If you find that your refinement is unstable and keeps diverging, one option is to use Marquardt damping in the least squares. This can be useful if having difficulties with fitting an initial pattern profile. To enable this, go into the Model, General and under refinement strategy, select the Marquardt Automatic. Be wary that the Marquardt algorithm is guaranteed to converge but may do so in a false minimum. Thus (as always) it is best to take care and be wary of the results. Perhaps starting from several starting points to check the robustness of the refinement.