Model architectures

We adopt a U-Net model architecture³⁵ similar to that used by Lehtinen et al.⁹ except that the input and output feature maps are one-dimensional (n = 1 to match monochrome micrographs) and we replace the first two width 3 convolutional layers of Lehtinen et al. with a single width 11 convolutional layer (Supplementary Fig. ¹⁹) similar to other convolutional neural networks used in cryoEM¹⁶. This model contains five max pooling downsampling blocks and five nearest-neighbor upsampling blocks with skip connections between down- and up-sampling blocks at each spatial resolution. We refer to this as the U-net model. For comparison, we also consider a smaller U-net model with only 3 downsampling and upsampling blocks which we refer to as the U-net (small) model. We also compare with a fully convolutional neural network consisting of three convolutional layers of width 11 × 11 with 64 filters each and leaky rectified linear unit activations, termed FCNN, and an affine model with a single convolutional filter of width 31 × 31.