We address a learning-based quantum error mitigation method, which utilizes deep neural network applied at the postprocessing stage, and study its performance in presence of different types of quantum noises. We concentrate on the simulation of Trotterized dynamics of 2D spin lattice in the regime of high noise, when expectation values of bounded traceless observables are strongly suppressed. By using numerical simulations, we demonstrate a dramatic improvement of data quality for both local weight-1 and weight-2 observables for the depolarizing and inhomogeneous Pauli channels. train-zen.ipynp DNN structure is in train-zen notebook. The number of input (output) neurons is equal to the number of input (output) quantities, such as individual spin magnetizations. It also has three hidden layers each containing 1000 neurons. The sigmoid activation function is used for both input and output layers, while ReLu activation function is used for hidden layers. data12.py dataset generaion train12.py training set generation 6zz_data_16_10000_0.csv set example structure