Abstract: Seismic exploration is a non-destructive detection technology widely applied in oil and gas exploration or geological structure exploration. The seismic signals are quite difficult to be interpreted due to the complexity of underground structures and the inhomogeneous media. Limitations of current methods used in seismic signals interpretation and needs for expert knowledge considerably reduce the efficiency of interpretation. We established a nonlinear model using neural network to predict geological information. Firstly, we generated thousands of random layered geological models of media in rock, water and air. Then we got one signal data each model by using finite difference method to simulate the propagation of seismic waves. At last, 80% signals and geological models data are used as training data sets and 20% as test data sets. As the tests show, this neural network accurately predicts signals the depths of underground media interface and velocity of the media between. We didn't use any filtering or feature extraction method in the work above. This method improves the seismic exploration economically and efficiently. The method of machine learning nicely fits in the simplified geological situation in karst area. This research is a preliminary discussion on the feasibility of applying machine learning method into seismic exploration.