Artificial Neural Network–Based Modeling of Spark-Ignition Engines to Reduce Experimental Effort in Engine Calibration

The calibration of modern internal combustion engines requires extensive experimental campaigns in order to characterize engine behavior across a wide range of operating conditions. These activities are time-consuming and costly, motivating the development of predictive models capable of supporting or partially replacing experimental testing. In this context, data-driven modeling techniques based on machine learning offer a promising approach for capturing the complex nonlinear relationships between engine control variables and performance indicators. This thesis presents the development of a data-driven model of a spark-ignition internal combustion engine based on Artificial Neural Networks. The objective of this work is to evaluate the capability of neural network models to reproduce engine behavior using experimental data acquired during dedicated test-bench campaigns. The modeling activity is based on a dataset obtained through dynamic spark sweep tests, which allow efficient exploration of the engine operating space while limiting the duration of experimental measurements. The collected data is processed and used to train feedforward neural networks aimed at approximating the relationship between engine control inputs and combustion-related outputs. Particular attention is devoted to the definition of the input and output variables used for training, and to the influence of dataset size and neural network hyperparameters on model performance. The results show that excluding the air–fuel ratio from the input variables can lead to improved predictive accuracy for the considered operating conditions. Furthermore, the analysis of dataset sensitivity indicates that the amount of experimental data required for model training can be significantly reduced, suggesting that comparable model performance can be achieved with a reduction of experimental activity of up to approximately 50–65%. Additional analysis highlight the importance of selecting suitable output variables for the learning process. In particular, the use of indicated mean effective pressure as a training target provides more consistent model behavior compared to using engine torque directly. The study also shows that once appropriate hyperparameters are identified, the training process can be effectively limited to the best network configurations without requiring extensive additional tuning. Overall, the results demonstrate that neural network models can accurately reproduce engine behavior while substantially reducing the amount of experimental data required. The proposed methodology therefore represents a promising tool for supporting engine calibration activities and reducing test-bench development effort.