Physics-Guided Machine Learning for Performance Modelling of the Nacra 17 Class

Abstract Polar diagrams, which map wind speed and angle to boat speed, are fundamental tools for sailing performance analysis, race preparation, and coaching. Yet data-driven polar models typically produce point predictions without principled uncertainty, leaving coaches unable to distinguish genuine crew differences from noise or to set statistically grounded training targets. Furthermore, Olympic classes lack manufacturer-supplied polars, so sailors must rely on on-water testing and racing to benchmark against competitors. This paper presents the Sailing Performance Additive Model (SPAM), a 28-parameter model that addresses both prediction quality and uncertainty quantification through a basis design informed by sailing physics. SPAM combines physics-motivated basis functions, including power-law wind scaling, Fourier angular harmonics, and their interactions, fitted via Ridge regression. This architecture admits an exact Bayesian interpretation, yielding closed-form prediction intervals without sampling or simulation. A residual-based noise surface separates model uncertainty from inherent performance variability, making SPAM a coherent probabilistic model. Evaluated on GNSS tracking data from 21 Olympic Nacra 17 crews across 9 international regattas including the Paris 2024 Olympic Games, SPAM captures 95% of learnable variance and achieves 15–18% lower prediction error than Random Forest and GAM baselines when extrapolating to held-out regattas. Practical applications are demonstrated: condition-dependent confidence bands, percentile polars that distinguish a crew’s best achievable performance from their average day on the water, and crew comparison with statistically grounded confidence intervals. Keywords sailing performance; polar diagram; physics-guided machine learning; uncertainty quantification; prediction intervals; Nacra 17; foiling catamaran; Paris 2024 Olympics