Fig. 4 shows the calibrated thickness map reconstructed from the DF tilt series. One can see that the thickness increases gradually from the edge into the sample, verifying that the sample is indeed wedgeshaped, as we would expect. One should also note that, as with any other thickness measuring technique, the thickness estimated this way represents the projected thickness of a potentially tilted slab. Measured normal to the specimen surface the actual thickness of the crystal may differ from the values shown here. However, as described above, in the current experiment the sample was cut in such a way that the nominal specimen surface normal is parallel to the [110] zone axis, so that such projection effects can largely be neglected.
To verify that the final thickness map truly corresponds to the best possible solution we carried out a multiple local minima analysis. The cost-function we are trying to minimize by calculating the best fit can show several local minima. To determine the parameters corresponding to all of these minima by an extensive multidimensional search at each pixel in the data stack is a computationally very expensive procedure and is therefore considered inefficient. However, we did check our results using the MATLAB function MultiStart [27] from the Global Optimization Toolbox. MultiStart runs the local solver for many randomly generated sets of starting points, sampling as many possible basins of attraction as possible. The thickness map presented earlier did indeed correspond to the best solutions found by this global search scheme. We also estimated the thickness of the sample by conventional EELS-based t/λ mapping. Values calculated by both methods are in good agreement with each other