The authors compared both pre-modeling bias correction & downscaling and post-modeling bias correction & downscaling to analyze its effectiveness in hydrologic impact-based studies. They found that post-modeling bias correction (bias correcting streamflow) is slightly more effective than bias-correcting precipitation et al.
Using Machine Learning to Cut the Cost of Dynamical Downscaling
The authors used Multilayer Perceptron (MLP) and Random Forest to emulate the performance of a regional climate model and tested the transferrability of the trained model on some independent inputs. They used this ML approach on evaportranspiration only, but has the potential to expand to other variables.