Hi,
I am not familiar with the approach… But I suppose that in the VAR, to obtain what you call ‘‘empirical IRF’’, you do not just hit an exogenous variable… You must have an identification scheme. The identification scheme you consider is different, a priori, from the one embodied in a DSGE model (which comes with more restrictions than the VAR model). But the game is to minimize the gap between the identification schemes by adjusting the values of the DSGE’s parameters.
You may obtain more informations in this thread DSGE based-on IRF from VAR's.
Best,
Stéphane.