Offshore wind turbine monopile foundations frequently encounter subsurface obstacles such as boulders during installation, with project economics often favouring driving through resistance rather than relocation. This raises critical concerns about pile tip damage that could compromise structural integrity and fatigue life. While Pile Driving Analysis (PDA) provides rapid measurements during installation, no accurate methodology exists to detect or characterize tip damage and deformed shape from these dynamic responses. This work presents a data-driven framework combining finite element simulation with machine learning to enable tip damage detection and geometry reconstruction from standard PDA instrumentation. The approach generates a dictionary of damaged tip configurations using FEA, featuring a three-dimensional model, and considering steel plasticity, and nonlinear soil springs. Each scenario undergoes staged analysis: (1) gravity loading to establish equilibrium at embedment depth, (2) quasi-static loading inducing plastic tip deformations representative of boulder impact, and (3) dynamic analysis capturing the damaged structure's response to continued hammer impacts. Undamaged reference cases, created by omitting the damage step, enable binary classification training. While this quasi-static approach does not capture transient impact dynamics of the damaging phase, it efficiently generates realistic damage geometries and their corresponding dynamic measurement signatures for supervised learning. Two machine learning classifiers trained on strain and acceleration time histories operate sequentially: first detecting damage presence (binary classification), then estimating a probability distribution over dictionary geometries for detected damage cases. The proposed strategy enables rapid assessment during time-critical post-installation evaluation when iterative finite element analysis would be impractical. The site-specific nature of the training data, generated for assumed pile geometries and soil conditions, focuses computational effort on the most relevant damage scenarios while maintaining model fidelity. This minimizes extrapolation in the parameter space, though validation against real damaged structures remains challenging. The present study establishes the computational framework and demonstrates damage scenario generation capability, with experimental validation using laboratory tests from parallel research at NGI planned for future work.