Umm.. if you're specific data is now explained well after tuning, but you apply your model to canonical flows and it fails, I would not trust such a tuning as anything beyond the specific dataset (not even for other people's research on related problems, since the tuning has made the model incapable of explaining basic characteristics of experiments upon which these models have been gradually built)