Abstract: Soil, as a porous medium, exhibits significant randomness in pore structure.Existing methods for generating porous media typically yield permeability values that are random, making it challenging to produce random porous media with a specified permeability.To address this issue, this paper proposes a machine learning-based approach capable of efficiently generating random porous media with a target porosity, characteristic pore size, and permeability.A series of numerical models of random porous media were constructed using the truncated Gaussian random field method, where pore structures are quantitatively controlled by two modeling parameters: porosity and characteristic pore size.The permeability of the generated media was computed using the lattice Boltzmann method.Based on 3, 000 datasets describing the relationship between pore structure and permeability, empirical formulas were fitted to predict the upper and lower bounds of permeability of the random porous media.Additionally, a convolutional neural network was trained to accurately predict permeability directly from images of porous media.Results show that the coefficient of determination, mean absolute error, and root mean square error between the predicted and test values are 0.992 2, 0.048 0, and 0.113 0, respectively.Prediction time is approximately 0.08 seconds, and the generation time for a single model is about 1 minute.The proposed method enables efficient and accurate generation of random porous media, providing a powerful tool for seepage simulation.