How to separete networks from losses

When improving the loss of a PDE computed using the outputs of multiple networks by training, we would like to make sure that only certain networks are trained. For example, in the turbulent_channel example, the k-ε model is trained to reduce the residuals in addition to the NavierStokes equation. In this example, the network for predicting flow velocity and the networks for predicting k and ε are prepared separately. In my understanding, the residuals of the NavierStokes equation should not be used to train the k or ε network, and similarly, the residuals of the k-ε model should not be used to train the flow velocity network. Is there any way to achieve this in modulus?