Model complexity optimisation of multibody dynamical systems for real-time driving simulation with combinatorial Bayesian optimisation

Advanced multibody vehicle models for driving simulators must balance fidelity and real-time capability. Selecting which model components to simplify is a combinatorial problem that grows exponentially with the number of configurable elements. In this work, we apply CBOSS, a constrained combinatorial Bayesian optimisation method, to optimise the structural complexity of a multibody vehicle model in a 50-dimensional binary design space of more than 1015 possible configurations. We propose the Perception Threshold Violation Frequency (PTVF), a metric that quantifies how often model deviations from reference trajectories exceed human perception thresholds. On two use-cases (highway lateral dynamics and ride comfort), the optimiser reduced the computational cost (real-time index) by 21% and 12%, respectively, while keeping perception-relevant deviations below 5% of the simulation time. A subjective validation study with two expert drivers provided moderate Bayesian evidence that drivers could not distinguish the optimised from the high-complexity reference model.