{"id":"e4eb37f1-8b61-4319-ab82-3698a0aba4ba","arxiv_id":"2411.13953","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper proposes ML-guided simulation for marble sludge mix design but only benchmarks porosity regression on an existing concrete dataset, leaving the central reuse claim untested.","lead":"This position paper proposes using machine learning models with Bayesian optimization to predict material properties and guide mix designs for recycling marble sludge, but it only tests the idea on a small public concrete dataset and provides no sludge-specific experiments. A smart generalist might read it as a proposal for cheaper, faster construction waste reuse, but the core claim is not demonstrated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'simulation' in the title and Fig. 1 is never built: the ML model is a regressor trained on 240 concrete records, so the promised 'large dataset generated through simulations' is unsupported, and no sludge-specific transfer is tested.","rationale":"The reader's weakest assumption was the concrete-to-sludge transfer. I partially agree, but I see a more fundamental gap: the paper never actually builds the simulator that is the basis of its title and main advantage. Section 3.4 conflates prediction with simulation: a regressor trained on historical data cannot generate new labeled data. This is not a dispute with consensus; it is an internal inconsistency between the claimed 'simulator' and the described method. The only experiment is a small concrete benchmark with a single random split and no error bars, so even the reported MSE of 0.1077 is not robust evidence. The transfer to marble sludge is therefore doubly unsupported: there is neither a generative simulation nor a valid proxy demonstration. A position paper may propose speculative ideas, but the abstract and Section 1 make concrete quantitative claims ('MSE of 0.1077', 'large dataset can be generated') that are presented as results. Because these claims are load-bearing for the central argument, the appropriate verdict remains REJECT. The proposed end-to-end test would settle the matter: if the authors can generate novel sludge mix designs and validate them, the idea becomes testable; if not, the paper is only a proposal.","tokens_in":8809,"tokens_out":5470,"duration_ms":52981,"concrete_test":"Have the authors provide the executable simulator and run it end-to-end: use the tuned GB model to propose a set of novel marble-sludge mix designs (e.g., 10,000 sampled compositions with sludge, binder, water, and curing parameters) and physically measure porosity for 20 of the proposed designs. If the pipeline cannot propose new compositions, or if measured porosities deviate from the model's predictions by more than the reported test MSE of 0.1077, the 'large dataset from simulation' claim and the transfer to sludge are falsified. Report the code, the generated input distribution, and the 20 measured values.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central promise is that 'through simulations, a large dataset can be generated' (Abstract) and that an 'ML-guided Simulation' can replace trial-and-error mix design (Fig. 1). No such simulator is provided. In Section 3.4 the 'simulator' is a gradient-boosting regressor trained on 240 historical concrete records; the text even claims it estimates porosity 'without relying on historical data and using only current material inputs,' but the model is literally built from historical data. A supervised regressor predicts porosity for input feature vectors; it cannot generate new ground-truth data points. No generative model, physical forward model, or sampling procedure is described, so the 'large dataset' advantage reduces to interpolation within the existing feature distribution. This is load-bearing because the title, abstract, and Figure 1 all promise simulation-driven synthesis, and the proposed transfer to marble sludge depends on having a working simulator. Additionally, no sludge-specific features (e.g., CaCO3 content, particle size distribution, sludge-to-binder ratio) are used, and no transfer experiment appears: Section 4 validates only concrete porosity as a proxy. Thus the conclusion that ML 'can estimate the correct quantity of stone-cutting sludge' rests on an unbuilt simulator and an untested analogy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper proposes using machine learning, with hyperparameter optimization via 'meta-learning' (identified as Bayesian optimization), to guide material mix design for marble sludge reuse. The authors train several regression models (SVM, LR, RF, GB, XGB) on a public dataset of 240 concrete records and report that after Bayesian optimization, gradient boosting predicts porosity with a test MSE of 0.1077 and R² of 0.8887, claiming this unlocks simulation-driven mix design and that the approach can be transferred to stone-cutting sludge. The paper presents the concrete experiment as a proxy because no sludge-specific dataset is available.","tokens_in":9095,"tokens_out":6237,"duration_ms":56367,"significance":"If fully supported, the paper would offer a low-cost, data-efficient path for estimating concrete porosity and a plausible first step toward ML-guided mix design for recycled stone sludge. The manuscript has clear strengths: it uses a public dataset, reports standard metrics (MSE, R², MAPE) with a held-out test set, compares five models, and documents the practical motivation for stone-cutting sludge reuse in detail. However, the central contribution is not delivered: the 'simulator' is a supervised regressor and cannot generate the promised large dataset, no sludge-specific data or transfer experiment is provided, and the 'meta-learning' claim is a mislabeling of Bayesian hyperparameter optimization. The paper therefore does not substantiate its headline claims.","major_comments":[{"comment":"The 'simulator' is not a simulator and cannot generate large datasets as the Abstract promises. The text claims that the model 'enabl[es] the estimation of porosity without relying on historical data and using only current material inputs,' but the model's parameters are learned entirely from 240 historical concrete records. A supervised regressor can only interpolate predictions for new feature combinations; it cannot generate new ground-truth data points. No generative model, physical forward model, or sampling procedure is described. Consequently, the Abstract's statement that 'through simulations, a large dataset can be generated' is unsupported, and the title's promise of 'material synthesis through simulations' is not met.","section":"Section 3.4"},{"comment":"The transfer to stone-cutting sludge is untested and rests on an unverified analogy. Section 4 justifies the concrete dataset by stating 'Concrete quality estimation is a very similar task to the one we aim to solve,' but no sludge-specific features (e.g., CaCO3 content, particle-size distribution, sludge-to-binder ratio) are used, and no transfer experiment is performed. The conclusion that ML 'can estimate the correct quantity of stone-cutting sludge to be used in aggregates' (Abstract) is therefore a speculation about transfer across materials with different chemistry, variability, and processing, not a demonstrated result.","section":"Section 4"},{"comment":"The feature set includes Strength as an input to predict porosity. In a mix-design workflow, strength is a target property that would not be known before the mix is produced; using it as an input makes the model an inverse or diagnostic predictor, not a forward design simulator. The paper does not explain how a user of the proposed simulator would supply 'Strength' before making and testing the concrete, which undermines the claimed applicability of the simulator for discovering optimal mix designs.","section":"Table 1"},{"comment":"The identification of Bayesian hyperparameter optimization as 'meta-learning' is a terminology error with consequences for the paper's central claim. Meta-learning refers to learning across tasks (e.g., few-shot learning, learning-to-learn); Bayesian optimization is a standard hyperparameter-search technique applied to a single dataset. No multi-task or few-shot experiments are reported, so the paper's advertised advantage of using meta-learning in data-constrained settings is not demonstrated.","section":"Section 3.2"},{"comment":"The evaluation protocol is underspecified regarding the test set. The paper does not state whether Bayesian optimization used the same test partition for feedback (Figure 2 shows 'evaluation feedback' but not its source); if the test set was used to select hyperparameters, the reported test MSE of 0.1077 would be optimistically biased. Moreover, the random 80/20 split on 240 records from only 74 unique mixture designs may place the same design in both training and testing, inflating performance. A grouped split by mixture design should be reported as a check.","section":"Section 3.3 and Figure 2"}],"minor_comments":[{"comment":"The term 'Z-Normalization' is nonstandard; use 'z-score normalization' or 'standardization'.","section":"Section 3.1"},{"comment":"The headings 'T raining' and 'T esting' contain spurious spaces that should be removed.","section":"Sections 3.2 and 3.3"},{"comment":"The figure references appear inconsistent: the text says 'Figure 4 displays the performance ... before any hyper-parameter optimization,' but Figure 3 is captioned 'without meta-learning' and Figure 4 'with meta-learning.' Please correct the numbering or the verbal description.","section":"Section 4.1"},{"comment":"The discussion of GB states 'GB performs the worst among the models, with an (R2 = 0.4738)' without specifying whether this is the training or test value; from Table 2 the value appears to be the test R², but the sentence should state this explicitly.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The paper is a position paper with a preliminary experiment, but its central claims are not supported by the evidence. The 'simulator' is a mislabeled regressor, the sludge-transfer claim is untested, and the 'meta-learning' terminology is unjustified. These are load-bearing issues that cannot be fixed by local revisions; they would require either a proper generative or forward model and a sludge-specific transfer study, or a substantial reframing of the contribution. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's title and abstract promise ML-guided simulation for material synthesis, but what you actually get is a regression benchmark: four standard models tuned with Bayesian optimization on a public concrete porosity dataset (240 records) from Cao 2023. The best test MSE is 0.1077 for gradient boosting after tuning. That part is fine: public data, standard splits, multiple metrics, and a clear table. The paper is also honest about being a position paper and about the scarcity of sludge data. It cites prior ML mix-design work correctly, including Liu and Cao, so the citation pattern looks solid.\n\nWhat is not fine is the gap between the claims and the evidence. The central promise, that \"through simulations, a large dataset can be generated,\" is unsupported. The \"simulator\" in Section 3.4 is a supervised regressor trained on historical data; it predicts porosity for input feature vectors but cannot generate new ground-truth points. The text even claims the simulator works \"without relying on historical data,\" which is contradicted by its own training procedure. No generative model, physical forward model, or sampling scheme is described. And the transfer to marble sludge is never tested: no sludge-specific features (CaCO3 content, particle size distribution, sludge-to-binder ratio), no sludge data, no transfer experiment. The strongest support for the sludge application is the sentence in Section 4 that concrete quality estimation is \"a very similar task,\" which is an assumption, not a result.\n\nThe concrete regression itself is plausible, but it has soft spots. There are no error bars or confidence intervals, and the paper does not compare its numbers with those in Cao 2023, so it is hard to tell whether the Bayesian tuning actually beats the source paper's baseline. The feature set is selected by Pearson correlation with the target, which is a form of target leakage in feature selection, and including compressive Strength as an input for porosity is at least debatable since strength is a consequence of curing, not a mix input. These are not fatal to the benchmark, but they weaken the \"simulation\" story further.\n\nWho should read this? People working on ML for construction materials who want a quick look at how Bayesian hyperparameter optimization affects standard regressors on a small concrete dataset. It could be a workshop paper or a short conference paper if reframed. As it stands, the title, abstract, and Figure 1 overpromise. If I were an editor, I would send this to peer review because the benchmark is reproducible and the topic is relevant, but I would expect major revision: either build a real simulator (generative or physical) or drop the simulation language, and either add sludge data or soften the transfer claim to a proposal. My own verdict would be skeptical, but the paper deserves a serious look rather than a desk reject.","headline":"The paper's real content is a small, reproducible regression benchmark on concrete porosity with Bayesian tuning; the 'simulation' and sludge-reuse promise are unsupported, so it reads as an overclaimed position paper.","tokens_in":748,"tokens_out":1337,"would_cite":false,"duration_ms":29899,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A tuned gradient-boosting model predicts concrete porosity with a mean squared error of 0.1077, suggesting machine-learning-guided simulations could replace costly trial-and-error mix design for stone-cutting sludge reuse.","keywords":["marble sludge recycling","mix design","machine learning","Bayesian optimization","gradient boosting","porosity prediction","simulation-guided materials","position paper"],"falsifier":"Collect a dataset of marble-sludge concrete mixes with measured porosity (or any target property), train the same gradient-boosting model with Bayesian optimization on it, and evaluate on a held-out test set; if the test MSE is substantially worse than the 0.1077 achieved on concrete, or no better than predicting the mean, the transfer claim collapses. A simpler check: compare the model's predicted optimal sludge mix against a randomly chosen mix in a physical experiment; if the model-guided mix does not reliably outperform the baseline, the practical value is unproven.","tokens_in":8608,"feed_emoji":"🧱","tokens_out":2290,"duration_ms":25633,"temperature":0.7,"pith_summary":"This position paper argues that machine learning, combined with Bayesian-optimization-based hyperparameter tuning (which the authors call meta-learning), can predict material properties accurately enough to guide the mix design of concrete made with stone-cutting sludge. To support this claim in the absence of sludge-specific data, the authors train several regression models on a public concrete dataset and show that gradient boosting achieves a test mean squared error of 0.1077 for porosity prediction. They propose this as a proof-of-concept that simulations can generate large datasets and identify optimal sludge mix designs, reducing the need for lengthy and costly physical experiments. The paper's contribution is thus a proposal backed by a preliminary demonstration, not a validated industrial workflow.","feed_headline":"Tuned ML model predicts concrete porosity with MSE 0.1077","feed_subtitle":"A proof-of-concept that machine-learning-guided simulations could replace costly trial-and-error mix design for stone sludge reuse.","key_machinery":"The central mechanism is the machine-learning-based 'simulator': a regression model (gradient boosting, after tuning) that takes material characteristics — water-to-binder ratio, binder content, fly ash, slag, superplasticizer, aggregate ratio, curing days, and strength — as inputs and outputs a predicted porosity. Hyperparameter tuning is carried out by Bayesian optimization, which the paper describes as a meta-learning approach: it builds a Gaussian-process surrogate of the objective function and selects promising hyperparameter combinations, balancing exploration and exploitation. This tuning step is what lifts gradient boosting from the worst-performing model (test R² 0.2820) to the best (test R² 0.8887), and it is the load-bearing procedure for the paper's claim that ML can be made trustworthy for mix-design guidance.","core_discovery":"The paper's central claim is that machine-learning-guided simulation can replace much of the empirical trial-and-error in material mix design, specifically for reusing marble sludge. On the concrete dataset, the authors apply five regression models (linear regression, support vector regression, random forest, gradient boosting, and XGBoost) and then tune them with Bayesian optimization. After tuning, gradient boosting performs best, reaching a test MSE of 0.1077, a test R² of 0.8887, and a test MAPE of 0.6933. The authors conclude that this level of accuracy 'unlocks the possibility of trusting the simulation' and performing only a restricted number of real experiments. They propose extending the same approach to stone-cutting sludge mix design once relevant data become available, arguing that the concrete task is a similar proxy.","pith_inferences":["The paper's use of the term 'meta-learning' is nonstandard: Bayesian optimization for hyperparameter tuning is not learning from multiple related tasks in the usual meta-learning sense, so readers should interpret the claim accordingly.","The strongest test of the proposal would be to run the identical pipeline on an actual marble-sludge mix-design dataset; until that is done, the concrete result remains a proxy demonstration, and the transfer is an assumption rather than an established finding.","If sludge-specific data are too scarce for training, the paper's own logic suggests a more direct route: collect a modest set of sludge mixes, tune the model with Bayesian optimization, and check whether the predicted optimal mix outperforms a random or heuristic baseline in physical testing.","The approach could be extended to active learning, where the simulator suggests which experiments to run next, using the model's uncertainty to minimize the number of physical tests needed."],"forward_implications":["If the approach transfers, stone producers could screen candidate sludge mix designs computationally before running any physical tests, cutting both cost and waste.","The same tuned-regression pipeline could be applied to other material properties — not just porosity — such as compressive strength or durability, whenever a relevant dataset exists.","Simulation-driven mix design could be adapted to different regional building regulations by retraining on local data, reducing the need for redundant experiments in each jurisdiction.","The demonstrated performance jump from Bayesian optimization suggests that many existing ML applications in materials may be underperforming simply due to poor hyperparameter choices.","A reliable porosity simulator would let researchers explore a far wider mix-design space than laboratory budgets allow, potentially discovering better-performing recipes than current heuristics."],"supporting_citations":[{"why":"Supplies the public concrete dataset with porosity as target and the eight material features used to train and test all models.","marker":"[7]"},{"why":"Provides the Bayesian optimization method that the paper uses as its 'meta-learning' hyperparameter-tuning engine.","marker":"[8]"},{"why":"Precedent for applying ML (SVR and gradient boosting) to asphalt mix design, supporting the claim that ML can optimize mix designs.","marker":"[11]"},{"why":"Documents the challenge of hyperparameter optimization in machine learning, motivating the paper's use of Bayesian optimization.","marker":"[20]"}],"fun_headline_variants":["ML-guided simulations slash trial-and-error in material mix design","AI predicts concrete porosity, cutting need for physical experiments","Tuned gradient boosting yields low error on concrete porosity tests","Simulation plus meta-learning optimizes stone sludge mix designs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that concrete porosity prediction is a valid proxy for marble-sludge mix design, so a model trained on concrete data will transfer to sludge-based mixtures with different chemistry and variability — but no sludge-specific data or transfer experiment is provided.","fun_headline_variants_meta":{"raw":{"variants":["ML-guided simulations slash trial-and-error in material mix design","AI predicts concrete porosity, cutting need for physical experiments","Tuned gradient boosting yields low error on concrete porosity tests","Simulation plus meta-learning optimizes stone sludge mix designs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1612,"prompt_tokens":920,"completion_tokens":692,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":626}},"tokens_in":536,"tokens_out":692,"duration_ms":7412,"temperature":1.0,"reasoning_tokens":626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:41:35.862882+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a dataset of marble-sludge concrete mixes with measured porosity (or any target property), train the same gradient-boosting model with Bayesian optimization on it, and evaluate on a held-out test set; if the test MSE is substantially worse than the 0.1077 achieved on concrete, or no better than predicting the mean, the transfer claim collapses. A simpler check: compare the model's predicted optimal sludge mix against a randomly chosen mix in a physical experiment; if the model-guided mix does not reliably outperform the baseline, the practical value is unproven.","supporting_citations":[{"cited_title":"Information Sciences 591, 155–175 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian optimization method that the paper uses as its 'meta-learning' hyperparameter-tuning engine."},{"cited_title":"Construction and Building Materials325, 126607 (2022)","cited_arxiv_id":null,"evidence_quote":"Precedent for applying ML (SVR and gradient boosting) to asphalt mix design, supporting the claim that ML can optimize mix designs."},{"cited_title":"Neurocomputing415, 295–316 (2020)","cited_arxiv_id":null,"evidence_quote":"Documents the challenge of hyperparameter optimization in machine learning, motivating the paper's use of Bayesian optimization."}],"review_version":1}