Setting limits on the predictions

I’m training a network on the synthetic data that are a decent but not a perfect approximation of the real observations. The input labels have a particular range, but it’s not uncommon for the predictions for the real observations can frequently be outside of that range (e.g. becoming negative, which would not have a physical meaning), and that may be trowing the other parameters off.

Is there a way to force lower and upper limits on the predicted values?