This study investigates the use of artificial neural network (ANN)-based surrogate models for predicting residual velocities in impact problems under varying conditions. Building upon the combination of FEM simulations and incubation time fracture criteria, this paper extends the methodology by exploring multiple ANN architectures tailored to capture complex impact dynamics, including cases involving composite targets and variable impactor velocities. Special attention is given to identifying failure modes of the FEM model. The results demonstrate that appropriately configured ANN surrogates can effectively reduce computational costs and achieve accurate predictions in the case of high-velocity impacts.