{"id":"36d9f91a-d08a-4693-9524-e1b169e70abf","arxiv_id":"2412.07881","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A random forest predicts NOx emissions from machine states on two biochar pyrolysis plants with R2 scores of 0.97 and 0.84, and the fitted model is used as a surrogate for constrained optimization of operating parameters.","lead":"This paper trains a random forest model on sensor and control data from two biochar pyrolysis plants to predict NOx emissions, then uses the model as a surrogate in an optimization to suggest lower-emission operating points. A generalist might care because continuous NOx sensors are often absent on biochar machines, and cheap ML monitoring could help the industry scale while meeting emission limits.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The optimization claim is unvalidated: Fig. 3c is a path through model predictions only, and random forests cannot extrapolate to operating points the optimizer may select.","rationale":"The reader's weakest assumption is that the optimizer's suggestion is trusted to reduce real NOx even though the surrogate may extrapolate poorly outside its training distribution. My stress-test identifies the same load-bearing point: Section 4 reports no physical validation of the optimization path, and Figure 3c is generated from model predictions alone. I also note a supporting technical reason that strengthens the concern: random forests are not extrapolators, so a surrogate-based optimizer can exploit leaf boundaries and out-of-range corners without any physical basis. The additional ambiguity about temporal train/test splitting is relevant because the reported R2 values are the only quantitative evidence that the surrogate is accurate enough to be trusted; if those figures are inflated by autocorrelation leakage, the foundation for the optimization step is weaker still. Both issues are addressable with a clean temporal holdout and an on-machine trial, which is exactly why the appropriate verdict remains CONDITIONAL rather than REJECT. The paper deserves credit for demonstrating the pipeline on two different reactors from different manufacturers, but that demonstration confirms predictive correlation on historical data, not the causal optimization claim that is central to the stated purpose.","tokens_in":4966,"tokens_out":3152,"duration_ms":35920,"concrete_test":"On the PYREG and ARTi datasets, re-run the published pipeline with a blocked time-series evaluation: train on the first 70% of each continuous recording and test on the remaining 30%, with no overlapping timesteps and a purge of at least one hour around the boundary. Report held-out R2 and MAE, and compare against a persistence baseline (predict the last observed NOx) and against a model trained on time-shuffled data to quantify autocorrelation leakage. Then take the optimizer's recommended operating point from Figure 3c, apply it to the reactor for at least one full stabilization period, and compare the measured NOx with the surrogate's prediction and with the baseline operating point. If the held-out R2 collapses toward persistence or the measured NOx at the proposed point is not below baseline, the central surrogate-optimization claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central use case is stated in Section 4 and the Abstract: the Random Forest surrogate should guide a numerical optimization that reduces NOx emissions. Figure 3c shows an optimization path over ARTi data, but it is computed entirely from the surrogate's predictions; no experiment applies the resulting operating point to either the PYREG or the ARTi reactor and measures NOx. This matters because a Random Forest Regressor is piecewise constant and does not extrapolate: outside the training support its prediction is the mean of the training targets in the nearest leaf, and inside a leaf the objective is flat. An optimizer can therefore push inputs to leaf boundaries or out-of-range values that satisfy the O2/CO2 constraints but have no physical justification. The reported R2 values of 0.97 and 0.84 (Section 4) do not close this gap, because they are obtained on historical data without a stated temporal train/test split; since the deployed model was retrained every two hours on recent data, the R2 figures may partly reflect temporal autocorrelation rather than genuine generalization. The two-machine demonstration is a useful positive step, but it supports predictive correlation, not the central claim that the optimizer's recommendations reduce real NOx emissions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Random Forest Regressor to predict NOx emissions from machine states (control settings) in two biochar pyrolysis reactors, PYREG and ARTi. It reports R2 values of 0.97 and 0.84, respectively, and shows an optimization path over the surrogate's predictions to minimize NOx under constraints on O2 and CO2 (Figure 3c). The authors claim that this approach can enable continuous NOx monitoring and surrogate-based process optimization on standard IoT devices. The paper is a short proof-of-concept based on historical sensor data from two plants with different manufacturers.","tokens_in":5211,"tokens_out":5061,"duration_ms":47582,"significance":"If the predictive performance were rigorously established, the paper would provide a useful proof-of-concept for a real industrial gap: continuous NOx monitoring and operating-point optimization in biochar plants without dedicated NOx sensors. The dual-plant demonstration and the IoT deployment framing are positive features, as is the use of a standard, easily interpretable model. However, the evaluation methodology is insufficiently described, and the optimization component is not validated against either physical reactor. As it stands, the paper supports a correlation claim, not the stronger claim that the proposed surrogate can reliably guide emission-reducing control decisions. The contribution is incremental but potentially relevant to the applied machine-learning and process-control communities.","major_comments":[{"comment":"The reported R2 scores (0.97 for PYREG, 0.84 for ARTi) are stated without specifying whether they come from a held-out test set, cross-validation, or the training data. Given that the data are time series and the deployed model is retrained every two hours on recent data (Section 3), random splitting or in-sample evaluation could substantially inflate R2 via temporal autocorrelation. Please describe the exact evaluation protocol (e.g., temporal split, walk-forward validation, or cross-validation with leakage prevention) and report the corresponding R2, RMSE, MAE, and the number of test samples for both reactors.","section":"Section 4, Figures 3a and 3b"},{"comment":"The optimization path is computed entirely from the Random Forest surrogate's predictions; no experiment on the PYREG or ARTi reactor confirms that the suggested operating point actually reduces NOx. Since Random Forest regression is piecewise constant and does not extrapolate reliably outside the training support, the optimizer may select out-of-distribution machine states whose predicted low NOx is an artifact of the model. The manuscript should either provide an experimental validation of the optimized settings (even a brief plant test) or explicitly limit the claim to predictive modeling and describe the optimization as illustrative.","section":"Section 4, Figure 3c"},{"comment":"The relationship between the server-side pre-trained model and the deployed IoT model is ambiguous. The text states that the pre-trained model is simplified to a maximum tree depth of 2 for the IoT device, but it is unclear which model produced the R2 figures in Section 4 and which model was used for the optimization in Figure 3c. Please specify the model version (hyperparameters, training data, evaluation split) for each reported result, and clarify whether the optimized operating points are within the training distribution of the deployed model.","section":"Section 3"},{"comment":"The NOx measurements for the ARTi reactor cover only two days of operation. This limited dataset makes the reported R2 of 0.84 statistically fragile, yet the discussion does not address the number of independent samples, the temporal coverage, or confidence intervals. Please report the sample size, the time span, and the variability of the performance metric (e.g., error bars or a small cross-validation study) to support the generalization claim.","section":"Section 4, ARTi result"}],"minor_comments":[{"comment":"The heading misspells 'PYREG' as 'PRYEG'; please correct this typo.","section":"Section 2, heading"},{"comment":"The sentence 'The optimization was done by constraining the prediction of O2 and CO2 of the Random Forest Regressor should be constrained by 0% < O2 < 10 % and 0% < O2 < 20 %' is grammatically incomplete and appears to contain a typo: the first inequality likely refers to CO2, not O2, matching the caption of Figure 3c.","section":"Section 4, optimization paragraph"},{"comment":"The text says the model was first evaluated using 'consistently available sensor data' and shows a temperature prediction in Figure 2b; please clarify whether this prediction is in-sample or out-of-sample, as this affects the interpretation of the figure.","section":"Section 3, evaluation of Figure 2b"},{"comment":"The caption describes reactor temperature as 'a key feature of how much NOx is produced'; for precision, consider rephrasing to 'a key predictor of NOx production'.","section":"Figure 2b caption"},{"comment":"The reference for the global biochar market report incorrectly merges two entries: 'European Biochar Industry' and 'International Biochar Initiative'; please split or correct the author listing.","section":"References"},{"comment":"The caption says 'we encounter the reader to visit the websites'; this should be 'we encourage the reader'.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a proof-of-concept with a useful application and a real dataset, but the evaluation protocol must be made explicit and the optimization claim needs to be validated or substantially softened. The paper would benefit from a longer-form treatment with a proper temporal train/test split and a discussion of out-of-distribution behavior for random forests. If the authors can provide those elements, a revision would be worth considering; otherwise, the central claims outstrip the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a workshop-quality proof-of-concept, not a benchmark paper. The genuinely new bit is applying a standard random forest to NOx prediction on biochar pyrolysis reactors, and they show it on two machines from different manufacturers. That's a real gap in the literature, and the paper says so plainly.\n\nWhat works: the pipeline is simple and sensible—train on historical states, retrain every two hours on the IoT device, evaluate on a two-week test period. The two demonstrations (R2=0.97 on PYREG, 0.84 on ARTi) are credible as evidence that NOx can be predicted from machine states in this setting. The choice of random forest is justified by prior uses in industrial control, and the caveats about limited ARTi data are honest.\n\nWhere it's soft: the evaluation details are thin. The R2 values are reported without stating whether they come from a held-out time split, cross-validation, or the in-sample fit. With time-series data and a model retrained every two hours, the risk of leakage or autocorrelation inflating the numbers is real. The optimization figure (3c) is the biggest gap: it walks an optimization path over model predictions only, and no one validates that the suggested operating point actually lowers NOx on either machine. Since a random forest is piecewise constant and doesn't extrapolate, the optimizer can find points at leaf boundaries that satisfy the constraints but are physically meaningless. The paper should have either run a confirmation experiment or explicitly labeled the optimization as a hypothetical illustration.\n\nMinor: the constraints in the text are written confusingly (0% < O2 < 10% and 0% < O2 < 20%—presumably one is CO2), and there are typos like 'PRYEG' in the section heading.\n\nOverall, the predictive claim is plausible and the direction is useful, but the optimization claim is not yet demonstrated. The paper is honest about being a 'preliminary test,' and the authors clearly know the difference.\n\nWho this is for: practitioners in biochar or process industries who want a cheap soft sensor. It deserves a serious referee if the venue expects rigorous evaluation, but as a workshop paper it's fine as a poster.\n\nRecommendation: I would accept it for peer review at a workshop or short-paper venue, but for a full journal I'd ask for a proper train/test split, error bars, and an on-machine validation of the optimizer's recommendation.","headline":"A useful proof-of-concept for NOx soft-sensing in biochar plants, with an unvalidated optimization step that should be clearly labeled as illustrative.","tokens_in":5735,"tokens_out":2172,"would_cite":false,"duration_ms":21537,"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 standard Random Forest Regressor trained only on machine states predicts NOx emissions from biochar pyrolysis reactors accurately enough to act as a surrogate in numerical optimization, with an R2 score of 0.97 on one reactor and 0.84…","keywords":["NOx emission prediction","biochar production","pyrolysis reactor","random forest regression","surrogate model","numerical optimization","IoT monitoring","continuous emission monitoring"],"falsifier":"On either physical reactor, take the operating point suggested by the optimizer, hold it long enough for conditions to stabilize, measure NOx with an external reference sensor, and compare to the model's prediction and to the current baseline. If measured NOx does not decrease, or the prediction error at the suggested point is far larger than the training error, the surrogate cannot be trusted to guide real emission reductions.","tokens_in":4811,"feed_emoji":"🌱","tokens_out":7381,"duration_ms":63317,"temperature":0.7,"pith_summary":"Biochar plants are scaling up quickly, but their pyrolysis reactors are hard to control and most do not measure nitrogen oxide (NOx) emissions continuously. This paper argues that a standard Random Forest Regressor—an ensemble of decision trees whose predictions are averaged—can learn the mapping from machine settings such as valve positions, ventilator settings, and oxygen flows to sensor outputs well enough to predict NOx in real time. On two reactors from different manufacturers, the trained model reached R2 values of 0.97 and 0.84. The same model then serves as a surrogate in a numerical optimizer that minimizes predicted NOx while keeping other outputs like CO2 and O2 within user-set bounds. If the claim holds, continuous emission monitoring and operating-point optimization become feasible on ordinary IoT hardware without installing new NOx sensors.","feed_headline":"Random forest predicts biochar NOx at R2 = 0.97 on two reactors","feed_subtitle":"A model trained on machine settings can monitor emissions and guide NOx-reducing optimization on IoT hardware.","key_machinery":"The central object is a Random Forest Regressor, an ensemble of decision trees whose output is the average of the individual trees' predictions. The model is trained with machine states as input features and sensor readings—NOx, CO2, O2, temperature—as target values. It carries the argument by acting as a data-driven surrogate for the reactor: it maps control settings to emissions, allowing a numerical optimizer to search for settings that lower predicted NOx under user-set constraints. On the IoT device, the deployed model is simplified to a maximum tree depth of 2 to fit the device's limited resources.","core_discovery":"The central claim is that the complex physics of a pyrolysis machine can be approximated from operational data by a random forest, and that this approximation is precise enough to stand in for the machine during optimization. The paper demonstrates this by training the model on historical machine states from two commercial reactors, the PYREG and ARTi machines. It reports strong prediction of measured NOx, with R2 = 0.97 on two months of PYREG data and R2 = 0.84 on two days of externally measured ARTi data. Using the model as a surrogate, the optimizer walks the predicted response surface toward lower NOx while keeping CO2 and O2 within user-defined ranges. The authors present the result as a proof of concept: the same pipeline can be retrained every two hours on an IoT device and can later be extended to other target outputs such as biochar yield.","pith_inferences":["The optimizer's suggested point is chosen from model predictions only; until a physical reactor run verifies that emissions actually drop, the predicted optimum could be an artifact of extrapolation.","A natural next experiment is to run the optimized settings on the ARTi or PYREG reactor, measure NOx with an external sensor, and compare against baseline—this would directly test whether the surrogate stays accurate off-distribution.","The difference between the two R2 scores suggests that data quantity and coverage, rather than the method itself, currently limit performance, so collecting more labelled NOx data could close the gap.","Because random forests provide no smooth gradient, the optimizer's path is step-like; incorporating uncertainty estimates or active learning would give more reliable guidance at the edges of the training region."],"forward_implications":["Plants without continuous NOx sensors can get near-real-time NOx estimates from readings they already collect, enabling continuous emission monitoring.","A numerical optimizer can use the surrogate to suggest operating points that minimize predicted NOx while satisfying user-set CO2 and O2 constraints.","Because the pipeline works on two reactors from different manufacturers, the approach is not tied to one machine design.","Retraining the model every two hours on the IoT device keeps it fresh as operating conditions drift.","The same surrogate machinery can be extended to other production goals, such as maximizing biochar yield while controlling emissions."],"supporting_citations":[{"why":"Supplies the random forest method that the paper uses as its predictor.","marker":"Ho, 1995"},{"why":"Documents random forest success in optimizing operational parameters in industrial settings, motivating the method choice.","marker":"Alshraideh et al., 2020"},{"why":"Reports random forest proficiency in controlling and monitoring industrial machinery.","marker":"Chai & Zhao, 2020"},{"why":"Provides the scikit-learn implementation the pipeline is built on.","marker":"Pedregosa et al., 2011"},{"why":"Guides structuring the historical operation data into a flat table format suitable for random forest processing.","marker":"Brownlee, 2020"},{"why":"Identifies the PYREG reactor whose two months of NOx data produce the R2 = 0.97 result.","marker":"PYREG, 2024"},{"why":"Identifies the ARTi reactor used for the two-day NOx measurement and the R2 = 0.84 result.","marker":"ARTi, 2024"}],"fun_headline_variants":["Random forest hits R2 0.97 on biochar NOx","AI model guides biochar NOx reduction","Biochar NOx prediction powers IoT optimization","Machine learning tames biochar plant emissions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the model stays accurate at the new operating points chosen by the optimizer, even though those points can lie outside the historical data it was trained on and no physical test has confirmed that the proposed settings actually reduce NOx.","fun_headline_variants_meta":{"raw":{"variants":["Random forest hits R2 0.97 on biochar NOx","AI model guides biochar NOx reduction","Biochar NOx prediction powers IoT optimization","Machine learning tames biochar plant emissions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1495,"prompt_tokens":942,"completion_tokens":553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":492}},"tokens_in":558,"tokens_out":553,"duration_ms":5613,"temperature":1.0,"reasoning_tokens":492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:26:36.270136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On either physical reactor, take the operating point suggested by the optimizer, hold it long enough for conditions to stabilize, measure NOx with an external reference sensor, and compare to the model's prediction and to the current baseline. If measured NOx does not decrease, or the prediction error at the suggested point is far larger than the training error, the surrogate cannot be trusted to guide real emission reductions.","supporting_citations":[{"cited_title":"Scikit-learn: Machine learning in python","cited_arxiv_id":null,"evidence_quote":"Provides the scikit-learn implementation the pipeline is built on."},{"cited_title":"Process control via random forest classification of profile signals: An application to a tapping process","cited_arxiv_id":null,"evidence_quote":"Documents random forest success in optimizing operational parameters in industrial settings, motivating the method choice."},{"cited_title":"Enhanced random forest with concurrent analysis of static and dynamic nodes for industrial fault classification","cited_arxiv_id":null,"evidence_quote":"Reports random forest proficiency in controlling and monitoring industrial machinery."},{"cited_title":"Time series, 2020","cited_arxiv_id":null,"evidence_quote":"Guides structuring the historical operation data into a flat table format suitable for random forest processing."},{"cited_title":"PYREG GmbH , 2024","cited_arxiv_id":null,"evidence_quote":"Identifies the PYREG reactor whose two months of NOx data produce the R2 = 0.97 result."},{"cited_title":"ARTi , 2024","cited_arxiv_id":null,"evidence_quote":"Identifies the ARTi reactor used for the two-day NOx measurement and the R2 = 0.84 result."}],"review_version":1}