Developing new deep-drawn sheet metal parts is a challenging task due to conflicting demands for low costs, durability, and crash properties. Ensuring manufacturability throughout geometrical changes adds to the complexity, leading engineers to rely on experience-driven iterative design changes that compromise requirements and lack reproducibility [1]. Finite Element (FE) simulation models are employed to ensure manufacturability, albeit at the expense of high computational costs and delays in part development. To improve efficiency, a Machine learning (ML)-centered approach was proposed to ensure manufacturability [2]. However, the limited availability of data raises uncertainty about whether a low- or high-dimensional ML approach is most suitable for drawability assessment [3]. This work compares the accuracy of a low-dimensional, feature-based Linear Support Vector surrogate and an adapted high-dimensional PointNet model [4] under different dataset sizes. The dataset is composed of parametrically generated, U-shaped structural sheet metal parts. We use a one-step simulation scheme and evaluate results with a Forming Limit Diagram (FLD) to label drawability [5]. Results show the point of transition to be at about 500 training samples, from which onwards Deep learning is advantageous. Moreover, the generalizability of these models is tested on a second dataset with topologically similar components. This is to assess the potential for a geometrically more comprehensive evaluation. We discuss several influences on model performances and outline future potentials.