Machine learning-enabled uncertainty quantification for thermo-catalytic reactors: a study on fugitive methane oxidation in monolith reactors

Abstract

Ultra-lean methane oxidation via catalytic combustion is critical for mitigating greenhouse gas emissions from fugitive methane sources. However, the catalytic oxidation process exhibits significant uncertainties that hinder its widespread implementation. To address this challenge, the present study develops a robust machine learning-based framework for quantifying combustion uncertainties, enabling more effective emission control strategies. The work presents a novel hybrid methodology integrating polynomial chaos expansion (PCE) with artificial neural networks (ANN), achieving real-time prediction of methane conversion rates and their uncertainties in monolith reactors. The machine learning model reduces computational time from hours to secondswhile achieving excellent agreement with detailed 1D plug-flow reactor simulations. The investigation revealsthat variations in methane concentration (0.2 %–1.3 %, ± 10 %), inlet temperature (800–1000 K, ± 2 %), and inlet velocity (0.8–1.2 m/s, ± 5 %) significantly influence conversion uncertainty, with inlet temperature identified as the dominant parameter (CV ≈ 75 %). Stability improves at elevated temperatures (>950 K) and lower flow velocities (CV ≈ 10 %) compared to higher velocities (CV = 17 %–22 %). Additionally, catalyst deactivation, represented by reduced coating length, decreases methane conversion rates and increases uncertainty, with longer coatings providing greater stability at higher inlet temperatures. This work advances the fundamental understanding of uncertainty propagation in ultra-lean catalytic methane combustion and establishes a generalisable, computationally efficient PCE-ANN framework applicable to catalytic combustion of diverse fuels.

More information

Identifiers

ORCID for Nader Karimi: orcid.org/0000-0002-4559-6245

Catalogue record

Export record

Altmetrics