Hybrid Physics-Based and Data-Driven Modelling for Vehicle Dynamics Simulation

In the vehicle dynamic sector, the study of phenomena affecting the dynamics of cars is often performed by means of physics-based simulation models. Nowadays, artificial neural networks are largely used as predictive tools thanks to their enormous versatility and efficiency in emulating several system behaviours. In this paper, the implementation of a dynamic vehicle model is investigated, by combining physics-based and Deep Learning subsystem modelling in the Simcenter AMEsim software. Specifically, Artificial Neural Networks are proposed for data-driven modelling of the powertrain sub-system, comprising the engine, the gearbox and the differential, while a physics-based vehicle model is employed to simulate the dynamic behaviour of the vehicle. After trained with a sufficiently large amount of simulated data, the implemented neural network is embedded in a hybrid full vehicle model, in which the powertrain surrogate interacts with the physics-based model of the vehicle remainder in a closed-loop. The accuracy of the hybrid model is investigated by assessing its fidelity in terms of predicted drive torque, engine speed and longitudinal velocity of the vehicle.