Predicting the productivity of a geothermal system is a challenging task due to complex reservoir properties and operational conditions. Numerical simulation is an accurate and widely used method to predict geothermal productivity. However, it typically has low efficiency and is very time consuming. A Long Short-Term Memory (LSTM) neural network is able to accurately and efficiently predict the sequential data. In this study, the feasibility of using a LSTM to predict geothermal productivity is discussed. The geothermal productivity prediction performance of LSTM and Multi-Layer Perceptron (MLP) neural networks are compared. It indicates that an LSTM network can accurately and consistently predict geothermal productivity. The prediction accuracy and stability on the geothermal productivity of the LSTM are better than those of MLP.