{"id":"24232272-2bd3-4842-9650-07a35e3a721a","arxiv_id":"2412.11981","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Machine learning models trained on two years of industrial cement plant data predict clinker mineral phases more accurately than the standard Bogue equation, though the headline numbers partly rely on post-production oxide inputs.","lead":"This paper trains eight machine learning models on two years of data from an industrial cement plant to predict the mineral phases in clinker, and reports that they beat the standard Bogue equation. A generalist might read it because accurate real-time clinker phase prediction could cut waste and CO2 in cement production, which is responsible for about 8% of global emissions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fixed ~37-minute synchronization in §2.1 omits rotary kiln residence time, so inputs are likely aligned too early; every reported accuracy depends on this alignment.","rationale":"The reader's weakest-assumption analysis correctly targets temporal synchronization, and I agree that the random split of an autocorrelated series and the aggressive percentile filter inflate apparent accuracy. My concern sharpens this: the paper's own process timeline omits the rotary kiln residence time entirely. In a dry-process cement plant, material spends roughly 20–30 minutes in the rotary kiln after the preheater and before the cooler; the manuscript lists only 16 minutes preheater plus 20 minutes cooler, totaling 37 minutes, which is physically implausible. This is not a disagreement with domain consensus but an internal inconsistency in the method. If the alignment lag is wrong by the kiln dwell time, then every model—especially the PP-only models that support the real-time control claim—is trained on inputs that do not correspond to the output they predict. The two-month held-out period does not fix this, because the same misalignment applies to all samples. The concrete cross-correlation test would settle whether the fixed lag is defensible; if the empirical lag is much longer than 37 minutes, the models must be retrained with the correct lag and re-evaluated on a strictly temporal split before the 'unprecedented real-time prediction' claim can stand. I therefore retain the paper's CONDITIONAL verdict: the direction is plausible and the plant-specific equations are a useful contribution, but the central real-time claim is conditional on correcting the synchronization and re-validating.","tokens_in":26699,"tokens_out":5891,"duration_ms":53635,"concrete_test":"Estimate the true feed-to-clinker lag empirically before any model training: compute cross-correlation between a fast-responding raw process variable (e.g., calciner fuel consumption P14 or kiln-inlet O2 P15, at 1-minute resolution) and clinker alite, testing lags from 0 to 120 minutes. If the peak lag is not near 37 minutes but closer to 60–90 minutes (preheater + kiln + cooler), the synchronization protocol is wrong and the reported MAPE values are not a valid measure of real-time predictive skill.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing premise is the fixed input–output synchronization used to build every training sample. In §2.1 and §3.1 the stated timeline is: kiln feed at t0, 1-minute buffer, 16-minute preheater residence, and 20-minute clinker-cooler retention, giving a total of ~37 minutes from feed to clinker. The rotary kiln—where clinker phases actually form, with a typical material residence time of 20–30 minutes—is missing from this timeline. The 37-minute figure is therefore not merely uncertain; it is internally inconsistent: the aligned input for a clinker sample is shifted earlier by roughly the kiln dwell time. Because input features are then averaged into 2-hour intervals, a 20–30 minute lag error may be partly masked, but it systematically biases the learned mapping, especially during the rapid transients the paper highlights (e.g., 15 wt.% alite swings in a day). The two-month 'unseen' evaluation cannot repair this because the misalignment is in the data construction, not in the train/test split. The random 70/30 split and 19.4% percentile filtering further inflate apparent skill, but the omitted kiln residence time is the more fundamental flaw.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops machine-learning models to predict clinker phase compositions (alite, belite, ferrite) from two years of operational data from an industrial cement plant. Fifteen feature combinations built from process parameters (PP), kiln feed (KF), hot meal (HM), and clinker oxides (CO) are tested across nine regression architectures, with performance measured by MAPE, MAE, and R2. The authors report that the best full-feature models achieve MAPEs of 1.24% (alite), 6.77% (belite), and 2.53% (ferrite), that PP-only models still outperform the plant-specific Bogue equation (e.g., alite MAPE 3.14% vs 7.79%), and that data-driven plant-specific linear equations provide a middle ground between Bogue and black-box ML. SHAP analysis is used to attribute predictions to clinker oxides. The central claim is that ML-based digital twins can enable real-time clinker quality control and outperform conventional Bogue calculations.","tokens_in":26961,"tokens_out":5245,"duration_ms":47754,"significance":"If the reported results are valid, the paper would offer a practically valuable demonstration that data-driven models can predict clinker mineralogy from process data in a real industrial setting, with clear implications for quality control and emissions reduction. Strengths include the unusually large industrial dataset (two years, ~8,654 curated measurements), a systematic comparison across many model architectures and feature sets, a direct benchmark against the industry-standard Bogue equation, and public release of the code. The SHAP analysis provides a useful interpretability layer. However, the significance is currently limited by several evaluation issues that affect the trustworthiness of the headline numbers: the temporal alignment of inputs and outputs omits the rotary kiln residence time; the train/test split is random on an autocorrelated series and is described inconsistently; and the best accuracy figures come from models that include clinker oxides measured on the same sample as the XRD phase labels, which is a post-production mapping rather than a forward prediction. These issues need to be addressed before the claimed 'unprecedented accuracy' can be accepted.","major_comments":[{"comment":"The synchronization timeline is internally inconsistent. The stated sequence is 1-minute buffer, 16-minute preheater residence, and 20-minute cooler retention, totaling ~37 minutes, but the rotary kiln—where clinker phases actually form and which has a typical material residence time of 20–30 minutes—is absent from the timeline. As a result, process and feed inputs are aligned to clinker outputs roughly one kiln dwell time too early. Because every training sample depends on this alignment, all reported metrics (including the two-month 'unseen' evaluation in Fig. 2) are affected. The 2-hour averaging may partially mask the error, but it will systematically bias the learned mapping, particularly during the rapid transients the paper emphasizes (e.g., 15 wt.% daily alite swings). The authors must correct the timeline with plant-specific residence-time data, or at minimum perform a sensitivity analysis over plausible kiln residence times and show that the conclusions are unchanged.","section":"§2.1 and §3.1"},{"comment":"The data-splitting protocol is reported inconsistently and is inappropriate for a time series. Section 2.5 states an 80:20 train-test split, while Section 3.1 states a 70:30 split. More importantly, the random split of a two-year autocorrelated series means that temporally adjacent points fall into both training and test sets, inflating apparent accuracy through information leakage. The two-month January–February 2020 evaluation is a good idea, but it does not repair the leakage from the random split used for model selection and hyperparameter tuning. The authors should use a strictly temporal split (e.g., train on an early contiguous block and test on a later contiguous block) and report the autocorrelation of the target series and of the residuals, so that the reader can judge the effective number of independent test samples.","section":"§2.5 and §3.1"},{"comment":"The headline accuracy (alite MAPE 1.24%) comes from models that include clinker oxides (CO) as inputs. Since CO and the XRD phase labels are measured on the same clinker sample, this is a post-production mapping from XRF chemistry to XRD phase composition—similar in information content to the Bogue equation, which also takes clinker oxides as inputs. Framing this as 'predicting' clinker phases 'while requiring minimal inputs' (Abstract) is misleading; the CO-based models are not usable for real-time process control, as the paper itself acknowledges. The fair comparison to Bogue for forward prediction is the PP-only model (alite MAPE 3.14%), which still shows a meaningful improvement. The paper should restructure the claims so that the post-production XRF-to-XRD mode is clearly separated from the forward predictive mode, and should avoid citing the 1.24% figure as the central achievement.","section":"§3.2.1, Fig. 2, and Table 2"},{"comment":"The 0.01–99.99 percentile filtering removes 2,910 rows (19.4% of the data after earlier cleaning, and about 42% of the raw 14,985 clinker measurements), and it is applied uniformly to all 59 input features and 3 outputs. This approach can systematically exclude the very transients that the model is supposed to predict—the paper highlights 15 wt.% alite swings as a success, but those swings may be underrepresented after filtering. The authors should report model performance without this percentile filter (or with a milder filter) and demonstrate that the removed points are measurement artifacts rather than genuine process variability. At a minimum, the sensitivity of the reported MAPEs to the filtering threshold should be documented.","section":"§2.2 and Table A.5"}],"minor_comments":[{"comment":"The definition of R2 and the auxiliary quantities is garbled. The text defines TSS as the sum of squared differences between predictions and a quantity Äüyý that is itself defined as the mean of squared prediction errors, which is not the mean of the true values. The standard definition should be TSS = Σ(y_t(i) − Ä«y_t)², with Ä«y_t the mean of the true targets.","section":"§2.4, Eq. (3)"},{"comment":"The data-driven clinker equations contain typographical errors that make them unusable as printed. In Case 1, the alite equation is written as Alite = 2.97CaO − 4.5SiO2 − 7.25SiO2 + 0.05Fe2O3, with SiO2 appearing twice; the coefficient −7.25 is almost certainly intended for Al2O3. The corresponding matrix also lists two SiO2 columns. Additionally, Cases 2 and 4 report identical intercept vectors (−166.9, −219.4, −45), which is suspicious because the two cases use different feature sets. These equations should be corrected and re-validated.","section":"Appendix D"},{"comment":"The text in Section 2.3 says 'eight machine learning architectures' and lists linear regression, lasso, elastic net, support vector regression, random forest, XGBoost, neural networks, and Gaussian processes, while Fig. 2 and Table 2 include ridge in addition (nine models total). The count and the model lists should be harmonized.","section":"§2.3 and Fig. 2 caption"},{"comment":"The displayed Shapley value formula is not written correctly: the summation index and the power set notation are confusing, and the expression is missing the necessary subtraction and averaging structure of the standard Shapley formula. Please replace it with a standard, correctly typed version (or refer to the original reference without reproducing a malformed equation).","section":"§2.6, Eq. (4)"},{"comment":"The underlying dataset is proprietary and not available, while the code is public. This limits reproducibility; at least the preprocessing and synchronization steps should be described with enough detail (including exact timestamps, units, and merging rules) that the pipeline could be re-run on a similar plant dataset.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses an important practical problem and has a solid core of model comparison, but the evaluation protocol currently overstates the results. The synchronization omission (rotary kiln residence time) and the random temporal split are load-bearing issues that affect every reported number. I would encourage the editor to ask for a re-analysis with a corrected, plant-validated residence-time model and a strictly temporal split, and to ensure the claims distinguish post-production XRF-to-XRD mapping from forward prediction. Given the data are not public, the authors should provide as much detail as possible on the alignment and filtering steps so that reviewers can assess sensitivity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is an honest attempt to benchmark ML phase prediction against Bogue on a real cement plant, and the plant-specific equations are a nice practical add. But the headline numbers—MAPE 1.24% for alite—are built on inputs that are as \"post-mortem\" as Bogue's, and the input-output timing is off by the rotary kiln dwell time. The real-time claim doesn't stand as written.\n\nWhat's actually new: a two-year, 8,654-sample industrial dataset, which is larger than what's in the cited clinker-prediction papers. The systematic comparison across 15 feature sets, 8 architectures, and the derivation of plant-specific linear equations (Appendix D) is useful and reproducible, even if the equations are only reported, not shipped. The SHAP analysis agrees with known clinker chemistry, which is a good sanity check.\n\nThe soft spots, in rough order of severity:\n\n1. The 37-minute synchronization timeline omits the rotary kiln. They list preheater (16) + clinker cooler (20), but nothing for kiln residence. Actual feed-to-clinker is closer to 60 minutes. Since all input features are aligned to that 37-min assumption, every learned mapping is systematically misaligned. The 2-hour averaging masks part of it, but the claimed skill on 15 wt% alite swings in a day is not believable under this misalignment. This is the load-bearing flaw.\n\n2. The random 70/30 split on an autocorrelated series leaks neighbors into test. The two-month January–February 2020 \"unseen\" period is better, but the data construction error remains.\n\n3. Models that include clinker oxides are not \"predictions\" in the real-time sense—they are a nonlinear version of Bogue, both using post-production XRF on the same sample. The paper admits this in Section 3.2.1, but the abstract and conclusions don't. The PP-only results (3.14% alite MAPE) are the honest headline, and they are still better than Bogue.\n\n4. Minor: split ratio is 80:20 in Section 2.5 and 70:30 in Section 3.1. The 0.01–99.99 percentile filter drops 19.4% of rows; the excluded points are exactly the transients you'd want to predict.\n\nOverall: this deserves a serious referee, but it needs a major revision: re-do the alignment with kiln residence time known (or at least treat lag as uncertain), validate with a temporally blocked split, and reframe the headline around process-parameter inputs. The dataset and the equation results are worth keeping.\n\nWho is this for: process control engineers and ML researchers working on industrial time series. I wouldn't cite it until the evaluation issues are fixed.\n\nRecommendation: send to peer review, but flag the temporal alignment and split as must-fix.","headline":"Useful industrial dataset and a fair Bogue comparison, but the headline accuracy claims are inflated by post-production inputs and a mis-specified time alignment; needs revision before the real-time claim holds.","tokens_in":27516,"tokens_out":3688,"would_cite":false,"duration_ms":32990,"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":"Trained on two years of data from a working cement plant, a machine-learning pipeline predicts the three main clinker mineral phases from process readings, beating the industry-standard Bogue equations used for quality control.","keywords":["machine learning","cement clinker phases","Bogue equation","alite","belite","ferrite","digital twin","process parameters"],"falsifier":"Hold out the entire second year (or any continuous 60-day block), train on the remaining 18 months, and compute the alite MAPE on that held-out block; if it climbs well above 1.24% and toward the 3–4% range, the published accuracy reflects split leakage rather than forecasting skill. A second check: rerun the pipeline with the 19.4% of rows removed as “outliers” restored and see whether the model still tracks the rapid alite swings it is credited with, or whether those transients were excluded precisely so the model would not have to predict them.","tokens_in":26537,"feed_emoji":"🏭","tokens_out":10298,"duration_ms":85936,"temperature":0.7,"pith_summary":"The paper claims that data-driven models trained on two years of operation at an industrial cement plant can predict the three main clinker mineral phases — alite, belite, and ferrite — accurately enough for real-time quality control. With all 59 plant measurements as input, a neural network predicts alite with a mean absolute percentage error of 1.24%, a Gaussian-process model predicts belite at 6.77%, and a support-vector model predicts ferrite at 2.53%; restricting inputs to process parameters alone still beats the Bogue equations, for example alite at 3.14% error versus Bogue's 7.79%. The paper also fits plant-specific linear “clinker equations” from the same data that keep Bogue's pen-and-paper simplicity while roughly halving its error. If these results hold, a cement plant could estimate clinker quality minutes before the clinker is made, reduce off-specification waste, and adjust the process proactively instead of after the fact.","feed_headline":"1.24% error: ML predicts cement clinker phases in real time","feed_subtitle":"Two years of plant data let a neural network beat the 90-year-old Bogue equation used for cement quality control.","key_machinery":"The machinery is the synchronized, cleaned plant archive combined with per-phase model selection. The paper's temporal alignment protocol assigns each clinker composition to process conditions roughly 37 minutes earlier (a 1-minute buffer, 16-minute preheater residence, 20-minute cooler retention, and 20-minute sampling delay) and averages all signals to two-hour windows; after removing duplicates, missing entries, negative values, and values outside the 0.01–99.99 percentile range, 8,654 rows remain. Eight regression architectures are compared on fifteen input-feature combinations, and the best per-phase models are chosen by cross-validation. The Bogue equation — the classical linear recipe that derives clinker phase percentages from oxide content — serves as the baseline to beat; the paper also extracts plant-specific linear equations from the same data as an intermediate tool between Bogue and full black-box models.","core_discovery":"The central discovery the authors claim is that the mineral composition of industrial clinker is learnable from operational records with enough accuracy to replace the Bogue calculation in practice. They assembled a two-year dataset of 8,654 complete measurements after synchronizing process readings to clinker samples through a fixed residence-time chain of about 37 minutes (plus sampling delay), and found that non-linear non-parametric models — a neural network for alite, a Gaussian-process model for belite, and support vector regression for ferrite — capture compositional swings that the linear Bogue equations systematically miss. On a two-month hold-out period the models reduced prediction error several-fold relative to Bogue, and the improvement persisted even when the inputs were limited to process parameters that are available before the clinker is produced. A post-hoc Shapley-value analysis shows the learned drivers match established clinker chemistry — CaO and SiO$_2$ dominate alite, Fe$_2$O$_3$ governs ferrite — which the authors read as evidence the models learned real physical relationships rather than noise.","pith_inferences":["The stated accuracy likely depends on split physics more than the paper reports: clinker is sampled hourly and the series is strongly autocorrelated, so neighboring records land on opposite sides of the random 70/30 split and shadow each other; a time-ordered test (train on the first year, predict the second) would reveal how much of the 1.24% alite error is true forecasting.","The 19.4% of rows dropped as percentile “outliers” are precisely the plant upsets the model is credited with tracking, including the rapid one-day alite swings; retraining on the unfiltered record would show whether the model learned to predict transients or simply had them removed.","If the process-parameters-only result survives a blocked temporal test, it implies that kiln temperature, pressure, fuel, and airflow readings encode most of the phase-formation information, which would guide other plants toward sensor-first data collection without expensive online XRD.","Because the plant-specific linear equations fit from routine XRF/XRD records already beat Bogue, a short campaign of paired measurements at any plant could yield a local, low-cost quality-control formula even before a full machine-learning digital twin is built."],"forward_implications":["Real-time control: a model fed only process parameters and kiln-feed composition can estimate clinker quality before the clinker exists, so operators could adjust the kiln instead of discarding out-of-specification product.","Cheap plant-specific equations: a plant can fit its own linear “clinker equation” from routine XRF/XRD records and immediately beat generic Bogue calculations, with no machine-learning infrastructure required.","Transferable digital twin: because the framework is trained purely on plant records, the same pipeline can be retrained for other plants and extended toward properties such as 28-day strength and alternative fuels and raw materials.","Interpretable chemistry: the learned feature attributions give quantitative, plant-specific rankings of how each oxide drives phase formation, complementing textbook clinker chemistry with measured local relationships."],"supporting_citations":[{"why":"Prior machine-learning prediction of alite and belite from optical clinker images; the accuracy bar this plant-level work extends.","marker":"[4]"},{"why":"Documents the 15-minute to 4-hour XRD measurement delays that motivate real-time phase prediction.","marker":"[5]"},{"why":"Response-surface methodology previously used to model alite, a first-principle baseline the data-driven models outperform.","marker":"[18]"},{"why":"Prior machine-learning clinker quality prediction with model-predictive control design, the application this work targets.","marker":"[21]"},{"why":"Precedent for machine learning on industrial process data at scale, which this paper transfers to cement manufacturing.","marker":"[22]"},{"why":"The Shapley additive explanations framework used to attribute oxide contributions to phase predictions.","marker":"[31]"},{"why":"Gradient-boosted tree method, one of the eight architectures benchmarked for phase prediction.","marker":"[48]"},{"why":"Gaussian-process regression formulation used by the winning belite model.","marker":"[50]"},{"why":"The classical clinkering-reaction calculation behind Bogue's equations, the baseline the paper compares against.","marker":"[51]"}],"fun_headline_variants":["ML beats Bogue equation for cement clinker phase prediction","ML predicts cement clinker phases from two years of plant data","Cement clinker phases predicted in real time with ML","Machine learning replaces 90-year-old Bogue equation for cement quality","Cement clinker phase prediction ML cuts error vs Bogue"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a fixed time delay of about 37 minutes (plus sampling lag) pairs each clinker measurement with its true process inputs, and that a random 70/30 split of the resulting auto-correlated two-year series yields independent training and test samples.","fun_headline_variants_meta":{"raw":{"variants":["ML beats Bogue equation for cement clinker phase prediction","ML predicts cement clinker phases from two years of plant data","Cement clinker phases predicted in real time with ML","Machine learning replaces 90-year-old Bogue equation for cement quality","Cement clinker phase prediction ML cuts error vs Bogue"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000616,"raw_usage":{"total_tokens":2860,"prompt_tokens":942,"completion_tokens":1918,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1833}},"tokens_in":558,"tokens_out":1918,"duration_ms":12186,"temperature":1.0,"reasoning_tokens":1833,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:23:54.729287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out the entire second year (or any continuous 60-day block), train on the remaining 18 months, and compute the alite MAPE on that held-out block; if it climbs well above 1.24% and toward the 3–4% range, the published accuracy reflects split leakage rather than forecasting skill. A second check: rerun the pipeline with the 19.4% of rows removed as “outliers” restored and see whether the model still tracks the rapid alite swings it is credited with, or whether those transients were excluded precisely so the model would not have to predict them.","supporting_citations":[{"cited_title":"Negash, G","cited_arxiv_id":null,"evidence_quote":"Documents the 15-minute to 4-hour XRD measurement delays that motivate real-time phase prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Response-surface methodology previously used to model alite, a first-principle baseline the data-driven models outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Precedent for machine learning on industrial process data at scale, which this paper transfers to cement manufacturing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Shapley additive explanations framework used to attribute oxide contributions to phase predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gaussian-process regression formulation used by the winning belite model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The classical clinkering-reaction calculation behind Bogue's equations, the baseline the paper compares against."}],"review_version":1}