The projection of extreme convective precipitation by global climate models (GCM) exhibits significant uncertainty due to the coarse resolution of GCMs, which cannot resolve fine-scale processes. Direct dynamical downscaling (DDD) of regional climate at convection-permitting resolutions (~1 km) provides valuable insight into the potential changes in extreme precipitation, but its computational expense is formidable. Here we document the effectiveness of machine learning in enabling smart dynamical downscaling (SDD), which performs downscaling only for a small subset of GCM data. Trained with reanalysis and satellite data for three Asian cities, support vector machines can filter out approximately 87% to 94% of circulation data, which are irrelevant to extremes. Deep convolutional neural networks, trained with larger data sets, can filter out more than 97% of circulation data and in the selected subset, retrieve 72% to 81% of the circulation patterns responsible for extreme events (rain intensity higher than the 99th percentile).