In this paper, we construct a deep learning-based neural network (DNN) for scheme-wise investigation of 3D NAND flash memories, with randomization of pulse parameters (PP) including pulse duration (tpulse), step voltage (Vstep), and initial program voltage (VPGM,init). A training dataset is generated by technology computer-aided design simulations which are calibrated with experimental data of three-gate test elements with macaroni configuration, and long short-term memory (LSTM) layers are implemented to train the sequential application of PP. It is shown that a DNN based on LSTM layers provides an accurate prediction for random incremental step pulse programming (ISPP) transient with mean squared error (MSE) of 0.0014, and also provides an accessibility to test various ISPP schemes. An example of a conventional ISPP scheme is introduced with a concept of binary periodic scheme, by constructing a Monte Carlo simulation environment based on deep learning-based results.