The iterative design of active tool surfaces for multi-stage deep drawing is currently bottlenecked by the computational cost of high-fidelity Finite Element (FE) simulations. While accurate, the time-intensive nature of explicit or implicit solvers significantly restricts the exploration of complex tool geometries and delays the optimization of the design prior to prototyping. To address this latency, this work proposes a rapid, geometry-aware surrogate modeling framework capable of identifying "risky" process and design features in real-time, thereby drastically shortening the design feedback loop. We introduce a novel Self-Supervised Learning (SSL) pipeline trained on a comprehensive synthetic dataset generated via an automated, parametric Abaqus simulation workflow. Within this environment, ground truth for failure modes—specifically tearing and wrinkling—is established using material-specific Forming Limit Diagrams (FLD). To effectively capture the nuances of complex tool topologies, we move beyond scalar geometric descriptors by adapting a two-stage structured latent space representation, inspired by the Trellis architecture for 3D generation. This rich geometric embedding is processed via a Sparse Flow Transformer-based encoder, allowing for the dense integration of high-level topological features with standard machine process parameters (e.g., blank holder forces, stroke rates). The primary contribution of this research is the integration of a **Gradient-Based Explainable AI (XAI) module utilizing Fourier coefficient analysis**. Unlike traditional black-box surrogates, this architecture enables the rigorous backtracking of predicted instabilities to their precise origins. By analyzing the spectral properties of the gradients, the framework can pinpoint specific geometric features or local curvatures that predispose the part to failure. This allows for the ex-ante detection of problematic design elements before a full-scale simulation is initiated. The proposed methodology not only facilitates a nearly instantaneous assessment of design feasibility but also provides interpretable, localized feedback to the tool designer, bridging the gap between high-fidelity physics and data-driven design optimization.