Machine learning tracks prompt NOx in hydrogen-enriched boiler

Machine-learning models built from a small set of simulated operating conditions can estimate prompt nitrogen oxides in a hydrogen-enriched industrial boiler using eight inputs that are available from ordinary plant instrumentation, a study of a 207-megawatt water-tube boiler reports. Under leave-one-out cross-validation, regularized linear regression achieved an R² of 0.815 for prompt NOx without using simulated furnace or exhaust temperatures, while Gaussian process regression achieved 0.788. The same operating variables were much less effective for thermal NOx, revealing a sharp division between an emissions component that could potentially support a soft sensor and one that remains dependent on information unavailable in a real boiler.

The work examined hydrogen and methane combustion in a Saudi Aramco water-tube boiler manufactured by Mitsubishi Heavy Industries, using 24 unique computational fluid dynamics cases. The boiler has two vertically aligned burners, each with eight fuel nozzles, and was modeled at a fixed firing rate of 207 megawatts. Thirteen hydrogen-fraction cases came from an earlier validated CFD study, nine excess-air cases came from a related study using the same model, and three additional hydrogen-fraction cases were generated for the present research. The simulations used ANSYS Fluent 2023 R2 and a 371,000-cell mesh selected after a refinement analysis.

The operating data came from two parameter sweeps. One varied the hydrogen volume fraction from 0% to 100% while holding the excess-air factor, or lambda, at 1.15. The other varied lambda from 1.05 to 1.45 while maintaining a fuel mixture of 90% methane and 10% hydrogen. One case was shared by the sweeps, leaving 24 unique records. The full dataset contained 13 candidate inputs and two outputs: average thermal NOx and average prompt NOx. Prompt NOx was evaluated for the 23 methane-containing cases because its formation through reactions involving hydrocarbon radicals and molecular nitrogen is not defined for the pure-hydrogen case.

Thermal NOx rose sharply after hydrogen was introduced. The simulated average increased from 37 parts per million for pure methane to 1,326 ppm at 1% hydrogen and 1,284 ppm at 2%, before reaching 1,474 ppm in the pure-hydrogen case. The source simulation associated this change with the greater diffusivity of hydrogen, which shifts reactions closer to the burner and raises local temperatures. Thermal NOx formation through the Zeldovich mechanism increases exponentially with temperature. By contrast, prompt NOx in the excess-air sweep declined as lambda rose, falling from 58 ppm at lambda 1.05 to 1 ppm at 1.45 under the fixed 90% methane and 10% hydrogen mixture.

The researchers compared four regression approaches: Ridge regression, support vector regression with a radial basis function kernel, random forest, and Gaussian process regression with a Matérn kernel. Because the dataset was small, every case was held out once in turn while the model was trained on all remaining cases. Input standardization and the calculation of the support-vector penalty parameter were performed only within each training fold, preventing information from the held-out case from leaking into model preparation. The models were implemented in Python and scikit-learn without a hyperparameter search, and the random forest used 300 trees.

Two feature sets were specified in advance for physical rather than target-dependent reasons. The full set included 13 inputs: fuel composition, hydrogen mass fraction, air and fuel flow rates, air-to-fuel ratio, stoichiometry, excess-air factor, mixed-fuel heating value, and maximum and average furnace and exhaust temperatures. The field-measurable set retained eight inputs and removed the four temperatures and constant stoichiometry. The temperatures were computed alongside NOx in the CFD model, but they are not spatial quantities that a plant can directly measure in the same way. Correlation analysis also showed redundancy among fuel composition, fuel flow, air flow, and heating value, although no variables were removed on the basis of those correlations.

For prompt NOx, the best result with the full 13-input set came from Gaussian process regression, with an R² of 0.838, a mean absolute error of 2.53 ppm, and a root-mean-square error of 7.28 ppm. Random forest and Ridge followed with R² values of 0.834 and 0.817. Removing the simulated temperatures did not materially reduce Ridge performance: its R² was 0.815, mean absolute error was 4.35 ppm, and root-mean-square error was 7.77 ppm. Gaussian process regression without temperatures produced an R² of 0.788, a mean absolute error of 4.07 ppm, and a root-mean-square error of 8.32 ppm. Support vector regression improved from an R² of 0.541 with temperatures to 0.728 without them, while random forest declined from 0.834 to 0.773.

Thermal NOx showed the opposite pattern. With the temperature co-outputs included, Ridge regression achieved an R² of 0.742, a mean absolute error of 83.6 ppm, and a root-mean-square error of 153.3 ppm. Gaussian process regression reached an R² of 0.543, while support vector regression and random forest achieved 0.237 and 0.277. Without temperatures, Ridge and support vector regression each fell to an R² of about 0.22, while random forest and Gaussian process regression produced negative R² values. The researchers therefore treated thermal NOx prediction from field-measurable variables as a limitation rather than a deployable result. Even with temperatures, the target was less accurately predicted than prompt NOx.

Feature-ablation tests supported the distinction between the two NOx components. A statistical selection method that retained variables above mean importance, applied separately within each cross-validation fold, gave prompt NOx R² values between 0.723 and 0.847. Selecting variables once over the full dataset before cross-validation overstated thermal NOx performance by about 0.17 in R², illustrating how preprocessing outside the training folds can inflate results. The Gaussian process also supplied uncertainty estimates: its 95% predictive intervals covered 22 of 23 prompt NOx cases, or 96%, with both feature sets. A resampling bootstrap gave a 95% interval of 0.59 to 0.94 for Ridge prompt NOx R² with the 13 inputs and 0.59 to 0.95 with the eight inputs, underscoring the uncertainty associated with a dataset of this size.

The findings suggest that a compact prompt-NOx soft sensor may be feasible within the simulated operating envelope, but they do not demonstrate readiness for industrial deployment. All 24 records were CFD results for a single boiler, and no new direct experimental validation was performed on the exact training cases. The models were not tested beyond the sampled fuel compositions, temperatures, flow conditions, and excess-air range, and their transfer to other boilers or burners remains unknown. The pure-methane case was also isolated from much of the prompt-NOx data and was not reproduced closely. Independent CFD, experimental, or plant measurements would be needed to establish performance under sensor noise, missing inputs, operational drift, and conditions outside the dataset.

Subject of Research: Machine-learning prediction of prompt and thermal nitrogen oxides in a hydrogen-enriched industrial boiler

Article Title: Machine-learning prediction of nitrogen oxides in a hydrogen-enriched industrial boiler

Article References: Haeruman, A., Elshafei, M., Aliyu, M., Nemitallah, M., Alowaifeer, M., Habib, M., & Alzahrani, F. (2026). Machine-learning prediction of nitrogen oxides in a hydrogen-enriched industrial boiler. Results in Engineering, 32, Article 113668. https://doi.org/10.1016/j.rineng.2026.113668

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113668

Keywords: hydrogen combustion, industrial boilers, nitrogen oxides, prompt NOx, thermal NOx, machine learning, soft sensors, CFD modeling, cross-validation, Machine-learning, prediction, nitrogen