In site seismic characterization, peak ground acceleration (PGA) is an important parameter, along with response spectrum/time history, etc. Expanding strong motion network has been complemented by data sharing through satellite links. PGA is generally estimated using empirical expressions developed from this data (earthquake magnitude, distance from the site, etc.) and these have been updated periodically with fresh records. Certain studies added new factor/s specific to site, soil conditions, etc. for getting improved estimates of PGA. However, as estimation of PGA involves multiple uncertainties, the accuracy achieved by these equations leaves ample scope of improvement. In this article, development of data-driven model/s for estimation of PGA is presented using artificial neural network, using readily available recorded data from the literature. The best ANN models had a correlation between 0.86 and 0.89, mean absolute error between 18 and 25 gals, and root mean square error between 36 and 44 gals, indicating good performance. Out of 32 test data, only 2 (6%) of the estimates fell outside the one standard deviation limits on either side of the perfect fit for Set 1, one (3%) each in case of Set 2 and Set 3, and this can be considered to be a pretty good fit. The case study for Himalayan region in India would demonstrate comprehensive model evaluation followed by an intuitive ranking system to select the best model from a few candidate models.