Network architecture to retrieve rotation axes. The feature recognition part uses periodic convolutions and the feature combinations part consists of densely connected layers with PReLU activation functions and dropout layers to prevent over-fitting (the number of nodes is indicated above the lower layers). When using polarization, the input shape is 8 ☓ 8 ☓ 12, otherwise 8 ☓ 8 ☓ 6. The periodic convolutions maintain the 1st and 2nd data dimensions since the first N − 1 values along the rotation axis are appended to the end before convolution. The number of filters determines the 3rd dimension of the output. The number of trainable parameters is 667907 (with polarization) and 667619 (without polarization).