Machine learning surrogate model for athlete aerodynamics with application to cycling

Aerodynamic drag dominates the resistive forces in many sports at racing speeds, and small athlete posture changes can produce practically meaningful changes in performance. Standard approaches like field testing, wind-tunnel testing and computational fluid dynamics can be reliable, but time-consuming. Recent advances in scientific machine learning have enabled surrogate models that predict flow quantities at a fraction of the computational cost of a full computational fluid dynamics simulation. However, most applications currently focus on industrial geometries, such as cars or airplanes, where the surface is morphed to generate large datasets for training. It remains unclear how well these methods transfer to athlete geometries where variability is dominated by articulated pose changes rather than smooth shape morphing. In this work, a dataset for cyclist aerodynamics is generated by combining 12 scanned athlete geometries with 20 postures per athlete. Simulations are performed for these 240 geometries, which are then used to train a state-of-the-art surrogate model. An investigation is made on the generalization of the model to an unseen geometry and the balance between number of positions and number of unique geometries in the dataset. The surrogate reproduces the drag area with good agreement, with all predictions within ±10 % and a mean absolute percentage error of 2.69 %. The results also indicate that increasing the number of unique athlete geometries improves performance more effectively than increasing the number of postures per athlete at fixed dataset size. These results support the feasibility of scalable surrogate models for flow over athletes, enabling posture screening and comparative assessment at interactive speeds. This may open up new workflows and new application areas of drag calculations in the domain of sports aerodynamics.