Pith. sign in

REVIEW 5 major objections 6 minor 72 references

AIMatDesign: Knowledge-Augmented Reinforcement Learning for Inverse Materials Design under Data Scarcity

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read AIMatDesign is a reinforcement learning framework that claims data scarcity in inverse materials design can be offset by combining a difference-based Trustworthy Experience Pool, LLM-driven Automatic Model Refinement, and a…

desk verdict A useful RL+LLM framework with real experimental validation of yield-strength guidance, but the headline success rates and ductility claims rest on a surrogate that is systematically optimistic. read the letter →

arxiv 2507.00024 v1 pith:D3WS5F4N submitted 2025-06-17 cs.LG cond-mat.mtrl-scics.AI

Yeyong Yu , Xilei Bian , Jie Xiong , Xing Wu , Quan Qian This is my paper
classification cs.LGcond-mat.mtrl-scics.AI
keywords inversematerialsdesignreinforcementlearningbulkmetallicglassesdatascarcityaugmentationlargelanguagemodelsknowledge-basedrewardautomaticmodelrefinement
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

AIMatDesign is a reinforcement learning framework for inverse materials design that claims data scarcity can be offset by three ingredients: a difference-based Trustworthy Experience Pool built from pairs of real compositions, an LLM-driven Automatic Model Refinement that retrains the surrogate when its predictions look unreliable, and a Knowledge-Based Reward that injects expert rules into the reward signal. The paper argues that these mechanisms let an RL agent explore a high-dimensional composition space efficiently, generating candidates that satisfy bulk metallic glass constraints and hit multiple performance targets. Against grid search, NSGA-II, and standard RL baselines, the framework reports the highest success rates at the same number of surrogate evaluations, and its proposed Zr-based alloys were experimentally cast, yielding a top composition with 1.7 GPa yield strength and 10.2% elongation. The measured yield strengths across eight compositions average 4.9% relative error versus predictions, while elongations were systematically overpredicted because process conditions were not modelled. The central claim matters because it offers a path to closed-loop materials discovery when experimental data number only in the hundreds.

What carries the argument

The framework's load-bearing mechanism is a hierarchical reward built on two surrogates, a CatBoost classifier for BMG-likeness and an edRVFL regressor for seven properties, plus three corrective loops: the Trustworthy Experience Pool (TEP), which forms actions as differences between real samples and stores their rewards, turning n data points into n(n−1) experiences; Automatic Model Refinement (AMR), where an LLM selects 1–3 new features from a candidate pool and retrains the regressor when early-training prediction variance or mid-training reward/value correlation falls below thresholds; and a Knowledge-Based Reward (KBR), where an LLM scores compositions against expert rules in the final 20% of training when classification confidence exceeds 0.8. The reward also includes UCB1-based decay for re-discovered materials. Each loop is positioned to correct a distinct failure mode: data scarcity, surrogate unreliability, and missing domain knowledge.

What would settle it

Compute the surrogate's prediction variance and reward-value Pearson correlation at the W-series compositions during training: if the AMR triggers do not fire even though measured elongations are systematically overpredicted (for example, predicted 15.9% versus measured 5.0% for Zr57Cu15Al10Ni10Fe8), then the refinement mechanism does not detect systematic bias, only variance and correlation anomalies.

Watch

Extended reading notes

Core claim

The paper's central claim is that augmenting experimental data with pairwise differences, dynamically refining the surrogate via LLM-selected features, and adding expert-knowledge rewards converts a sparse dataset into a reliable guide for inverse design. The authors show that AIMatDesign reaches an epoch-level success rate of 50.32% and per-property success rates of 46.93% and 49.38% for yield strength and elongation, exceeding all tested baselines under identical 128,000 surrogate-evaluation budgets. They report that the design-to-synthesis loop produces BMGs with yield strength agreement at 4.9% average error, and an alloy Zr63Cu15Al10Ni10Fe2 with 1.7 GPa yield strength and 10.2% elongation, and they conclude that the framework can deliver BMGs whose yield strength agrees closely with predictions while retaining appreciable ductility without further processing. The paper also states that measured plastic strains are systematically lower than predicted, attributing this to unmodeled processing conditions and surface defects.

Load-bearing premise

The load-bearing premise is that the machine-learning models used to judge candidate materials are accurate enough to serve as the reward oracle; the agent is trained against their probabilities and predictions, and the headline success rates are computed from those same models, so if the surrogates are biased in unexplored regions the framework explores and succeeds relative to a distorted objective.

Editorial extensions

If this is right

  • With 128,000 surrogate evaluations, AIMatDesign reaches an epoch-level success rate of 50.32%, about 3.4 times NSGA-II's 14.71%.
  • The LLM-driven refinement raises per-property success rates for yield strength and elongation to 46.93% and 49.38%, over six percentage points above the best RL baseline.
  • The difference-based experience pool accelerates early training, increasing average reward by about 0.1 within the first 5,000 steps compared with TD3.
  • Experimentally, the framework's proposed Zr-based alloys yield 4.9% average relative error between predicted and measured yield strength, with the Fe2 composition reaching 1.7 GPa yield strength and 10.2% elongation.

Reading between the lines

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

  • Because elongations were systematically overpredicted in the very alloy family used for validation, the same kind of surrogate bias may exist for other properties in unexplored composition regions, which suggests adding real experimental feedback for closed-loop retraining rather than relying on LLM-only refinement.
  • The TEP's n(n−1) pairwise differences implicitly assume the property landscape is smooth enough that differences between real samples define meaningful action directions, and that assumption may fail for chemically distant pairs.
  • The framework's success metrics are computed against the same surrogates that guide the RL agent, so an independent evaluation with a different model family or with measured properties would likely reduce the reported advantage.
  • The three-component recipe of difference-based replay, LLM feature refinement, and knowledge rewards should transfer to other data-scarce design spaces such as high-entropy alloys or battery electrolytes, where expert rules are also available in text form.
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

5 major / 6 minor

Summary. The paper proposes AIMatDesign, a reinforcement learning framework for inverse materials design under data scarcity. It combines a difference-based Trustworthy Experience Pool (TEP) to augment limited experimental data, an LLM-driven Automatic Model Refinement (AMR) mechanism that dynamically modifies the surrogate property predictors, and a Knowledge-Based Reward (KBR) that injects expert or LLM judgments in the final training stage. The framework is applied to the design of Zr-based bulk metallic glasses (BMGs). The authors report that AIMatDesign outperforms traditional optimization methods (grid search, NSGA-II) and RL baselines (DQN, DDPG, TD3, SAC, PPO) on step-level success rates (SR80%) and epoch-level completion rate (SR_done) in Tables 2 and 3. They synthesize eight compositions and report a top-performing alloy Zr63Cu15Al10Ni10Fe2 with 1.7 GPa yield strength and 10.2% elongation; predicted yield strengths are on average 4.9% from measured values, while elongations are systematically overpredicted.

Significance. If validated, the framework would represent a meaningful advance in integrating LLM-based model refinement and knowledge rewards into RL for materials discovery. The inclusion of real experimental synthesis of eight Zr-based alloys is a genuine strength: it provides independent ground truth that partially confirms the yield-strength predictions and demonstrates a closed-loop design-to-synthesis workflow. The ablation study suggests that each component contributes positively. However, the headline success-rate comparisons are computed with the same surrogate models that generate the RL rewards, and the surrogate's known bias on elongation undermines the ductility-related claims. With a corrected evaluation protocol and more thorough statistical reporting, the contribution could be solid.

major comments (5)
  1. [§4.2.2, Eq. (2), Table 2] The step-level success rates SR80%(ε) and the epoch-level SR_done in Table 2 are computed with the same edRVFL model whose predictions supply the reward in Eq. (2). The independent measurements in Table 4 show that this surrogate systematically overpredicts elongation (e.g., predicted 18.8% vs measured 4.6% for Zr55Cu15Al10Ni10Fe10, and roughly a factor of two on average), and Table 6 reports a MAPE of 84% for ε. The apparent advantage of AIMatDesign in ε-specific success rates (49.38% vs 43.43% for TD3) and the overall SR_done=50.32% therefore reflect consistency with a biased oracle rather than externally validated ductility. Please report success rates on the experimentally measured properties of the synthesized compositions, or re-evaluate with a bias-corrected elongation model, before claiming superiority in ductility optimization.
  2. [§4.3, Tables 2 and 3] All success rates in Tables 2 and 3 are single-run point estimates. Because the RL evaluation involves stochasticity in the initial base selection, action sampling, and network initialization, the reported differences between AIMatDesign and the best baselines (roughly 3–6 percentage points in SR80% and 3–5 points in SR_done) may be within run-to-run variability. The authors should provide means and standard deviations over at least five independent seeds for the main methods, and ideally a paired significance test, to support the claim that AIMatDesign 'significantly surpasses' the baselines.
  3. [§4.6, Table 4] The conclusion that AIMatDesign reliably predicts 'appreciable ductility' is not supported by the experimental data. For the headline alloy Zr63Cu15Al10Ni10Fe2, the predicted yield strength (1485 MPa) is 11% below the measured 1671 MPa, and the predicted elongation (14.3%) is 40% above the measured 10.2%. The 4.9% average relative error cited in the text applies only to yield strength; for elongation the average error across Table 4 is roughly 130%, and the predicted monotonic increase of ε with Fe content is contradicted by the measured peak at x=2. The abstract and conclusion should restrict the 'close agreement' claim to yield strength, or the elongation predictor must be corrected before ductility-guided design claims are made.
  4. [§3.1.2] The Trustworthy Experience Pool is motivated by claiming that difference-based experiences have 'significantly higher' trustworthiness than model-generated data, but the reward labels for these experiences are computed with the same CatBoost and edRVFL surrogates via Eq. (2). The trustworthiness argument applies to the compositional state transitions (derived from real sample pairs), not to the reward supervision, which inherits all surrogate bias. The paper should state this distinction explicitly and provide evidence—for example, a comparison of TEP reward labels against held-out experimental measurements—that the reward labels are sufficiently reliable to warrant the 'trustworthy' designation.
  5. [§3.2 and §4.3] The Automatic Model Refinement procedure modifies the regression model during training, so AIMatDesign optimizes against a non-stationary reward while the baseline methods are evaluated against a static surrogate. Table 2 thus conflates the benefit of the RL policy with the benefit of the evolving reward oracle. The authors should specify whether all methods are finally evaluated on the same (post-refinement) model, and ideally include a control in which a strong baseline such as TD3 is retrained after the model refinement, to isolate the algorithmic contribution of the RL component.
minor comments (6)
  1. [§4.2.2, Fig. 6, Table 6] The R² for ε is 0.93 in Fig. 6(g) but 0.92 in Table 6; unify these values, and note the 84% MAPE for ε when justifying the choice of edRVFL.
  2. [Abstract and §5] The phrases 'closely matching predictions' and 'retaining appreciable ductility' overstate the agreement for elongation; specify that yield strength, not ductility, is the property with close agreement.
  3. [Eq. (2) and §4.3] The thresholds τ_t and weights w_t are used in the reward but defined only implicitly through the SR80% metric; define them explicitly in the results setup.
  4. [Table 2] The baseline labeled 'Random[27]' is a random search, not an RL method; rename to 'Random search' and adjust the baseline categorization text accordingly.
  5. [Throughout] There are several formatting and typographical issues, including 'T able' in table captions, 'for for Material Exploration' in the Fig. 3 caption, 'a innovative' in Section 1, and inconsistent capitalization of 'Trustworthy Experience Pool'.
  6. [Data availability] The text states the code is available at a GitHub URL but later says it 'will be made publicly available upon acceptance'; clarify the current access status of the repository.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline success-rate comparisons are computed from the same edRVFL surrogate that defines the RL reward (Eq. 2), making SR80% and SR_done self-referential; only the yield-strength experimental check partially breaks the loop.

  1. self definitional [Section 3.1.1, Eq. (2), and Section 4.3, Evaluation Metrics / Table 2]
    "rreg(s, s′) = Σ_{t∈T} w_t·tanh((ŷ_t(s′)−ŷ_t(s))/max{τ_t, ŷ_t(s)}) + Σ_{t∈T} w_t·I[ŷ_t(s′)≥τ_t] (2) ... SR80%: The step-level success rate of generating samples that meet the top 80% of key performance indicators in the original dataset, including maximum diameter (Dmax), glass transition temperature ratio (Tg/Tl), yield strength (σY), Young's modulus (E), and elongation (ε(%)."

    The reward the agent maximizes is built from edRVFL predictions: w_t·I[ŷ_t(s')≥τ_t] in Eq. (2) is a threshold-success indicator on the surrogate outputs. The headline metrics SR80% and SR_done are defined as rates of 'generating samples that meet the top 80% of key performance indicators'; since every design step is a virtual environment query, those thresholds are evaluated with the same edRVFL outputs. Hence SR80%(ε)=49.38% and SR_done=50.32% count how often the policy satisfies the surrogate that defines its own reward, and the ≥6-point advantages over DQN/DDPG/TD3/PPO are optimizer-vs-optimizer comparisons against a shared fitted oracle, not externally grounded success.

full rationale

The central derivation chain is: edRVFL is fit to the BMG property dataset; its predictions enter the RL reward (Eq. 2); the agent is trained to maximize that reward; the success metrics in Tables 2 and 3 are computed by checking the same threshold conditions on the same edRVFL predictions. That makes the reported success rates self-consistent with the optimization objective by construction, not independent evidence of material-design quality. The experimental section (Table 4) is the only external link: yield-strength predictions track measurements with roughly 4.9% average error, a genuine non-circular validation of the σY channel. However, the same experiment shows the ε channel is systematically overpredicted and trend-inverted (predicted rise with Fe content versus measured peak at x=2), so the ductility-related success advantages claimed from the surrogate are not confirmed. No load-bearing self-citation chain was found: the cited prior work ([17], [35]) supplies components, but the paper's own equations and experiments carry the argument. Score 6 because the central comparative success-rate claim reduces, by construction, to optimizing a fitted surrogate whose objective and evaluation metric are the same function, while the yield-strength validation keeps the work from being wholly circular.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the surrogate models as a reward oracle, several hand-set thresholds and weights, and knowledge sources whose reliability is not independently quantified. The experimental validation supplies independent evidence but only for one family of Zr-based alloys.

free parameters (7)
  • Per-property reward thresholds tau_t = Top 80% quantiles of the training data for each property
    The r_t term in Eq. (2) and the SR80% metric are both defined by these thresholds, so success is measured against fitted dataset statistics rather than an independently set target.
  • Per-property reward weights w_t = Not reported
    Equation (2) weights the objectives, but the values and tuning procedure are omitted.
  • Knowledge-based reward weight beta = Not reported
    Algorithm 1 mixes the base reward with the LLM reward through beta, and the paper does not state the chosen value.
  • Variance threshold tau and correlation threshold rho = Not reported
    AMR triggers in Algorithm 1 and Fig. 4 depend on these thresholds; numerical values are not given.
  • Stage trigger fractions for AMR and KBR = 0.2, 0.8, and 0.8 of training steps
    The 20% and 80% training-stage boundaries are hand-set constants that control when each refinement mechanism activates.
  • Per-base composition bounds and action step size = Not reported
    The RL agent's additive and subtractive actions require component limits for each exploration base; the text describes ranges only qualitatively.
  • UCB1 exploration constant = Not reported
    Equation (3) adjusts the existing-material reward with UCB1, but the exploration coefficient is unspecified.
assumptions (5)
  • domain assumption Cross-validated performance of the surrogate models transfers to unexplored composition space
    The virtual environment and AMR acceptance criteria in Sections 4.2 and 4.4 use 5-fold and 10-fold R^2 to certify the guiding models, implicitly assuming generalization beyond the training region.
  • domain assumption Literature BMG data are representative of the experimental casting conditions used for validation
    The authors state in Section 4.6 that process parameters are often missing and that measured plastic strain is systematically lower than predicted, so this assumption is only partially satisfied.
  • domain assumption Element-wise additive and subtractive actions span the relevant design space
    The RL agent changes compositions by adding or removing elements with masked steps, which may miss optima requiring simultaneous large changes.
  • domain assumption The candidate feature pool of Xiong et al. contains features sufficient for refinement
    AMR lets the LLM select only 1 to 3 features from this fixed pool, so any informative feature outside the pool cannot be added.
  • domain assumption LLM knowledge retrieval provides scientifically calibrated reward and feature-selection signals
    KBR and AMR rely on GPT-4o with retrieval-augmented generation, without a quantitative check of the LLM's scoring accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AIMatDesign: Knowledge-Augmented Reinforcement Learning for Inverse Materials Design under Data Scarcity." pith.science (2026). https://pith.science/paper/D3WS5F4N

@misc{pith2026250700024,
  author       = {Pith},
  title        = {Pith review of: AIMatDesign: Knowledge-Augmented Reinforcement Learning for Inverse Materials Design under Data Scarcity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D3WS5F4N}},
  note         = {Machine review of arXiv:2507.00024}
}
read the original abstract

With the growing demand for novel materials, machine learning-driven inverse design methods face significant challenges in reconciling the high-dimensional materials composition space with limited experimental data. Existing approaches suffer from two major limitations: (I) machine learning models often lack reliability in high-dimensional spaces, leading to prediction biases during the design process; (II) these models fail to effectively incorporate domain expert knowledge, limiting their capacity to support knowledge-guided inverse design. To address these challenges, we introduce AIMatDesign, a reinforcement learning framework that addresses these limitations by augmenting experimental data using difference-based algorithms to build a trusted experience pool, accelerating model convergence. To enhance model reliability, an automated refinement strategy guided by large language models (LLMs) dynamically corrects prediction inconsistencies, reinforcing alignment between reward signals and state value functions. Additionally, a knowledge-based reward function leverages expert domain rules to improve stability and efficiency during training. Our experiments demonstrate that AIMatDesign significantly surpasses traditional machine learning and reinforcement learning methods in discovery efficiency, convergence speed, and success rates. Among the numerous candidates proposed by AIMatDesign, experimental synthesis of representative Zr-based alloys yielded a top-performing BMG with 1.7GPa yield strength and 10.2\% elongation, closely matching predictions. Moreover, the framework accurately captured the trend of yield strength variation with composition, demonstrating its reliability and potential for closed-loop materials discovery.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 57 canonical work pages

  1. [1]

    Market drivers for materials and process development in the 21st century,

    F. Field III, J. Clark, and M. Ashby, “Market drivers for materials and process development in the 21st century,”MRS Bulletin, vol. 26, no. 9, pp. 716–725, 2001

    work page 2001

  2. [2]

    Trial-and-error-based innovation: Rapid materialisation as catalyser of perception and communication in design

    J. Capjon, “Trial-and-error-based innovation: Rapid materialisation as catalyser of perception and communication in design.” 2004. AIMatDesign21

    work page 2004

  3. [3]

    Advances and challenges in dft-based energy materials design,

    J. Kang, X. Zhang, and S.-H. Wei, “Advances and challenges in dft-based energy materials design,”Chinese Physics B, vol. 31, no. 10, p. 107105, 2022

    work page 2022

  4. [4]

    Materials design and discovery with high-throughput density functional theory: the open quantum materials database (oqmd),

    J. E. Saal, S. Kirklin, M. Aykol, B. Meredig, and C. Wolverton, “Materials design and discovery with high-throughput density functional theory: the open quantum materials database (oqmd),” Jom, vol. 65, pp. 1501–1509, 2013

    work page 2013

  5. [5]

    Combinatorial and high-throughput materials science,

    W. F. Maier, K. Stoewe, and S. Sieg, “Combinatorial and high-throughput materials science,” Angewandte chemie international edition, vol. 46, no. 32, pp. 6016–6067, 2007

    work page 2007

  6. [6]

    Mdts: automatic complex materials design using monte carlo tree search,

    T. M. Dieb, S. Ju, K. Yoshizoe, Z. Hou, J. Shiomi, and K. Tsuda, “Mdts: automatic complex materials design using monte carlo tree search,”Science and technology of advanced materials, vol. 18, no. 1, pp. 498–503, 2017

    work page 2017

  7. [7]

    Bayesian optimization for materials design,

    P. I. Frazier and J. Wang, “Bayesian optimization for materials design,”Information science for materials discovery and design, pp. 45–75, 2016

    work page 2016

  8. [8]

    Genetic algorithms in materials design and processing,

    N. Chakraborti, “Genetic algorithms in materials design and processing,”International Materials Reviews, vol. 49, no. 3-4, pp. 246–260, 2004

    work page 2004

Show all 72 references
  1. [9]

    Materials discovery and design using machine learning,

    Y. Liu, T. Zhao, W. Ju, and S. Shi, “Materials discovery and design using machine learning,” Journal of Materiomics, vol. 3, no. 3, pp. 159–177, 2017

    2017

  2. [10]

    Machine learning in materials informatics: recent applications and prospects,

    R. Ramprasad, R. Batra, G. Pilania, A. Mannodi-Kanakkithodi, and C. Kim, “Machine learning in materials informatics: recent applications and prospects,”npj Computational Materials, vol. 3, no. 1, p. 54, 2017

    2017

  3. [11]

    A generative approach to materials discovery, design, and optimization,

    D. Menon and R. Ranganathan, “A generative approach to materials discovery, design, and optimization,”ACS omega, vol. 7, no. 30, pp. 25 958–25 973, 2022

    2022

  4. [12]

    Generative adversarial networks (gan) based efficient sampling of chemical composition space for inverse design of inorganic materials,

    Y. Dan, Y. Zhao, X. Li, S. Li, M. Hu, and J. Hu, “Generative adversarial networks (gan) based efficient sampling of chemical composition space for inverse design of inorganic materials,”npj Computational Materials, vol. 6, no. 1, p. 84, 2020

    2020

  5. [13]

    Encoding and exploring latent design space of optimal material structures via a vae-lstm model,

    A. J. Lew and M. J. Buehler, “Encoding and exploring latent design space of optimal material structures via a vae-lstm model,”Forces in Mechanics, vol. 5, p. 100054, 2021

    2021

  6. [14]

    Vae for modified 1-hot generative materials modeling, a step towards inverse material design,

    K. El-Awady, “Vae for modified 1-hot generative materials modeling, a step towards inverse material design,”arXiv preprint arXiv:2401.06779, 2023

    2023 arXiv

  7. [15]

    A semi-supervised approach to architected materials design using graph neural networks,

    K. Guo and M. J. Buehler, “A semi-supervised approach to architected materials design using graph neural networks,”Extreme Mechanics Letters, vol. 41, p. 101029, 2020

    2020

  8. [16]

    Inverse design of glass structure with deep graph neural networks,

    Q. Wang and L. Zhang, “Inverse design of glass structure with deep graph neural networks,” Nature communications, vol. 12, no. 1, p. 5359, 2021

    2021

  9. [17]

    From small data modeling to large language model screening: A dual-strategy framework for materials intelligent design,

    Y. Yu, J. Xiong, X. Wu, and Q. Qian, “From small data modeling to large language model screening: A dual-strategy framework for materials intelligent design,”Advanced Science, vol. 11, no. 45, p. 2403548, 2024

    2024

  10. [18]

    Multi-objective optimization for materials design with improved nsga-ii,

    P. Zhang, Y. Qian, and Q. Qian, “Multi-objective optimization for materials design with improved nsga-ii,”Materials today communications, vol. 28, p. 102709, 2021. 22AIMatDesign

    2021

  11. [19]

    A comprehensive survey on nsga-ii for multi- objective optimization and applications,

    H. Ma, Y. Zhang, S. Sun, T. Liu, and Y. Shan, “A comprehensive survey on nsga-ii for multi- objective optimization and applications,”Artificial Intelligence Review, vol. 56, no. 12, pp. 15 217– 15 270, 2023

    2023

  12. [20]

    Deep reinforcement learning: An overview,

    Y. Li, “Deep reinforcement learning: An overview,”arXiv preprint arXiv:1701.07274, 2017

    2017 arXiv

  13. [21]

    Deep reinforcement learning for digital materials design,

    F. Sui, R. Guo, Z. Zhang, G. X. Gu, and L. Lin, “Deep reinforcement learning for digital materials design,”ACS Materials Letters, vol. 3, no. 10, pp. 1433–1439, 2021

    2021

  14. [22]

    Deep reinforcement learning for engineering design through topology optimization of elementally discretized design domains,

    N. K. Brown, A. P. Garland, G. M. Fadel, and G. Li, “Deep reinforcement learning for engineering design through topology optimization of elementally discretized design domains,”Materials & Design, vol. 218, p. 110672, 2022

    2022

  15. [23]

    Reinforcement learning applied to metamaterial design,

    T. Shah, L. Zhuo, P. Lai, D. La Rosa-Moreno, F. Amirkulova, P. Gerstoftet al., “Reinforcement learning applied to metamaterial design,”The Journal of the Acoustical Society of America, vol. 150, no. 1, pp. 321–338, 2021

    2021

  16. [24]

    Deep reinforcement learning for inverse inorganic materials design,

    C. Karpovich, E. Pan, and E. A. Olivetti, “Deep reinforcement learning for inverse inorganic materials design,”npj Computational Materials, vol. 10, no. 1, p. 287, 2024

    2024

  17. [25]

    Llmatdesign: Autonomous materials discovery with large language models,

    S. Jia, C. Zhang, and V. Fung, “Llmatdesign: Autonomous materials discovery with large language models,”arXiv preprint arXiv:2406.13163, 2024

    2024 arXiv

  18. [26]

    Beyond designer’s knowledge: Generating materials design hypotheses via large language models,

    Q. Liu, M. P. Polak, S. Y. Kim, M. Shuvo, H. S. Deodhar, J. Han, D. Morgan, and H. Oh, “Beyond designer’s knowledge: Generating materials design hypotheses via large language models,”arXiv preprint arXiv:2409.06756, 2024

    2024 arXiv

  19. [27]

    Grid search, random search, genetic algorithm: a big comparison for nas,

    P. Liashchynskyi and P. Liashchynskyi, “Grid search, random search, genetic algorithm: a big comparison for nas,”arXiv preprint arXiv:1912.06059, 2019

    1912 arXiv

  20. [28]

    Taking the human out of the loop: A review of bayesian optimization,

    B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. De Freitas, “Taking the human out of the loop: A review of bayesian optimization,”Proceedings of the IEEE, vol. 104, no. 1, pp. 148–175, 2015

    2015

  21. [29]

    A fast and elitist multiobjective genetic algorithm: Nsga-ii,

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: Nsga-ii,”IEEE transactions on evolutionary computation, vol. 6, no. 2, pp. 182–197, 2002

    2002

  22. [30]

    High-dimensional bayesian optimization for metamaterial design,

    Z. Tian, Y. Yang, S. Zhou, T. Zhou, K. Deng, C. Ji, Y. He, and J. S. Liu, “High-dimensional bayesian optimization for metamaterial design,”Materials Genome Engineering Advances, vol. 2, no. 4, p. e79, 2024

    2024

  23. [31]

    Topology optimization for auxetic metamaterials based on isogeometric analysis,

    J. Gao, H. Xue, L. Gao, and Z. Luo, “Topology optimization for auxetic metamaterials based on isogeometric analysis,”Computer Methods in Applied Mechanics and Engineering, vol. 352, pp. 211–236, 2019

    2019

  24. [32]

    Topological optimization design of multi-material phononic crystals with floating projection constraints to achieve ultra-wide band gap,

    L. Liao, S. Yao, and Y. Li, “Topological optimization design of multi-material phononic crystals with floating projection constraints to achieve ultra-wide band gap,”Composite Structures, vol. 346, p. 118387, 2024

    2024

  25. [33]

    Introduction to macroscopic optimal design in the mechanics of composite materials and structures,

    A. Muc, “Introduction to macroscopic optimal design in the mechanics of composite materials and structures,”Journal of Composites Science, vol. 5, no. 2, p. 36, 2021. AIMatDesign23

    2021

  26. [34]

    Designing multi-objective multi-armed bandits algorithms: A study,

    M. M. Drugan and A. Nowe, “Designing multi-objective multi-armed bandits algorithms: A study,” inThe 2013 international joint conference on neural networks (IJCNN). IEEE, 2013, pp. 1–8

    2013

  27. [35]

    A machine-learning approach to predicting and under- standing the properties of amorphous metallic alloys,

    J. Xiong, S.-Q. Shi, and T.-Y. Zhang, “A machine-learning approach to predicting and under- standing the properties of amorphous metallic alloys,”Materials & Design, vol. 187, p. 108378, 2020

    2020

  28. [36]

    Logistic regression,

    M. P. LaValley, “Logistic regression,”Circulation, vol. 117, no. 18, pp. 2395–2399, 2008

    2008

  29. [37]

    Linear discriminant analysis,

    P. Xanthopoulos, P. M. Pardalos, T. B. Trafalis, P. Xanthopoulos, P. M. Pardalos, and T. B. Trafalis, “Linear discriminant analysis,”Robust data mining, pp. 27–33, 2013

    2013

  30. [38]

    Online training of support vector classifier,

    K. Lau and Q. Wu, “Online training of support vector classifier,”Pattern Recognition, vol. 36, no. 8, pp. 1913–1920, 2003

    1913

  31. [39]

    Decision tree methods: applications for classification and prediction,

    L. Yinget al., “Decision tree methods: applications for classification and prediction,”Shanghai archives of psychiatry, vol. 27, no. 2, p. 130, 2015

    2015

  32. [40]

    Random forests,

    L. Breiman, “Random forests,”Machine learning, vol. 45, pp. 5–32, 2001

    2001

  33. [41]

    Gradient boosting machines, a tutorial,

    A. Natekin and A. Knoll, “Gradient boosting machines, a tutorial,”Frontiers in neurorobotics, vol. 7, p. 21, 2013

    2013

  34. [42]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” inProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 2016, pp. 785–794

    2016

  35. [43]

    Catboost: unbiased boosting with categorical features,

    L. Prokhorenkova, G. Gusev, A. Vorobev, A. V. Dorogush, and A. Gulin, “Catboost: unbiased boosting with categorical features,”Advances in neural information processing systems, vol. 31, 2018

    2018

  36. [44]

    Multi-class adaboost,

    J. Zhu, H. Zou, S. Rosset, T. Hastieet al., “Multi-class adaboost,”Statistics and its Interface, vol. 2, no. 3, pp. 349–360, 2009

    2009

  37. [45]

    K-nearest neighbors,

    O. Kramer and O. Kramer, “K-nearest neighbors,”Dimensionality reduction with unsupervised nearest neighbors, pp. 13–23, 2013

    2013

  38. [46]

    Fast gaussian na ¨ ıve bayes for searchlight classification analysis,

    M. Ontivero-Ortega, A. Lage-Castellanos, G. Valente, R. Goebel, and M. Valdes-Sosa, “Fast gaussian na ¨ ıve bayes for searchlight classification analysis,”Neuroimage, vol. 163, pp. 471–479, 2017

    2017

  39. [47]

    Multinomial naive bayes classification model for sentiment analysis,

    M. Abbas, K. A. Memon, A. A. Jamali, S. Memon, and A. Ahmed, “Multinomial naive bayes classification model for sentiment analysis,”IJCSNS Int. J. Comput. Sci. Netw. Secur, vol. 19, no. 3, p. 62, 2019

    2019

  40. [48]

    Naive bayes classifiers,

    K. P. Murphyet al., “Naive bayes classifiers,”University of British Columbia, vol. 18, no. 60, pp. 1–8, 2006

    2006

  41. [49]

    Linear vs. quadratic discriminant analysis classifier: a tutorial,

    A. Tharwat, “Linear vs. quadratic discriminant analysis classifier: a tutorial,”International Journal of Applied Pattern Recognition, vol. 3, no. 2, pp. 145–180, 2016

    2016

  42. [50]

    Ridge regression,

    G. C. McDonald, “Ridge regression,”Wiley Interdisciplinary Reviews: Computational Statistics, vol. 1, no. 1, pp. 93–100, 2009. 24AIMatDesign

    2009

  43. [51]

    Lasso regression,

    J. Ranstam and J. A. Cook, “Lasso regression,”Journal of British Surgery, vol. 105, no. 10, pp. 1348–1348, 2018

    2018

  44. [52]

    Regularization and variable selection via the elastic net,

    H. Zou and T. Hastie, “Regularization and variable selection via the elastic net,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 67, no. 2, pp. 301–320, 2005

    2005

  45. [53]

    Support vector regression,

    M. Awad, R. Khanna, M. Awad, and R. Khanna, “Support vector regression,”Efficient learning machines: Theories, concepts, and applications for engineers and system designers, pp. 67–80, 2015

    2015

  46. [54]

    Adaboost. rt: a boosting algorithm for regression problems,

    D. P. Solomatine and D. L. Shrestha, “Adaboost. rt: a boosting algorithm for regression problems,” in2004 IEEE international joint conference on neural networks (IEEE Cat. No. 04CH37541), vol. 2. IEEE, 2004, pp. 1163–1168

    2004

  47. [55]

    Ensemble deep random vector func- tional link neural network for regression,

    M. Hu, J. H. Chion, P. N. Suganthan, and R. K. Katuwal, “Ensemble deep random vector func- tional link neural network for regression,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 53, no. 5, pp. 2604–2615, 2022

    2022

  48. [56]

    Gpt-4o system card,

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radfordet al., “Gpt-4o system card,”arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  49. [57]

    Human-level control through deep reinforcement learning,

    V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Ried- miller, A. K. Fidjeland, G. Ostrovskiet al., “Human-level control through deep reinforcement learning,”nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  50. [58]

    Continuous control with deep reinforcement learning,

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforcement learning,”arXiv preprint arXiv:1509.02971, 2015

    2015 arXiv

  51. [59]

    Addressing function approximation error in actor-critic methods,

    S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approximation error in actor-critic methods,” inInternational conference on machine learning. PMLR, 2018, pp. 1587–1596

    2018

  52. [60]

    Soft actor-critic algorithms and applications,

    T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, V. Kumar, H. Zhu, A. Gupta, P. Abbeelet al., “Soft actor-critic algorithms and applications,”arXiv preprint arXiv:1812.05905, 2018

    2018 arXiv

  53. [61]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  54. [62]

    Algorithm as 136: A k-means clustering algorithm,

    J. A. Hartigan and M. A. Wong, “Algorithm as 136: A k-means clustering algorithm,”Journal of the royal statistical society. series c (applied statistics), vol. 28, no. 1, pp. 100–108, 1979

    1979

  55. [63]

    Retrieval-augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. K¨ uttler, M. Lewis, W.-t. Yih, T. Rockt¨ aschelet al., “Retrieval-augmented generation for knowledge-intensive nlp tasks,” Advances in neural information processing systems, vol. 33, pp. 9459–9474, 2020. AI...

    2020

  56. [64]

    Review and understand the selection criteria (RULE), identifying key indicators and requirements

  57. [65]

    Compare with similar BMGs, analyzing performance characteristics and experimental outcomes as benchmarks

  58. [66]

    reward”: Data point’s reward value, [-1, 1], rounded to two decimal places, “reason

    Evaluate the provided data point (DATA) against the selection criteria and reference data, assigning a reward value and justifying your reasoning. RULE: {rule} Similar Real BMGs: {similarreal bmg} DA T A: {data} The reward value should range from -1 to 1, where 1 indicates hig...

  59. [67]

    Analyze the current state of the Guiding Model, including the features used and the potential reasons for high prediction variance, to identify areas for improvement

  60. [68]

    Evaluate each candidate feature, considering its correlation with the current high-variance BMG compositions and performance, data quality, and its potential impact on model prediction ability

  61. [69]

    selectedfeatures

    Based on the evaluation of the model’s improvement direction and candidate features, select the 1-3 most promising features and provide a brief explanation of why these features were chosen. Reference Knowledge: {knowledge} Guiding Model Status: {model status} Candidate F eatu...

  62. [70]

    Analyze the current state of the Guiding Model, including the features used and the potential reasons for the low Pearson correlation between the reward curve (Rf) and the state value curve (Vf), and identify areas for improvement

  63. [71]

    Evaluate each candidate feature, considering its potential relationship with the current inconsistency in predictions, and assess whether adding the feature will improve the machine learning model’s performance, helping it align with the reinforcement learning model

  64. [72]

    selectedfeatures

    Select the most optimal features and justify your choice by considering the direction of model improvement and the evaluation of candidate features. Select the 1-3 most promising features and briefly explain the rationale behind these selections. Reference Knowledge: {knowledg...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.