<p>Fast, data-driven prediction of transport properties of porous media from 2D images is important for materials design and optimization. This work presents prediction of liquid transport in 3D materials using geometric characteristics of pore cross sections. A dataset of 1733 digital structures, of which about a half has impermeable non-percolated systems of isolated pores, was generated using coarse-grained simulations mimicking formation of porous glasses and porous polymer membranes. 2D pore cross sections were characterized by the surface area, perimeter, and asymmetry. Liquid permeability and self-diffusion coefficients of tracers of varying sizes were calculated using lattice Boltzmann and Monte Carlo simulations, correspondingly. On these simulation results, machine learning models were built for (a) distinguishing permeable (percolated) from impermeable (non-percolated) structures and (b) predicting transport properties from pore cross-section characteristics. Very reasonable accuracy was achieved, despite a wide permeability range of the dataset. Gradient boosting method was employed with two-step learning: First, classification model distinguished the permeable samples, and second, a regression model predicted the dynamic properties. The total pore volume and the non-circular shape of the pore cross sections made the strongest influence on the transport properties, but no single feature was dominant. The model trained on synthetic polymer structures was applied to several sandstone samples, with a reasonable predictive accuracy for permeabilities within the training range, and demonstrating the robustness of the proposed approach. In addition, a separate gradient boosting model was trained on a semi-real dataset based on augmented real images, achieving high determination score and confirming the potential of the method for broader range of materials.</p>