with DFT, or to some other measure.
  5. [Section 4.7] The MaxCut experiment does not specify the graph size, edge density, or number of vertices, which limits the usefulness of that ablation for assessing the generality of the proposed mechanisms.
  6. [Sections 3.1 and 3.3.4] The symbols alpha, beta, and gamma are used both for lattice angles in Section 3.1 and for energy weights in Section 3.3.4; this reuse is confusing and should be resolved by renaming one of the sets.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the optimizer comparison uses a shared ElemNet oracle and an explicit black-box rescoring step, so the reported ranking is not a restatement of a fitted quantity.

full rationale

The paper's derivation chain contains no step in which a claimed output is identical by construction to an input. The objective is defined in Section 3.1 as a black-box function given by ElemNet or Kaggle models, and all three compared optimizers are evaluated under the same oracle, making the relative ranking an algorithmic comparison rather than a circular fit. Extended FMA's flow in Section 3.4 includes decoding the best binary solution and re-evaluating it with the black-box function, so the reported best score is not merely the FM surrogate's training objective. The Eg = 4.0 eV candidates in Section 4.6 are predictions of the fitted ElemNet model, but the paper presents them as 'candidates' and does not claim DFT or experimental confirmation; overstating them as discoveries is a correctness and interpretability concern, not a circular reduction. No load-bearing self-citation chain appears: FMA is attributed to Kitai et al. and the local-search idea to Endo & Takahashi, but those citations are background and do not force the headline benchmark result. The serious reporting gaps—'score' is never defined, hardware and evaluation counts are unspecified, and the FMA instantiation is underspecified—undermine verifiability but do not make any equation equal to its own input, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the ElemNet surrogate being a faithful oracle, on the FM accurately approximating that surrogate, and on several hyperparameters (Hopfield weight, bit length, energy weights, flip probability) that are tuned or unspecified. No new physical entities are introduced.

free parameters (5)
  • Hopfield penalty coefficient lambda = 0.5 (selected as best among 0.0, 0.5, 5.0)
    The paper tests lambda values and picks 0.5 based on final score; this is a hyperparameter tuned to the benchmark.
  • Bit length for discretization = 8 bits (best among 4, 8, 12, 16)
    The bit length is chosen after observing performance; affects the granularity of the continuous-to-binary mapping.
  • Energy weights alpha, beta, gamma = not fully specified; gamma corresponds to Hopfield penalty
    The energy function E(s) = alpha*y + beta*Psi_lagrange + gamma*Psi_hopfield uses weights that must be set; the paper only reports varying gamma (Hopfield) and does not state alpha and beta values.
  • Adaptive random flip probability p = not specified
    The flip probability is adapted based on search stagnation but the schedule is not described, adding an undocumented free parameter.
  • FM regularization lambda = not specified
    The FM loss includes an L2 regularization term lambda, but no value or tuning procedure is given.
assumptions (4)
  • domain assumption The ElemNet model trained on Nomad2018 data provides a faithful surrogate for true bandgaps and formation energies of (Al,Ga,In)2O3.
    The paper treats ElemNet as the black-box oracle and never compares its predictions to DFT or experiment for the optimized compositions.
  • domain assumption An 8-bit binary encoding of continuous variables yields a search space whose optimum corresponds to the true continuous optimum.
    The paper selects 8 bits based on benchmark performance, but no error analysis shows that the discretization error is negligible.
  • domain assumption The factorization machine, trained on a finite sample of ElemNet evaluations, accurately approximates the black-box over the entire search space, including extrapolation regions.
    The search optimizes the FM's QUBO, and the paper acknowledges the FM does not precisely represent the original problem.
  • standard math Simulated annealing with the specified schedule and random flips reaches the global minimum of the QUBO with high probability.
    Simulated annealing is a heuristic; no guarantee or convergence analysis is provided, but this is standard practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extended Factorization Machine Annealing for Rapid Discovery of Transparent Conducting Materials." pith.science (2026). https://pith.science/paper/SLERG7VV

@misc{pith2026250723160,
  author       = {Pith},
  title        = {Pith review of: Extended Factorization Machine Annealing for Rapid Discovery of Transparent Conducting Materials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SLERG7VV}},
  note         = {Machine review of arXiv:2507.23160}
}
abstract

The development of novel transparent conducting materials (TCMs) is essential for enhancing the performance and reducing the cost of next-generation devices such as solar cells and displays. In this research, we focus on the (Al$_x$Ga$_y$In$_z$)$_2$O$_3$ system and extend the FMA framework, which combines a Factorization Machine (FM) and annealing, to search for optimal compositions and crystal structures with high accuracy and low cost. The proposed method introduces (i) the binarization of continuous variables, (ii) the utilization of good solutions using a Hopfield network, (iii) the activation of global search through adaptive random flips, and (iv) fine-tuning via a bit-string local search. Validation using the (Al$_x$Ga$_y$In$_z$)$_2$O$_3$ data from the Kaggle "Nomad2018 Predicting Transparent Conductors" competition demonstrated that our method achieves faster and more accurate searches than Bayesian optimization and genetic algorithms. Furthermore, its application to multi-objective optimization showed its capability in designing materials by simultaneously considering both the band gap and formation energy. These results suggest that applying our method to larger, more complex search problems and diverse material designs that reflect realistic experimental conditions is expected to contribute to the further advancement of materials informatics.

Figures

Figures reproduced from arXiv: 2507.23160 by the authors.

Figure 1
Figure 1. Comparison among FMA, NSGA-II, and TPE From these results, for the bandgap maximization problem in the (AlxGayInz)2O3 system, extended FMA shows superior search performance compared to the other two methods. In the following sections, we further investigate how this proposed method (ex￾tended FMA) achieves its results by examining (1) Hopfield penalty coefficients, (2) the number of bi￾nary variables, and (3) encodi… view at source ↗
Figure 2
Figure 2. Average score transitions at each iteration [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Differences in optimization performance by [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Example of performance transitions when varying the bit length (4, 8, 12, 16) 4.5 Comparison of Different Encoding Methods In this section, we examine how the choice of en￾coding method for converting continuous values to binary variables affects optimization performan…
Figure 5
Figure 5. Figure 5: Example of multi-objective optimization results. Red circles represent the Pareto front, green stars denote newly discovered candidates, and the gray dashed line indicates Eg = 4.0 eV. 4.7 Validating Random Flips and Lo￾cal Search in the MaxCut Prob￾lem The objective o…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 45 canonical work pages

  1. [1]

    Machine learning in materials informatics: re- cent applications and prospects

    R. Ramprasad, R. Batra, G. Pilania, et al. “Machine learning in materials informatics: re- cent applications and prospects”. npj Comput. Mater. (2017)

    work page 2017

  2. [2]

    A Review of the Progress of Thin- Film Transistors and Their Technologies for Flexible Electronics

    M. J. Mirshojaeian Hosseini and R. A. Nawrocki. “A Review of the Progress of Thin- Film Transistors and Their Technologies for Flexible Electronics”. Micromachines (2021)

    work page 2021

  3. [3]

    Ultrabright and stable top- emitting quantum-dot light-emitting diodes with negligible angular color shift

    Mengqi Li et al. “Ultrabright and stable top- emitting quantum-dot light-emitting diodes with negligible angular color shift”. Nat. Com- mun. (2024)

    work page 2024

  4. [4]

    Achievements, challenges, and future prospects for industrialization of perovskite solar cells

    Chuang Yang et al. “Achievements, challenges, and future prospects for industrialization of perovskite solar cells”. Light Sci. Appl. (2024). 10

    work page 2024

  5. [5]

    From Design to Device: Challenges and Opportunities in Com- putational Discovery of p-Type Transparent Conductors

    Rachel Woods-Robinson et al. “From Design to Device: Challenges and Opportunities in Com- putational Discovery of p-Type Transparent Conductors”. PRX Energy (2024)

    work page 2024

  6. [6]

    p- type electrical conduction in transparent thin films of CuAlO 2

    H. Kawazoe, M. Yasukawa, H. Hyodo, et al. “p- type electrical conduction in transparent thin films of CuAlO 2”. Nature (1997)

    work page 1997

  7. [7]

    AZO (Al:ZnO) thin films with high figure of merit as stable indium free transpar- ent conducting oxide

    Bikash Sarma, Dipak Barman, and Bimal K. Sarma. “AZO (Al:ZnO) thin films with high figure of merit as stable indium free transpar- ent conducting oxide”. Applied Surface Science (2019). issn: 0169-4332

    work page 2019

  8. [8]

    Crowd-sourcing materials-science chal- lenges with the NOMAD 2018 Kaggle compe- tition

    C. Sutton, L. M. Ghiringhelli, T. Yamamoto, et al. “Crowd-sourcing materials-science chal- lenges with the NOMAD 2018 Kaggle compe- tition”. npj Comput. Mater. (2019)

    work page 2019

Show all 46 references
  1. [9]

    Prop- erties of orthorhombic Ga2O3 alloyed with In2O3 and Al2O3

    S. Seacat, J. L. Lyons, and H. Peelaers. “Prop- erties of orthorhombic Ga2O3 alloyed with In2O3 and Al2O3”. Appl. Phys. Lett. (2021)

    2021

  2. [10]

    Projector augmented-wave method

    P. E. Bl¨ ochl. “Projector augmented-wave method”. Phys. Rev. B (1994)

    1994

  3. [11]

    Efficient itera- tive schemes for ab initio total-energy calcula- tions using a plane-wave basis set

    G. Kresse and J. Furthm¨ uller. “Efficient itera- tive schemes for ab initio total-energy calcula- tions using a plane-wave basis set.” Phys. Rev. B (1996)

    1996

  4. [12]

    Computational Screening of p- Type Transparent Conducting Oxides Using the Optical Absorption Spectra and Oxygen- Vacancy Formation Energies

    Yu Kumagai. “Computational Screening of p- Type Transparent Conducting Oxides Using the Optical Absorption Spectra and Oxygen- Vacancy Formation Energies”. Phys. Rev. Appl. (2023)

    2023

  5. [13]

    A general-purpose machine-learning frame- work for predicting properties of inorganic ma- terials

    L. Ward, A. Agrawal, A. Choudhary, et al. “A general-purpose machine-learning frame- work for predicting properties of inorganic ma- terials”. npj Comput. Mater. (2016)

    2016

  6. [14]

    A multi- fidelity machine learning approach to high throughput materials screening

    C. Fare, P. Fenner, M. Benatan, et al. “A multi- fidelity machine learning approach to high throughput materials screening”. npj Comput. Mater. (2022)

    2022

  7. [15]

    Identifying domains of applicability of ma- chine learning models for materials science

    C. Sutton, M. Boley, L. M. Ghiringhelli, et al. “Identifying domains of applicability of ma- chine learning models for materials science”. Nat. Commun. (2020)

    2020

  8. [16]

    Learning from data to design functional mate- rials without inversion symmetry

    P. Balachandran, J. Young, T. Lookman, et al. “Learning from data to design functional mate- rials without inversion symmetry”. Nat. Com- mun. (2017)

    2017

  9. [17]

    Accelerated discovery of metallic glasses through iteration of machine learning and high-throughput experiments

    Fang Ren et al. “Accelerated discovery of metallic glasses through iteration of machine learning and high-throughput experiments”. Sci. Adv. (2018)

    2018

  10. [18]

    Factorization Machines

    Steffen Rendle. “Factorization Machines”. 2010 IEEE International Conference on Data Min- ing. 2010

    2010

  11. [19]

    BPR: Bayesian person- alized ranking from implicit feedback

    Steffen Rendle et al. “BPR: Bayesian person- alized ranking from implicit feedback”. Pro- ceedings of the 25th Conference on Uncertainty in Artificial Intelligence . Arlington, Virginia, USA: AUAI Press, 2009

    2009

  12. [20]

    Optimization by Simulated Annealing

    Scott Kirkpatrick, Charles D. Gelatt, and Mario P. Vecchi. “Optimization by Simulated Annealing”. Science (1983)

    1983

  13. [21]

    Quantum an- nealing in the transverse Ising model

    T. Kadowaki and H. Nishimori. “Quantum an- nealing in the transverse Ising model”. Phys. Rev. E (1998)

    1998

  14. [22]

    Designing metamaterials with quantum annealing and factorization ma- chines

    Koki Kitai et al. “Designing metamaterials with quantum annealing and factorization ma- chines”. Phys. Rev. Res. (2020)

    2020

  15. [23]

    Quan- tum annealing-assisted lattice optimization

    Z. Xu, W. Shang, S. Kim, et al. “Quan- tum annealing-assisted lattice optimization”. npj Comput. Mater. (2025)

    2025

  16. [24]

    A super- conducting copper oxide compound with elec- trons as the charge carriers

    Y. Tokura, H. Takagi, and S. Uchida. “A super- conducting copper oxide compound with elec- trons as the charge carriers”. Nature (1989)

    1989

  17. [25]

    Transparent p-Type Conducting Oxides: De- sign and Fabrication of p-n Heterojunctions

    H. Kawazoe, H. Yanagi, K. Ueda, et al. “Transparent p-Type Conducting Oxides: De- sign and Fabrication of p-n Heterojunctions”. MRS Bull. (2000)

    2000

  18. [26]

    Transpar- ent conducting materials discovery using high- throughput computing

    G. Brunin, F. Ricci, V. A. Ha, et al. “Transpar- ent conducting materials discovery using high- throughput computing”. npj Comput. Mater. (2019)

    2019

  19. [27]

    Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Ma- terial Properties

    Tian Xie and Jeffrey C. Grossman. “Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Ma- terial Properties”. Phys. Rev. Lett. (2018)

    2018

  20. [28]

    Graph Networks as a Univer- sal Machine Learning Framework for Molecules and Crystals

    Chi Chen et al. “Graph Networks as a Univer- sal Machine Learning Framework for Molecules and Crystals”. Chem. Mater. (2019)

    2019

  21. [29]

    Pre- dicting materials properties without crystal structure: deep representation learning from stoichiometry

    Rhys E. A. Goodall and Alpha A. Lee. “Pre- dicting materials properties without crystal structure: deep representation learning from stoichiometry”. Nat. Commun. (2020)

    2020

  22. [30]

    Generative models for inverse design of inorganic solid materials

    L. Chen et al. “Generative models for inverse design of inorganic solid materials”. J. Mater. Inf. (2021)

    2021

  23. [31]

    An improved genetic algorithm for crystal structure predic- tion

    L. Chen, Y. Li, X. Zhang, et al. “An improved genetic algorithm for crystal structure predic- tion”. Curr. Appl. Phys. (2017)

    2017

  24. [32]

    Crystal structure pre- diction via particle-swarm optimization

    Yanchao Wang et al. “Crystal structure pre- diction via particle-swarm optimization”.Phys. Rev. B (2010). 11

    2010

  25. [33]

    Accelerated search for materials with targeted properties by adaptive design

    D. Xue et al. “Accelerated search for materials with targeted properties by adaptive design”. Nat. Commun. (2016)

    2016

  26. [34]

    Bayesian optimization of chemical composition: A comprehensive frame- work and its application to RFe12-type magnet compounds

    Taro Fukazawa et al. “Bayesian optimization of chemical composition: A comprehensive frame- work and its application to RFe12-type magnet compounds”. Phys. Rev. Mater. (2019)

    2019

  27. [35]

    Ultranarrow-Band Wavelength-Selective Thermal Emission with Aperiodic Multilayered Metamaterials De- signed by Bayesian Optimization

    A. Sakurai et al. “Ultranarrow-Band Wavelength-Selective Thermal Emission with Aperiodic Multilayered Metamaterials De- signed by Bayesian Optimization”. ACS Cent. Sci. (2019)

    2019

  28. [36]

    Deep Reinforcement Learning for Inverse Inorganic Materials Design

    Elton Pan, Christopher Karpovich, and Elsa A. Olivetti. “Deep Reinforcement Learning for Inverse Inorganic Materials Design”. arXiv preprint arXiv:2210.11931 (2022)

    2022 arXiv

  29. [37]

    Reinforcement learn- ing in crystal structure prediction

    Elena Zamaraeva et al. “Reinforcement learn- ing in crystal structure prediction”. Digital Discovery (2023)

    2023

  30. [38]

    Learning con- ditional policies for crystal design using of- fline reinforcement learning

    Prashant Govindarajan et al. “Learning con- ditional policies for crystal design using of- fline reinforcement learning”. Digital Discovery (2024)

    2024

  31. [39]

    Deep reinforcement learning for inverse inor- ganic materials design

    C. Karpovich, E. Pan, and E. A. Olivetti. “Deep reinforcement learning for inverse inor- ganic materials design”. npj Comput. Mater. (2024)

    2024

  32. [40]

    High-Performance Trans- parent Radiative Cooler Designed by Quantum Computing

    Seongmin Kim et al. “High-Performance Trans- parent Radiative Cooler Designed by Quantum Computing”. ACS Energy Lett. (2022)

    2022

  33. [41]

    Quan- tum annealing-aided design of an ultrathin- metamaterial optical diode

    S. Kim, S. J. Park, S. Moon, et al. “Quan- tum annealing-aided design of an ultrathin- metamaterial optical diode”. Nano Conver- gence (2024)

    2024

  34. [42]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. “Adam: A Method for Stochastic Optimization”. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  35. [43]

    Multiplier and gradient meth- ods

    M. R. Hestenes. “Multiplier and gradient meth- ods”. J. Optim. Theory Appl. (1969)

    1969

  36. [44]

    Neural networks and physi- cal systems with emergent collective computa- tional abilities

    John J. Hopfield. “Neural networks and physi- cal systems with emergent collective computa- tional abilities”. Proc. Natl. Acad. Sci. U.S.A. (1982)

    1982

  37. [45]

    Function Smoothing Regularization for Pre- cision Factorization Machine Annealing in Continuous Variable Optimization Problems

    Katsuhiro Endo and Kazuaki Z. Takahashi. “Function Smoothing Regularization for Pre- cision Factorization Machine Annealing in Continuous Variable Optimization Problems”. arXiv preprint arXiv:2407.04393 (2024)

    2024 arXiv

  38. [46]

    Optuna: A Next- generation Hyperparameter Optimization Framework

    Takuya Akiba et al. “Optuna: A Next- generation Hyperparameter Optimization Framework”. Proceedings of the 25th ACM SIGKDD International Conference on Knowl- edge Discovery & Data Mining (2019). 12

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.