A Reduced Model is essentially a black box with a number of inputs and a number of outputs which as been constructed, or trained, using a full COMSOL model so as to return the same output values as the full model for inputs within some allowed range. During the training phase, the model reduction process must be able to evaluate the dependency of outputs on inputs. Such dependency evaluations can only be performed in general with respect to degrees of freedom in the model.

Therefore, any quantity that you want to use as input to a Reduced Model must be defined as a degree of freedom during the training phase. And it must also evaluate to a constant training value, independent of time or frequency. At the same time, it should behave like a normal user-defined variable when the full model is computed or when the Reduced Model is called online. For example, it can be time- or frequency-dependent in that situation.

Reduced model inputs are special variables which have these desired properties: they are defined by an online expression during normal computations, where they can be used exactly like any other user-defined variable. But they are internally redefined as degrees of freedom when necessary for training a Reduced Model. The Model Reduction study step (see Model Reduction) allows only such predefined reduced model inputs as inputs to the training process.