Abstract: Using machine learning models to predict the permeability of porous media is one of the key research directions of current pore-scale models. Since three-dimensional porous media data cannot be directly applied to classical machine learning models, it is necessary to perform feature extraction on the pore s pace structure. Deep learning, as the advancement of classical machine learning models, has achieved many successes in predicting permeability from three-dimensional digital images of porous media, but the computational cost is quite high. This study extracted pore structure features of slices of porous media, converting digital images into multidimensional vectors and serving as input to machine learning models. While reducing the amount of input data and greatly improving the training efficiency, the models maintained excellent prediction performance, where the long short -term memory (LSTM) neural network achieved the best results.

Key words: machine learning, long short-term memory neural network, porous media, permeability prediction, pore
structure features