A REINFORCEMENT-LEARNING-AUGMENTED LIQUID-FUELED REACTOR NETWORK MODEL FOR PREDICTING LEAN BLOWOUT IN GAS TURBINE COMBUSTORS

Predicting lean blowout (LBO) in liquid-fueled gas turbine combustors is computationally expensive using high-fidelity Computational Fluid Dynamics (CFD). While Chemical Reactor Networks (CRNs) offer a faster alternative, traditional methods for constructing these networks rely on distance-based clustering (like k-means) that often fails to capture the nonlinear thermochemical sensitivities required to accurately predict extinction limits. This study investigates whether a reinforcement learning (RL) agent can optimize reactor zone partitions to improve LBO prediction accuracy.

The researchers implemented a multi-stage framework. First, they used k-means clustering on CFD data to generate a large set of homogeneous micro-clusters. Second, an actor-critic RL agent was trained to merge these micro-clusters into optimal reactor zones. The RL agent's reward function was explicitly tied to the accuracy of the LBO prediction, allowing the network to adapt its topology based on physical performance rather than simple spatial proximity. The framework was validated using a 119-species Jet-A mechanism in a swirl-stabilized combustor.

The RL-augmented reactor network demonstrated superior predictive fidelity compared to standard k-means clustering. By explicitly accounting for the target metric (LBO accuracy) during the clustering process, the RL agent discovered physically relevant partitions that better captured the extinction behavior of the combustor. This approach provides a computationally efficient reduced-order model that maintains high accuracy, making it a viable tool for rapid design-space exploration where full CFD simulations would be prohibitively slow.

This work bridges the gap between computationally expensive high-fidelity simulations and overly simplified reduced-order models. By introducing a goal-oriented, RL-driven approach to reactor network construction, designers can perform rapid parametric studies of combustor performance near the lean limit without sacrificing the physical insights necessary for reliable design.