Mean square error （MSE） and mean absolute error （MAE） are usually used to train traditional multi-output deep neural networks in missile aerodynamic coefficient prediction. However， in the case of small sample size and no physical equation constraint， the constraint between MSE and MAE on missile performance coefficient and the distinction between different missile samples will be reduced. A K nearest neighbor large margin （KNNLM） loss function based on deep metric learning is proposed. The method uses the margin constraint to push the output samples with large differences away， and close the similar output samples. Taking the aerodynamic shape and working condition parameters of the missile as input and four aerodynamic coefficients as output， MSE， MAE and KNNLM are used for experimental comparison in backpropagation neural network （BPNN） and multi-task neural network （MTLNN）. The experimental results show that compared with MSE and MAE， KNNLM can improve the accuracy by 14.44% and 16.35% at most， and 3.72% at least in BPNN and MTLNN. And the KNNLM can better distinguish the missile samples in the case of fewer samples and no physical knowledge constraint， so that the prediction accuracy of the deep neural network model is higher and the robustness is stronger.