In this article, recent developments in construction of potential energy surfaces using neural networks are reviewed. Based on the previous studies, a systematical method is proposed for iteratively selecting new molecular conformations. The problem for data sampling is now fully resolved under strict standards. A series of potential energy surfaces have been constructed based on highly accurate ab initio energies, using this data selecting scheme and combined with several fitting methods as well as extensive dynamics calculations. Highly reliable dynamics results can be obtained from these surfaces.