The invention discloses a neural network based rock destroy strength determination method. The method is characterized in that a neural network is used to research the rock strength criterion under a uniaxial or triaxial loading condition, various collected rock related data are randomly divided into training and verifying subsets, the compressive strength and the smallest principal stress are used as input values and the largest principal stress value is used as an output value to train the neural network, and the trained neutral network is used to predict largest principal stress value during the test rock destroy. The method mainly comprises a step of experiment data analysis and a step of ANN based largest principal stress value prediction. The method shows the mean square deviation of the neutral network prediction result is reduced by 30-40%, and the determination coefficient is increased by 0.05-0.08 and is more close to 1. The use of the ANN to predict the rock strength has a wide adaptive loading range, so the method is suitable for rock-kind-changeable complex non-linear conditions, and is flexible and accurate.