Remarks on Automatic Adjoint Differentiation for gradient descent and models calibration