<p>Permeability is a crucial parameter in geology and petroleum engineering. Traditional methods for measuring the permeability of rocks primarily rely on physical experiments, which can easily damage the integrity of rocks, thus affecting subsequent analyses. Current approaches are mainly based on digital rock images using direct numerical simulation methods or deep learning methods to obtain permeability. Although these methods avoid destroying rocks, there are challenges of high computational complexity and capturing complex features on the microstructure for different rocks. To address these challenges, we design a novel network for the lightweight and efficient prediction of permeability in 3D porous media. We made our novel design in two steps: First, to reduce the computational burden, we introduced three-dimensional depthwise separable convolution into the network. Second, to enhance the capture of micro-features in digital rock, we incorporate an efficient multi-scale attention mechanism into the network. Finally, we propose the 3D-EmaSepNet. Experimental results demonstrate that 3D-EmaSepNet performs well on different rock types, such as sandstones and carbonates. Specifically, compared to traditional three-dimensional convolution, the computational cost of 3D-EmaSepNet using three-dimensional depthwise separable convolution is reduced by a factor of 9.6. Our 3D-EmaSepNet achieves coefficient of determination (R² score) of 94.57% and mean squared error (MSE) loss of 0.10 on the validation dataset for sandstone, while reaching R² score of 93.91% and MSE loss of 0.31 on the validation dataset for carbonate. These results show the potential of 3D-EmaSepNet for practical applications in geology, petroleum engineering, and other fields.</p>