pith. sign in

arxiv: 2604.19903 · v1 · submitted 2026-04-21 · 💻 cs.LG · cs.AI· cs.CY

A Multi-Plant Machine Learning Framework for Emission Prediction, Forecasting, and Control in Cement Manufacturing

Sheikh Junaid Fayaz , Nestor D. Montiel-Bohorquez , Wilson Ricardo Leal da Silva , Shashank Bishnoi , Matteo Romano , Manuele Gatti , N. M. Anoop Krishnan This is my paper

Pith reviewed 2026-05-10 02:17 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CY
keywords cement manufacturingNOx emissionsmachine learningemission forecastingsurrogate modelsprocess controlindustrial pollutionSNCR
0
0 comments X

The pith

Machine learning models on data from four cement plants forecast NOx emissions nine minutes ahead and project 34-64 percent reductions at the source.

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

The paper builds machine learning models from operational records across four cement plants to predict and control nitrogen oxide emissions that arise during production. Adding recent process measurements improves forecast accuracy by nearly three times for NOx, showing that emission levels depend on short-term history in a way that does not hold for carbon oxides. The resulting surrogate models then simulate operating adjustments that lower NOx formation before it occurs, reducing the need for downstream chemical treatment. If these projections hold, plants could cut annual NOx output by hundreds of tonnes and lower reagent expenses by tens of thousands of dollars while keeping the same product quality and without installing new hardware.

Core claim

Benchmarking nine machine learning architectures on data from four cement plants shows that prediction error varies three to five times across sites because of differences in data richness. Short-term process history nearly triples NOx prediction accuracy but does not improve CO or CO2 forecasts to the same degree, indicating that NOx formation carries substantial process memory on a timescale of minutes. Models trained this way forecast NOx overshoots as early as nine minutes ahead. Surrogate projections from the best models indicate that source-level control adjustments can reduce NOx by 34-64 percent while preserving clinker quality, which corresponds to roughly 290 tonnes of NOx avoided.

What carries the argument

The multi-plant surrogate modeling framework that trains machine learning architectures on operational variables plus short-term history to predict NOx and to optimize control setpoints that reduce emissions at their origin.

If this is right

  • NOx prediction accuracy increases nearly threefold when short-term process history is included in the models.
  • NOx overshoots can be anticipated up to nine minutes in advance, giving operators time to adjust.
  • Source control guided by the models yields 34-64 percent lower NOx output and corresponding reductions in ammonia reagent use.
  • The approach requires no structural plant modifications or additional hardware.
  • The framework can be applied across multiple plants using only existing operational data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same data-driven surrogate approach could be tested on other high-temperature industrial processes such as steel or glass production if comparable sensor records exist.
  • Coupling the models to automated control loops might allow continuous emission minimization rather than periodic adjustments.
  • Collecting data from additional plants or including more process variables could tighten the uncertainty in the projected savings.

Load-bearing premise

The control actions recommended by the surrogate models will transfer directly to live plant operation without lowering clinker quality or violating other unmodeled constraints.

What would settle it

Apply the surrogate-derived control setpoints to one operating cement plant for several months and measure the resulting NOx levels, ammonia consumption, and clinker quality metrics against the model forecasts.

Figures

Figures reproduced from arXiv: 2604.19903 by Manuele Gatti, Matteo Romano, Nestor D. Montiel-Bohorquez, N. M. Anoop Krishnan, Shashank Bishnoi, Sheikh Junaid Fayaz, Wilson Ricardo Leal da Silva.

Figure 1
Figure 1. Figure 1: Benchmarking nine ML architectures for predicting NO [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: EP vs EPH models for Plant 1. (A) Error vs. process history for the EPH model. (B) Comparison of computational load; circular arcs denote relative magnitudes and do not necessarily sum to 100%. (C–H) Performance of EP and EPH models for predicting NOx, CO, and CO2. Inset histograms in the parity plots show error distributions (ϵ = predicted−actual) on the test set with x-axis limits set to the 99.9% confid… view at source ↗
Figure 3
Figure 3. Figure 3: Auto-regressive forecast of NOx, CO and CO2. (A) Vertical patches indicate temporal gaps in the time-series due from to missing records or pre￾processing. Two major gaps occur during scheduled plant maintenance 13 Jan 2020 - 10 Feb 2020 and 13 Jan 2021 - 10 Feb 2021. (B) Raw and pre-processed NOx data. (C) Data-split strategy for developing the forecasters. 01 Jan 2020 - 30 Aug 2021 for training and 4-fold… view at source ↗
Figure 4
Figure 4. Figure 4: Integration of the NOx controller with plant operation. The controller optimizes seven decision variables (DVs) to minimize NOx emissions using a surrogate ML model (A). The initial DV states are derived from historical plant operation data. Post-hoc model explainability (B-C) ensures that ML-predicted trends are consistent with known process-emission relationships. After performing 35 optimization iterati… view at source ↗
Figure 5
Figure 5. Figure 5: Controller performance across different operating scenarios. A total of 3,761 optimization trials were performed. (A-F) Normal operation trials with initial state NOx ≤ 500 ppm. (A) Comparison of pre vs post–optimization f-CaO (left axis) shows that clinker quality remains unaffected. The right axis shows the relative change in clinker flow (∆ clinker-flow [%]) w.r.t the initial state, confirming process s… view at source ↗
read the original abstract

Cement production is among the largest contributors to industrial air pollution, emitting ~3 Mt NOx/year. The industry-standard mitigation approach, selective non-catalytic reduction (SNCR), exhibits low NH3 utilization efficiency, resulting in operational inefficiencies and increased reagent costs. Here, we develop a data-driven framework for emission control using large-scale operational data from four cement plants worldwide. Benchmarking nine machine learning architectures, we observe that prediction error varies ~3-5x across plants due to variation in data richness. Incorporating short-term process history nearly triples NOx prediction accuracy, revealing that NOx formation carries substantial process memory, a timescale dependence that is absent in CO and CO2. Further, we develop models that forecast NOx overshoots as early as nine minutes, providing a buffer for operational adjustments. The developed framework controls NOx formation at the source, reducing NH3 consumption in downstream SNCR. Surrogate model projections estimate a ~34-64% reduction in NOx while preserving clinker quality, corresponding to a reduction of ~290 t NOx/year and ~58,000 USD/year in NH3 savings. This work establishes a generalizable framework for data-driven emission control, offering a pathway toward low-emission operation without structural modifications or additional hardware, with potential applicability to other hard-to-abate industries such as steel, glass, and lime.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper develops a multi-plant data-driven ML framework for NOx emission prediction, short-term forecasting, and source-level control in cement manufacturing. Using operational data from four global plants, it benchmarks nine ML architectures, shows that incorporating process history nearly triples NOx prediction accuracy (revealing substantial process memory absent in CO/CO2), develops 9-minute-ahead overshoot forecasts, and employs surrogate models to project control policies that reduce NOx formation upstream of SNCR. The headline results are surrogate-estimated 34-64% NOx reductions (~290 t/year) and ~58k USD/year NH3 savings while preserving clinker quality, without hardware changes.

Significance. If the projected control policies transfer to live operation, the work supplies a practical, low-cost pathway for emission reduction in a hard-to-abate sector and demonstrates generalizability across plants. The multi-plant benchmarking, the empirical discovery of NOx process memory, and the forecasting buffer are concrete contributions that could inform similar surrogate-control approaches in steel, glass, or lime production. Reproducible code or open datasets would further strengthen the result.

major comments (2)
  1. [Abstract and §4 (Surrogate Control)] The central control claim (abstract and §4) rests on surrogate-model projections of 34-64% NOx reduction and clinker-quality preservation, yet the manuscript supplies no closed-loop execution of the derived policies, no verification that recommended set-point adjustments remain inside safe operating envelopes, and no live-plant measurements of clinker quality or unmodeled disturbances (e.g., raw-meal variability, ring formation). Without these, the projected ~290 t NOx and $58k NH3 savings cannot be assessed.
  2. [§3.2 and Table 2] Table 2 and §3.2 report 3-5× plant-to-plant variation in prediction error and state that the four-plant dataset supports generalization; however, no cross-plant error bars, leave-one-plant-out ablation, or sensitivity analysis on history-window length are provided. This variation directly affects the reliability of the surrogate policies extrapolated to new plants.
minor comments (2)
  1. [§4] Clarify the exact interface between the surrogate controller and the existing SNCR system (set-point changes vs. direct NH3 dosing) and state the assumed bounds on manipulated variables.
  2. [Figure 3] Figure 3 (forecasting results) would benefit from explicit uncertainty bands and a comparison against a simple persistence baseline.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We are grateful to the referee for providing insightful comments that have helped us improve the clarity and rigor of our manuscript. Below, we respond to each major comment in detail. We have made revisions to the manuscript to address these points to the extent possible given the nature of our study.

read point-by-point responses

  1. Referee: [Abstract and §4 (Surrogate Control)] The central control claim (abstract and §4) rests on surrogate-model projections of 34-64% NOx reduction and clinker-quality preservation, yet the manuscript supplies no closed-loop execution of the derived policies, no verification that recommended set-point adjustments remain inside safe operating envelopes, and no live-plant measurements of clinker quality or unmodeled disturbances (e.g., raw-meal variability, ring formation). Without these, the projected ~290 t NOx and $58k NH3 savings cannot be assessed.

    Authors: We agree that the control results are projections obtained from surrogate models trained and validated on historical operational data, rather than from closed-loop live-plant execution. Performing live interventions on industrial cement plants carries operational risks and falls outside the scope of this retrospective data-driven study. The surrogate models were rigorously validated on hold-out test sets drawn from all four plants and accurately reproduce observed NOx dynamics under varying conditions. In the revised manuscript we have added a new limitations subsection in §4 that explicitly states the assumptions of the surrogate approach, including dependence on historical data distributions. We have also appended an analysis confirming that all recommended set-point adjustments lie within the 5th–95th percentile ranges of historical operating conditions recorded at each plant, thereby providing quantitative evidence that the projections remain inside observed safe envelopes. These changes clarify the scope of the results while preserving the value of the surrogate-based estimates for guiding future pilot validation. revision: partial

  2. Referee: [§3.2 and Table 2] Table 2 and §3.2 report 3-5× plant-to-plant variation in prediction error and state that the four-plant dataset supports generalization; however, no cross-plant error bars, leave-one-plant-out ablation, or sensitivity analysis on history-window length are provided. This variation directly affects the reliability of the surrogate policies extrapolated to new plants.

    Authors: We concur that additional cross-validation strengthens the generalization claims. In the revised manuscript we have updated Table 2 to report error bars as the standard deviation of prediction errors across the four plants. We have also inserted a leave-one-plant-out ablation study in §3.2: models trained on any three plants are evaluated on the held-out fourth plant. The results indicate that multi-plant models retain competitive performance relative to single-plant baselines despite the observed variation, thereby supporting the framework’s applicability to new plants. Finally, we performed a sensitivity analysis on history-window length (5–30 min) and include the outcomes as a new supplementary figure; performance gains from process history plateau beyond approximately 15 min. These additions directly address the reliability of policies extrapolated beyond the training plants. revision: yes

standing simulated objections not resolved
  • Absence of closed-loop live execution and real-time measurements of clinker quality or unmodeled disturbances, which would require direct operational access and intervention at industrial facilities.

Circularity Check

0 steps flagged

No circularity: purely empirical ML framework with independent model projections

full rationale

The paper trains standard ML architectures on historical multi-plant operational data to predict NOx, CO, and CO2, then uses the fitted surrogates to simulate control adjustments that lower NOx while holding clinker quality metrics fixed. These projections are forward applications of the learned mappings rather than redefinitions or statistical re-use of the training targets themselves; no equation reduces the reported 34-64% reduction or the derived annual savings back to quantities defined by the same fitted parameters. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps, and the work contains no derivation chain that collapses to its inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

Abstract-only view limits visibility; the framework rests on standard supervised-learning assumptions plus the domain claim that plant data contain usable predictive signals for control.

free parameters (2)
  • ML architecture hyperparameters
    Nine architectures benchmarked; parameters fitted to each plant's data.
  • History window length and forecast horizon
    Short-term process history and nine-minute horizon chosen to capture NOx memory.
axioms (1)
  • domain assumption Operational sensor data from the four plants are sufficient and representative for cross-plant generalization
    Used to benchmark models and project control performance.

pith-pipeline@v0.9.0 · 5582 in / 1347 out tokens · 59722 ms · 2026-05-10T02:17:37.199337+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [1]

    Deutch, Is net zero carbon 2050 possible?, Joule 4 (11) (2020) 2237–2240

    J. Deutch, Is net zero carbon 2050 possible?, Joule 4 (11) (2020) 2237–2240

    work page 2050

  2. [2]

    C. ECRA, C. CSI, Development of state of the art techniques in cement manufacturing: trying to look ahead, Eur Cem Res Acad (2017)

    work page 2017

  3. [3]

    R. M. Andrew, Global co 2 emissions from cement production, Earth System Science Data 11 (2019) 1675–1710

    work page 2019

  4. [4]

    J. S. Kikstra, Z. R. Nicholls, C. J. Smith, J. Lewis, R. D. Lamboll, E. Byers, M. Sandstad, M. Meinshausen, M. J. Gidden, J. Rogelj, et al., The ipcc sixth assessment report wgiii climate assessment of mitigation pathways: from emissions to global temperatures, Geoscientific Model Development 15 (24) (2022) 9075–9109

    work page 2022

  5. [5]

    Habert, S

    G. Habert, S. A. Miller, V . M. John, J. L. Provis, A. Favier, A. Horvath, K. L. Scrivener, Environmental impacts and decarbonization strategies in the cement and concrete industries, Nature Reviews Earth & Environment 1 (11) (2020) 559–573

    work page 2020

  6. [6]

    Schorcht, I

    F. Schorcht, I. Kourti, B. M. Scalet, S. Roudier, L. D. Sancho, Best available techniques (bat) reference document for the production of cement, lime and magnesium oxide, European Commission Joint Research Centre Institute for Prospective Technological Studies, Luxembourg 506 (2013)

    work page 2013

  7. [7]

    rep., World Health Organization (2021)

    World Health Organization, Who global air quality guidelines, Tech. rep., World Health Organization (2021)

    work page 2021

  8. [8]

    Q. Xu, X. Hao, X. Shi, Z. Zhang, Q. Sun, Y . Di, Control of denitration system in cement calcination process: A novel method of deep neural network model predictive control, Journal of Cleaner Production 332 (2022) 129970

    work page 2022

  9. [9]

    Zheng, W

    J. Zheng, W. Du, Z. Lang, F. Qian, Modeling and optimization of the cement calcination process for reducing no x emission using an improved just-in-time gaussian mixture regression, Industrial & Engineering Chemistry Research 59 (11) (2020) 4987–4999

    work page 2020

  10. [10]

    Glassman, R

    I. Glassman, R. A. Yetter, N. G. Glumac, Combustion, Academic press, 2014

    work page 2014

  11. [11]

    S. R. Turns, et al., Introduction to combustion, V ol. 287, McGraw-Hill Companies New York, NY , USA, 1996

    work page 1996

  12. [12]

    A. I. Okoji, A. N. Anozie, J. A. Omoleye, A. E. Taiwo, D. E. Babatunde, Evaluation of adaptive neuro-fuzzy inference system-genetic algorithm in the prediction and optimization of nox emission in cement precalcining kiln, Environmental Science and Pollution Research 30 (19) (2023) 54835–54845

    work page 2023

  13. [13]

    Guseva, E

    T. Guseva, E. Potapova, I. Tichonova, K. Shchelchkov, Nitrogen oxide emissions reducing in cement production, in: IOP Conference Series: Materials Science and Engineering, V ol. 1083, IOP Publishing, 2021, p. 012083

    work page 2021

  14. [14]

    Y . Li, P. Jiang, Q. She, G. Lin, Research on air pollutant concentration prediction method based on self-adaptive neuro-fuzzy weighted extreme learning machine, Environmental Pollution 241 (2018) 1115–1127

    work page 2018

  15. [15]

    L. Tan, Y . Guo, Z. Liu, P. Feng, Z. Li, An investigation on the catalytic characteristic of no x reduction in scr systems, Journal of the Taiwan Institute of Chemical Engineers 99 (2019) 53–59

    work page 2019

  16. [16]

    L. Tan, Y . Guo, Z. Liu, P. Feng, Z. Li, An investigation on the catalytic characteristic of nox reduction in scr systems, Journal of the Taiwan Institute of Chemical Engineers 99 (2019) 53–59

    work page 2019

  17. [17]

    R. Dong, H. Lu, Y . Yu, Z. Zhang, A feasible process for simultaneous removal of co 2, so 2 and no x in the cement industry by nh 3 scrubbing, Applied Energy 97 (2012) 185–191

    work page 2012

  18. [18]

    Z. Li, X. M. Liu, D. H. Yang, W. J. Qin, G. S. Yang, D. L. Zhang, Research of the sncr process and its application, Advanced Materials Research 953 (2014) 1307–1314

    work page 2014

  19. [19]

    Environmental Protection Agency, Selective noncatalytic reduction (sncr) for nox control: Technical and cost guidance, Tech

    U.S. Environmental Protection Agency, Selective noncatalytic reduction (sncr) for nox control: Technical and cost guidance, Tech. rep., U.S. EPA, chapter 1 (SNCR process description, stoichiometry, NSR and reagent calculations) (2017). URLhttps://www.epa.gov/sites/default/files/2017-12/documents/sncrcostmanualchapter7thedition20162017revisions.pdf

    work page 2017

  20. [20]

    T. K. Hansen, Development of new diesel oxidation and nh 3 slip catalysts, Ph.D. thesis, Technical University of Denmark, Lyngby, Denmark (2017)

    work page 2017

  21. [21]

    X. Hao, Q. Xu, X. Shi, Z. Song, Y . Ji, Z. Zhang, Prediction of nitrogen oxide emission concentration in cement production process: a method of deep belief network with clustering and time series, Environmental Science and Pollution Research 28 (24) (2021) 31689–31703

    work page 2021

  22. [22]

    K. Li, S. Thompson, J. Peng, Modelling and prediction of no x emission in a coal-fired power generation plant, Control Engineering Practice 12 (2004) 707–723

    work page 2004

  23. [23]

    Liukkonen, E

    M. Liukkonen, E. H ¨alikk¨a, T. Hiltunen, Y . Hiltunen, Dynamic soft sensors for nox emissions in a circulating fluidized bed boiler, Applied Energy 97 (2012) 483–490

    work page 2012

  24. [24]

    X. Li, F. Li, S. Zheng, Q. Liu, Nox concentration prediction with a flexible cascaded echo-state network in a cement clinker calcination system, IEEE Transactions on Industrial Informatics 20 (7) (2024) 9644–9654

    work page 2024

  25. [25]

    Usman, I

    M. Usman, I. Ahmad, M. Ahsan, H. Caliskan, Prediction and optimization of emissions in cement manufacturing plant under uncertainty by using artificial intelligence-based surrogate modeling, Environment, Development and Sustainability (2024) 1–24

    work page 2024

  26. [26]

    Glassman, R

    I. Glassman, R. A. Yetter, N. G. Glumac, Combustion, 5th Edition, Academic Press, 2014. 17

    work page 2014

  27. [27]

    J. A. Miller, C. T. Bowman, Mechanism and modeling of nitrogen chemistry in combustion, Progress in Energy and Combustion Science 15 (1989) 287–338

    work page 1989

  28. [28]

    Y . B. Zeldovich, The oxidation of nitrogen in combustion and explosions, Acta Physicochimica URSS 21 (1946) 577–628

    work page 1946

  29. [29]

    Geological Survey, Mineral commodity summaries: Nitrogen (fixed) — ammonia, Tech

    U.S. Geological Survey, Mineral commodity summaries: Nitrogen (fixed) — ammonia, Tech. rep., USGS (2024). URLhttps://pubs.usgs.gov/periodicals/mcs2024/mcs2024-nitrogen.pdf

    work page 2024

  30. [30]

    S&P Global Platts, Ammonia price assessments: U.s. gulf coast and regional benchmarks,https://www.spglobal.com/energy/en/news- research/latest- news/energy- transition/051023- interactive- ammonia- price- chart- natural- gas- feedstock- europe- usgc- black-sea, accessed: 2026-01-11 (2025)

    work page 2026

  31. [31]

    S. J. Fayaz, N. Montiel-Boh ´orquez, S. Bishnoi, M. Romano, M. Gatti, N. A. Krishnan, Industrial-scale prediction of cement clinker phases using machine learning, Communications Engineering 4 (1) (2025) 94

    work page 2025

  32. [32]

    J. J. Monta ˜no Moreno, A. Palmer Pol, A. Ses´e Abad, Using the r-mape index as a resistant measure of forecast accuracy, Psicothema 25 (4) (2013) 500–506. doi:10.7334/psicothema2013.23

    work page doi:10.7334/psicothema2013.23 2013

  33. [33]

    VII. Mathematical contributions to the theory of evolution.—III. Regression, heredity, and panmixia

    K. Pearson, Mathematical contributions to the theory of evolution. iii. regression, heredity, and panmixia, Philosophical Transactions of the Royal Society of London A 187 (1896) 253–318.doi:10.1098/rsta.1896.0007

    work page doi:10.1098/rsta.1896.0007

  34. [34]

    S. M. Lundberg, S.-I. Lee, A unified approach to interpreting model predictions, Advances in neural information processing systems 30 (2017). URLhttps://proceedings.neurips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html

    work page 2017

  35. [35]

    Bansal, S

    A. Bansal, S. J. Fayaz, N. A. Krishnan, S. H. Kota, A. K. Nema, Interpretable machine learning for cost estimation in underground pipe installation, Journal of Pipeline Systems Engineering and Practice 16 (3) (2025) 05025004

    work page 2025

  36. [36]

    K. A. Waris, S. J. Fayaz, A. H. Reddy, B. M. Basha, Pseudo-static slope stability analysis using explainable machine learning techniques, Natural Hazards (2024) 1–33

    work page 2024

  37. [37]

    S. J. Fayaz, M3rg-iitd/industrial scale cement emissions: Prediction, forecast and control (v1.1),https://doi.org/10.5281/zenodo.18888654, zenodo, version 1.1 (2026). Acknowledgements The authors acknowledge the computational resources provided by HPC IIT Delhi, as well as the financial support and assistance with data acquisition from Innovandi–The Globa...

    work page doi:10.5281/zenodo.18888654 2026