Understanding the interaction between fracturing fractures and natural fractures (NF) is crucial for estimating fracture complexity and simulated reservoir volume (SRV). For the simulation of complex network models, the general time cost is enormous, therefore, it is very meaningful to consider using machine learning algorithms to predict fracture propagation. On the basis of establishing a fluid structure coupling model using the Cohesive module of ABAQUS and completing indoor experimental verification, this article introduces the global embedded cohesion zone model (CZM) combined with the natural fracture (NF) model (generated in Python) to establish a dispersed mesh model. Based on this model, compare the fracture propagation patterns of hydraulic fracturing and supercritical CO2 fracturing. Design experimental schemes to study the problem of fracture propagation under different engineering and geological conditions. Using machine learning (LSTM-Long Short-Term Memory algorithm) to predict the changes in fracture length under different time and conditions, and comparing it with actual simulation results to optimize the algorithm. The results indicate that supercritical CO2 is more inclined to penetrate natural fractures, leading to the expansion of fractures within the rock matrix. Hydraulic fracturing fractures combine well with natural fractures. The established prediction model can effectively predict the state of fracture propagation. This study combines numerical simulation and machine learning, which is of great significance for enriching the theory of supercritical CO2 fracturing.