Fast and reliable inference of gravitational wave source parameters is crucial for identifying short-
lived electromagnetic counterparts and avoiding systematic biases in large event catalogs. Neural
posterior estimation has recently emerged as a powerful inference method, where the model is
trained on simulations at considerable computational cost, and thereafter enables extremely fast
and inexpensive inference at test time. Here, we extend this approach by incorporating domain-
specific physical insights and methods in the model architecture. These include compressing the
data by heterodyning against a reference waveform chosen via approximate likelihood maximization,
removing parameter degeneracies through tailored coordinate systems, and eliminating known mul-
timodalities by folding the parameter space. As a result, the network is approximately equivariant
to changes in the source parameters, and achieves a reduced training cost and improved model inter-
pretability. Our implementation, called labrador,
a can be trained end-to-end on a 1-day timescale
on a hundred CPU-cores and a V100 GPU.