Investigating the interplay of adaptive optimizer hyperparameters on deep neural network predictive fidelity for simultaneous characterization of propane engine combustion, thermal efficiency and emissions

The optimization of propane-fueled spark-ignition engines is crucial for enhancing efficiency and reducing emissions. While deep neural network offers a powerful tool for this task, their performance is highly sensitive to the training hyperparameters of adaptive optimizers like Adam. This study systematically investigates the impact of the Adam optimizer's learning rate and first-moment coefficient on the performance of a deep neural network for predicting combustion, thermal efficiency, and emission characteristics. A high-fidelity engine simulation model, validated against experimental data, was used to generate a comprehensive dataset. The deep neural network's architecture was optimized, and its training was analyzed under varying combinations of learning rate (0.001, 0.01) and first-moment coefficient (0.8–0.95). Results indicate that a higher learning rate significantly accelerates convergence and improves predictive accuracy. The optimal configuration achieved the highest training accuracy of 93.75%, with coefficient of determination values exceeding 0.98 for all outputs. Transient parameters like effective release energy and NOx emissions were more sensitive to first-moment coefficient variation than integral metrics like brake mean effective pressure and thermal efficiency. The analysis demonstrates that adaptive control of these hyperparameters is critical for the deep neural network to resolve complex combustion phenomena accurately, thereby enabling faster calibration, improved energy efficiency, and more precise emission control in spark-ignition engines.