what level of discrepancy do I target for a good interpolation?

I'm performing some interpolation comparison and, of course, the quality of the training sample is a key parameter to survey. In this case I can create the dataset. For this reason, I try to create a good dataset ( = the minimum of sample that help me to have a predictive model).

What is the quantity of experiments required to generate a predictive model? To answer this question I managed to see how the data are nicely sparse in the design space and the discrepancy parameter looks to be a good indicator.

But I have no idea about what is a good value of discrepancy.” Any recommendation ? Any paper / article about that topic ?

a good start is here: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.qmc.discrepancy.html#scipy.stats.qmc.discrepancy