Impact of Ethanol and Methanol on NOx Emissions in Ammonia-Methane Combustion: ReaxFF Simulations and ML-Based Extrapolation

TL;DR

This study uses ReaxFF simulations and machine learning to analyze how ethanol and methanol additives reduce NOx emissions in ammonia-methane combustion, providing insights for cleaner fuel development.

Contribution

It introduces a hybrid ReaxFF-ML framework that predicts NOx emissions and demonstrates alcohol additives' effectiveness in emission suppression at high temperatures.

Findings

Alcohol additives significantly reduce NOx emissions at high temperatures.

ML models, especially Random Forest Regression, accurately predict NOx emissions with errors under 5%.

The framework aids in designing cleaner combustion fuels by combining simulations and predictive modeling.

Abstract

The development of ammonia-methane (NH3-CH4) combustion as a hydrogen-carrier energy source faces major challenges such as significant NOx emissions, hindering its practical implementation. This paper examines how ethanol (C2H6O) and methanol (CH4O) additives influence formation pathways of NOx using ReaxFF molecular dynamics (MD) simulations at temperatures of 2,000 K and 3,000 K. Ten carefully designed fuel mixtures (C1-C10) were evaluated across 0%, 5%, and 10% alcohol concentrations. The findings show that adding alcohol can effectively suppress NOx production, especially at elevated temperatures. At 3,000 K, 10% ethanol addition and 10% methanol addition reduced the production of NOx by approximately 39.6% and 30.1%, respectively, compared with the base fuel. This suppression is attributed to the charge redistribution and the redirection of nitrogen intermediates through stabilising pathways such as HNO, HNO2, and N2O. Simulation-derived descriptors served as the training data for machine learning (ML) models, including Random Forest Regression (RFR), Support Vector Regression (SVR), Gradient Boosting Regression (GBR), and Fully Connected Neural Networks (FCNN). RFR achieved superior performance compared with other models with an R2 of 0.993 and mean absolute error (MAE) of 0.661. The trained ML models successfully predicted NOx emissions for both simulated alcohol ratios (0%, 5% and 12%), and non-simulated alcohol ratios (2%, 7%, and 12%) which demonstrating how hybrid physics-informed ML algorithms can extrapolate complex chemical behaviours, along with prediction errors under 5% for most extrapolated ethanol cases. The results showcase that the ReaxFF informed ML framework successfully serves as a basis for designing cleaner fuels and is capable of establishing a reliable structure for future predictive models in combustion chemistry.