Neural Network-based model for burn rate prediction in large two-stroke diesel engine for marine applications

Accurate modelling of the combustion process is essential for improving efficiency and performance and to reduce emissions. In recent years, alongside significant advances in artificial intelligence, the use of machine learning for engine design and combustion analysis has grown, with the potential to significantly reduce the number of experimental tests required and, at the same time, the computational costs associated with numerical simulations both 1D models and 3D CFD, which could be very expensive. The objective of this thesis is to develop and validate a model based on artificial neural networks for predicting burn rate curves in a two-stroke marine diesel engine, starting from a selected set of operating variables. To this end, some characteristic engine parameters were initially provided, from which a feature selection procedure was carried out to identify the most representative variables for the purpose of combustion prediction. Subsequently, two parallel approaches were investigated: the first was semi-empirical, relying on the Wiebe function in both single and double forms, where the neural network was trained to estimate the coefficients required to reconstruct the burn rate curves; the second was purely data-driven, capable of directly estimating the burn rate instant by instant based on operating variables and the crank angle. The results showed that, in the case of a single Wiebe function, the neural network was able to correctly estimate the coefficients, but the function itself lacked the flexibility to reproduce the burn rate curves. The double Wiebe model offered a better fit of the experimental curves; however, the larger number of coefficients made it difficult for the network to reproduce them reliably, leading to overfitting and poor generalisation. On the other hand, the data-driven model demonstrated greater flexibility and predictive capability, managing to directly reconstruct the combustion curves with good accuracy. The performances of the data-driven model demonstrate the potential of machine learning algorithms, and in particular neural networks, as an effective tool for burn rate prediction and combustion characterization. This work paves the way for reducing reliance on expensive experimental investigations and computationally intensive simulations, and for the increasingly effective integration of artificial intelligence in the study and design of the engines.