The study of crystal structure in the process of plastic deformation is essential to analyse the properties of materials and help to understand the deformation mechanism. With the rapid development of computer hardware and algorithms, great efforts have been paid for atomic simulation with computer science. However, most existing methods can only identify known structures in advance and consider unknown structure as unrecognized ones. In this work, we propose a method to identify crystal structure among atomistic simulations of crystalline materials. First, we develop a new characterization to describe each atom’s local space information, i.e. local spatial characteristics and utilize mutual information to simplify the characteristics. Then, the multi-layer perceptron neural network is used for the classification on the simplified characteristics. With the proposed method, we can not only identify the crystal structures of the surface of atoms, but also obtain the value of probability of each crystal structure. Our work provides significant insights for finding new mechanisms in structure transformation in material science field.