Computational optimization of a vehicle dynamic simulation model

Simulation models in motorsport play a key role in race car design and performance optimization. However, the usability of these models is often limited by their deviation from real-world behavior — the so-called reality gap. Despite the significant progress in data acquisition and simulation technologies, closing the reality gap remains manual and time-consuming. It typically involves iterative tuning of model parameters and comparing simulation outputs with real-world vehicle data. This thesis investigates the major outstanding issues in simulation alignment and explores the adoption of an automated pipeline for parameter tuning to align the simulation model of a race car with real-world data. Usually manual, simulation alignment has heavily relied on expert knowledge and experience, making it hard to automate and reproduce. We propose a method to formalize this domain knowledge using “metrics” and integrate it directly into the automated pipeline. Powered by an optimization algorithm and a virtual driver, the proposed pipeline reduces human presence from the process and improves scalability, reproducibility, and consistency, thereby accelerating the overall race car design cycle. While the manual approach uses a driver-in-loop (DiL) methodology with a human driver, the proposed pipeline replaces this with a virtual driver based on the Data-enabled Predictive Control (DeePC) algorithm and its Fragmented extension, introduced and validated in this thesis. This algorithm learns vehicle dynamics directly from telemetry data, predicting responses to control actions and searching for an optimal control policy. By combining expert domain knowledge with the computational power of optimization algorithms and a data-driven virtual driver, the proposed pipeline enables faster, more consistent, and more scalable vehicle development workflows—moving one step closer to a fully autonomous, optimization-driven virtual engineering ecosystem.