Deep reinforcement learning-CFD framework for optimising NO _x emissions in combustion systems

This paper presents a novel Deep Reinforcement Learning (DRL) framework integrated with Computational Fluid Dynamics (CFD) simulations to optimise NOx emissions in natural gas burners. Unlike conventional static models, the proposed DRL-CFD framework employs a multi-layered neural network architecture with ten input features (including turbulence intensity, flame temperature, and air–fuel ratio) and three intermediate processing layers to capture complex nonlinear combustion dynamics. High-fidelity CFD data were generated using COMSOL Multiphysics with the k-ω turbulence model and premixed combustion modelling, representing diverse operational conditions across low, medium, and high NOx emission scenarios. The proposed model achieved a 25% reduction in NOx emissions compared to 15% and 20% reductions by baseline DRL models (DRL 1 and DRL 2, respectively), while maintaining superior predictive accuracy with Mean Absolute Error (MAE) of 0.03 versus 0.05 and 0.04 for baseline models. Through systematic evaluation across 800 training epochs with five-fold cross-validation, our framework demonstrated robust generalisation with R2 values of 0.98, 0.96, and 0.95 for low, medium, and high NOx scenarios respectively, outperforming both baseline models. The model achieves inference speeds of 15 ms per sample, making it computationally feasible for industrial implementation. Key innovations include: (1) integration of multiple combustion parameters for comprehensive control, (2) customised reward functions balancing emission reduction with system stability, and (3) enhanced feature extraction through deeper network architecture. This DRL-CFD framework provides a scalable, data-driven approach for adaptive combustion control, with potential applications across various industrial burner systems seeking to meet stringent environmental regulations while maintaining operational efficiency.