AI Combustion Control Tames Variable Fuel Reactivity with GRU-Guided RL

This paper tackles the challenge of controlling combustion phasing (CA50) in multi-fuel compression-ignition engines where fuel reactivity—measured by cetane number (CN)—varies unpredictably cycle-to-cycle. The authors frame the problem as a partially observable sequential decision problem and evaluate controllers of increasing complexity: LinUCB, history-augmented bandits, observation-only DDPG, recurrent DDPG, and their novel GRU-guided RL framework. A Gaussian-process surrogate trained on experimental engine data serves as the evaluation environment.

Results show myopic and fixed-history bandits degrade under CN variation, and observation-only RL suffers from latent-state aliasing. The proposed framework uses a GRU to estimate fuel reactivity from combustion history, then conditions both actor and critic on this estimated signal—not an oracle CN. By training the policy on the same imperfect information available at deployment, the controller avoids inconsistencies in conventional estimate-then-control pipelines. On unseen CN trajectories, it achieves stable CA50 regulation with mean absolute tracking error below 0.25° CA at the training setpoint, with smooth, physically consistent SOI and glow-plug-power actuation.

The work demonstrates that combustion control under latent, continuously evolving fuel dynamics requires more than standalone estimation or generic recurrence. Aligning fuel-reactivity inference with control policy learning creates a reactivity-aware decision-making system that uses the same estimated state available during real-world deployment. This approach could enable more efficient and flexible multi-fuel engines in trucks, generators, and marine applications where fuel quality varies.