Ballistic impact in carbon-fibre-reinforced polymer laminates involves interacting damage mechanisms whose relative contributions change with laminate thickness. Exploring the parameters governing these mechanisms through finite-element simulations is computationally demanding. Machine-learning surrogates have been applied to composite impact and terminal-ballistics prediction; their use in constitutive calibration also requires understanding how they represent the influence of damage parameters. This study compares multilayer perceptron, Gaussian-process, support-vector and gradient-boosted-tree regressors trained separately for cylindrical-projectile impact on 8-, 16- and 32-ply laminates. The finite-element configurations draw on published impact experiments. Each surrogate maps nine progressive failure parameters in LS-DYNA’s MAT_054 model to the projectile’s residual velocity. Learning curves, common test cases and permutation importance assess predictive accuracy, the benefit of additional simulations and consistency in the learned parameter dependence. Support-vector regression achieves the lowest test root-mean-square error for 8 plies, while Gaussian-process regression performs best for 16 and 32 plies. Adding a second 32-ply simulation campaign improves all four families and reduces Gaussian-process test error by 26.5% on the same 80 test cases. Despite their different prediction errors, all families identify effective failure strain as the leading input for 8 plies and retained fibre-compressive stress for 16 and 32 plies. This shift is consistent with the expected change in impact response as increasing laminate stiffness alters the balance between flexural deformation and local contact loading. These constitutive controls affect the work opposing penetration and the energy remaining for projectile exit. The results identify suitable forward surrogates and provide a mechanically consistent interpretation of parameter dependence within the investigated configurations, establishing a basis for subsequent inverse calibration.