Finite element (FE) simulations are widely employed to analyze sheet metal forming processes due to their high accuracy; however, achieving sufficiently high temporal resolution to capture the forming dynamics leads to high computational cost, which limits their applicability in real-time control, optimization, and interactive design. Data-driven surrogate models have therefore been introduced as efficient alternatives to FE simulations for complex mechanical systems. Nevertheless, many existing machine-learning-based surrogate models in sheet metal forming either predict only a small number of process responses or rely on direct full-field prediction, which limits scalability. This work presents an integrated surrogate modeling framework for a novel dieless sheet metal forming process employing two servo-driven rotating cones and a crease wheel. High-fidelity FE simulations conducted in LS-DYNA are used to generate time-dependent deformation data. Sequential machine learning models are trained to predict the deformation of a small set of physically meaningful edge nodes based on time-dependent machine parameters, including left and right cone rotation angles. The deformation of a small number of edge nodes, treated as virtual sensors, is predicted using Long Short-Term Memory (LSTM) networks based on time-dependent machine parameters. In parallel, Convolutional Neural Networks (CNNs) are used to predict the bending trajectory of the sheet, providing a compact representation of the forming behavior. To enable full-field deformation reconstruction from sparse edge predictions, Proper Orthogonal Decomposition (POD) is proposed as a reduced-order representation of the sheet deformation, following recent POD-based full-field reconstruction approaches in solid mechanics. In this framework, POD modes extracted from FE displacement snapshots are intended to provide a compact basis for representing the global deformation, while the corresponding modal coefficients can be inferred from the predicted edge-node displacements through a constrained least-squares formulation. This approach defines a clear pathway toward efficient full-field deformation reconstruction with significantly reduced computational cost. The results demonstrate the potential of combining reduced-order modeling and machine learning for efficient surrogate modeling of complex metal forming processes, supporting efficient design evaluation and future integration into digital-twin frameworks for