Link abstraction is an efficient way for predicting link-level performance (e.g., whether the receiver successfully receives packets through communication links). It can be used to support system-level simulation and facilitate link adaption. Nonetheless, it is challenging to quantify the collective impact of modulation, channel coding, interleaving, guard intervals, and other components on error performance of the communication link. In addition, the existing link abstraction schemes assume ideal hardware with perfect channel estimation. To address this issue, we design a deep learning approach to carry out link-level abstraction for millimeter wave (mmWave) communications. Specially, we conduct link-level simulation based on IEEE 802.11ay single carrier (SC) communication to collect training data, and describe the detailed training and prediction procedures in our approach. We further discuss model reuse and adaptation so that the trained model can be applied to different systems in different environments. Through extensive performance evaluation with data collected on 60 GHz mmWave channels from four environments, we show the prediction accuracy of our approach with respect to feature selection, learning model structure, and the amount of training data. We also demonstrate the efficacy of our approach in model reuse in cross-environments and cross-systems.