Preferred normalization of aerodynamic coefficients for ML training on DrivAerML

Determine whether normalizing aerodynamic force and moment coefficients using constant baseline reference quantities—specifically a fixed frontal area A_ref and wheelbase length L_ref for all variants—or using per-geometry reference frontal area and wheelbase for each DrivAer notchback variant is preferable for training machine-learning models on the DrivAerML dataset when constructing targets such as drag, lift, front-lift, rear-lift, and side-force coefficients.

Background

The dataset provides two versions of aerodynamic force and moment coefficients: one normalized with geometry-specific reference quantities (frontal area and wheelbase) for each variant, and another normalized with constant baseline references corresponding to the original DrivAer configuration. This dual provision aims to accommodate different industrial practices and analytical focuses.

Using constant reference area and length emphasizes absolute forces and increases the spread of coefficient values across variants, while geometry-specific normalization focuses on aerodynamic efficiency by separating shape effects from size changes. The authors explicitly note uncertainty about which normalization is better suited for machine-learning applications, motivating a systematic determination of the preferable approach for training and evaluation on DrivAerML.