Given a set of known input data X and output data Y, we should predict Y from X. Let Y_analytic be the prediction that the oldfashioned model gives. We take the difference between true and predicted: Y_diff = (Y - Y_analytic) and try to train the neural network to predict Y_diff from X. We name the prediction Y_ann
If the neural network does its job, the sum Y_analytic + Y_ann will be closer to the true Y than what the analytic method alone managed. Supposedly, this "hybrid approach" is better than ANN or a physical model on their own.
Seems sensible, but I've only managed to track down the method in papers about engineering applications, with no references to a thorough theoretical discussion of its advantages and drawbacks. Has anyone come across such a discussion?