Application of Machine Learning to Trajectory Analysis in Aerodynamics

Abstract

In this study, we aim to classify golf balls of different surface specifications using machine learning. We analyze trajectory data obtained from ITR (Indoor Test Range), in which three types of golf balls (C, D, and G) with the different specifications are used. Two balls (C and D) exhibit similar aerodynamic properties, while the third (G) shows rather different characteristics. When the original trajectory data are used as features, the classification accuracy is approximately 60%. Then, we employ the coefficients of the third-order interpolation of the trajectories as features. This results in the accuracy of over 80% for distinguishing between balls C/G and D/G. However, for the balls C and D with similar characteristics, the accuracy remains low, likely due to multicollinearity caused by correlations among the features.

To enhance the classification of balls with similar properties, we apply L1 regularization for dimensionality reduction. Our results show that a model using only three features—spin rate at launching, horizontal and vertical accelerations— can accurately classify the balls C/G and D/G. On the other hand, further improvements are needed to enhance classification accuracy for balls with similar characteristics.