Cross-Attractor Transformations: A Novel Machine Learning Framework to Minimize Forecast Error in the Presence of Model Bias