In order to explore the mechanical response characteristics of fractured sandstone under true triaxial different medium principal stresses, matdem particle flow software was used to study the mechanical response characteristics, fracture mechanism and damage evolution characteristics of sandstone specimens under the conditions of 30 MPa, 40 MPa and 50 MPa respectively. The simulation results are verified by true triaxial test. The results show that under true triaxial stress, the increase of medium principal stress is beneficial to increase the strength of sandstone. The fracture degree of the specimen increases with the increase of the intermediate principal stress, and finally the interlacing macroscopic cracks are formed. When the intermediate principal stress is perpendicular to the fracture strike, the fracture mode of sandstone is that the macroscopic fracture plane is perpendicular to the fracture strike, and the fracture mechanism of sandstone under true triaxial compression is mainly shear failure, accompanied by tensile failure. With the increasing of the intermediate principal stress, the fractal dimension of the fracture of sandstone specimen increases significantly and the degree of fracture deepens. Combined with the true triaxial test results, the rationality of particle flow simulation test is proved.

In order to accurately predict it, the maximum subsidence value of the surface of coal pillar replacement fill mining is examined using machine learning and numerical simulation. Exploratory factor analysis (EFA) is utilized to achieve the dimensionality reduction of influencing factors after carefully considering the seven key influencing factors. The findings of the numerical simulation prediction are then compared with the support vector regression machine model (EFA-tSSA-SVR), which is trained on various data sets and optimized using the enhanced sparrow search algorithm (tSSA). The surface subsidence in the first stage is calculated by the numerical model to be 2.00mm, and in the second stage, it is calculated to be 28.00mm, both of which are within 11.10% of the actual amount. The latter forecasts a relative inaccuracy that is about twice as large as the former. The findings show that the EFA-tSSA-SVR optimization model fits well, with an R²close to 1, and a predicted value of 26.40mm that is very close to the measured value of 25.20mm with a 5.90% error. The EFA-tSSA-SVR model can predict the maximum amount of surface sinking, and both models can provide reference values for surface subsidence prediction.