Neural Network-Enhanced Brake Thermal Modeling for Formula 1 Application

Accurate prediction of brake temperature is a critical requirement in Formula 1, where thermal management directly affects braking performance, component durability, and aerodynamic efficiency. This thesis addresses the development of a fast and reliable predictive framework for brake disc thermal behavior under race-representative operating conditions. The work is structured in two stages. First, the existing physics-based model is analyzed and improved through refined thermal formulation and implementation choices aimed at increasing predictive fidelity while preserving computational efficiency and compatibility with established team workflows. Second, machine learning strategies are investigated to capture residual and unmodeled effects that are difficult to represent explicitly in a reduced-order physical model. The data-driven investigation considers pure neural approaches and hybrid alternatives, with particular focus on residual-learning architectures that correct physics-model outputs using track-derived signals. The modeling pipeline includes event-based data partitioning across the 2025 championship season, robust preprocessing and normalization procedures, and validation protocols designed to assess generalization across circuits and operating conditions. Results indicate that the improved physics-based model provides a stronger baseline and that hybrid ML integration offers additional gains in scenarios affected by complex external influences (e.g., traffic and environment variability), while maintaining deployment feasibility for time-constrained engineering operations. The proposed framework demonstrates the practical value of combining physically grounded models with neural components for high-performance motorsport thermal prediction.