We show how multigrid preconditioners for the Wilson-clover Dirac operator can be constructed using gauge-equivariant neural networks. For the multigrid solve we employ parallel-transport convolution layers. For the multigrid setup we consider two versions: the standard construction based on the near-null space of the operator and a gauge-equivariant construction using pooling and subsampling layers. We show that both versions eliminate critical slowing down. We also show that transfer learning works and that our approach allows for communication-avoiding algorithms on large machines. In the outlook we discuss how the multigrid setup can be accelerated using gauge-invariant properties of the gauge field.