Quantifying Accuracy and Robustness to Guide Fidelity Selection for Multi-fidelity Vehicle Dynamics Models

Accurate and reliable simulation models are essential for design, development, and performance evaluation during virtual vehicle testing. However, fidelity assessment and validation remain a challenge. While error metrics are used to evaluate simulations, they alone do not capture how reliable predictions are, or how robust models are to varying driving scenarios and modeling assumptions. This work develops a systematic quantitative approach for evaluating vehicle dynamics model fidelity, moving beyond traditional visual or qualitative comparisons. A dimensionless fidelity metric is proposed that integrates error and uncertainty into a single measure, enabling objective accuracy assessment of variable-fidelity simulations. This framework supports fidelity selection in vehicle dynamics, providing clearer insight into tradeoffs between computational cost and achievable accuracy, and advancing the goal of reliable virtual testing. This approach is demonstrated on an open-loop vehicle dynamics simulation for a compact vehicle under both cruising and evasive maneuvers, comparing a library of multi-fidelity models developed through three approaches: Co-Kriging surrogate models that combine low and high-fidelity data, adaptive fidelity models that switch between fidelities during simulation, and mixed-component fidelity models that integrate low and high-fidelity subsystems. These models were executed under varying operating conditions to evaluate how driving scenarios influence accuracy. The fidelity metric incorporates three components: error relative to the high-fidelity reference, input parameter uncertainty propagated though Monte Carlo sampling, and model form uncertainty measured though Bayesian inference. Error and uncertainty results were combined in a single dimensionless fidelity metric that provides insight into both accuracy and robustness. Demonstration on open-loop driver simulations shows that models with low errors can still exhibit high uncertainty, indicating the importance of considering both error and uncertainty in models. The framework also enables evaluation of parametric scenarios through visual analytics, offering clearer insight into fidelity evaluation, and strengthening the role of virtual testing in vehicle development.