One crucial aspect that needs to be considered when studying the behaviour of historic masonry structures under earthquake ground excitation is the selection of appropriate intensity measures that can accurately capture the characteristics of seismic shaking. Among various features of ground motion, the cumulative absolute velocity (CAV) and peak ground velocity (PGV) have been identified as crucial factors affecting the response of masonry structures. These measures account for the amplitude, frequency content, and duration of ground motion, which can induce significant damage to masonry structures. These parameters can be assessed using either parametric linear or nonlinear regression techniques or non-parametric methods such as sophisticated machine-learning algorithms. However, parametric models, which are based on specific mathematical formulations, are prone to significant bias. In this study, we employ machine learning techniques that are more flexible and can capture the complex behaviour of ground motions, making them suitable for regions with high seismicity or geological complexity. Particularly, we employ the backpropagation neural network (BPNN) to develop models of CAV and PGV by considering an extensive dataset of Italian strong motions. A bias-variance trade-off was considered by tunning hyper parameters of the machine to ensure accurate predictions for future unseen data. Overall, the developed models demonstrate promising results compared to the prior models in the literature.