Compared with the traditional electrocatalyst screening of nitrogen reduction reaction (NRR), machine learning (ML) has achieved high-throughput screening with less computational cost. In this paper, 140 TM@g-CxNy single-atom catalysts (SACs) are constructed for NRR. The deep neural network (DNN) classification model and the extreme gradient boosting (XGBoost) model are proposed with 10 features obtained from anchoring TM atom, coordination environment and adsorption intermediates. The former model distinguishes qualified and non-qualified catalysts with an accuracy rate of 87.5% while the latter model predicts free energy of NRR with fitting coefficient of 0.82 on the test set. The N≡N bond length is found to be the most important feature for both models. Moreover, the N≡N bond length, adsorption energy of *N2H () and the number of outermost d electron of TM (Nd) are proved to reflect the degree of nitrogen (N2) activation and serve as NRR descriptors. The moderate activation and half-filled or nearly half-filled d-orbitals of TM atom (Nd≈4) favor NRR process. Among 20 screened catalysts, Mo@g-C4N3 shows the best catalytic activity, with a limiting potential (UL) of only 0.14 V. The activity origin is illustrated by electronic properties and bond changes of NRR intermediates. This research provides a new approach for high-throughput design and screening of SACs by ML based on DFT.