The invention discloses an optimal feature vector design method for carrying out reverse photolysis based on machine learning. The method comprises the following steps: dividing a design target pattern into a plurality of grid units; calculating a characteristic function set {Ki (x, y)} according to imaging conditions, i = 1, 2,... N1; establishing a neural network model, and selecting a trainingsample and a verification sample which need to be included in training; calculating a signal set {Si (x, y)} of each grid unit by using a characteristic function set {Ki (x, y)}; taking the value of the strict reverse photoetching at the corresponding position as a target value of neural network training; training, different input end dimensions N1, the number N2 of hidden layers and the number M1of neurons of each hidden layer are adopted; training the neural network model according to different combinations of the input end dimension N1, the hidden layer number N2 and the neuron number M1,M2,... MN2 by using a training sample, and verifying the neural network model by using a verification sample until the neural network model with satisfactory combinations of the input end dimension N1, the hidden layer number N2 and the neuron number M1, M2,... MN2 of each hidden layer is found. Therefore, according to the design method, the neural network does not need a feature extraction layerany more, so that the network architecture is simplified, and the training time is shortened.