EVALUATION OF NEURAL NETWORKS FOR DATA-DRIVEN MODELING: AN ENGINE MODEL AS A CASE STUDY

In recent years, Deep Learning has achieved remarkable success across diverse domains, rapidly expanding into both traditional and emerging applications. Data-driven models based on deep architectures are capable of learning complex features from large datasets, enabling faster predictions than conventional software approaches. This thesis investigates the opportunities and advantages arising from the use of such methods to predict outputs and simulate the behavior of complex systems, using an internal combustion engine model as a case study. Feedforward and Long Short-Term Memory neural networks are evaluated on datasets with increasing levels of complexity in order to identify optimal configurations. The model that offers the best performance is then employed as a surrogate model capable of reproducing the dynamics of the original simulation with significantly reduced computational cost. To ensure physical consistency, several constrained learning strategies are explored, including custom loss functions incorporating physics-informed penalties. The results demonstrate improved generalization, stronger adherence to engine dynamics, and the ability of the proposed models to successfully replicate the behavior of the original simulation, highlighting their potential for fast and real-time automotive applications.