Nr and Nrho too high in EAM potential file?

Dear Lammps user,

I'm constructing my own EAM potential files and I recently noticed a strange thing. I can change the number of points in the potential file (Nr and Nrho) and I thought the more points the better precision-wise. I know that Lammps uses a spline of third order to interpolate between the tabled values.

However, at least with one potential and Nr=Nrho=1E6 points I get strange instabilities when calculating elastic constants using /examples/ELASTIC. Some of the values are too high and the elastic tensor does not follow the cubic symmetry. So I thought it must be a numeric instability.

Indeed, decreasing the number of points in EAM file to Nr=Nrho=1E4 solves this instability. It seems, that the problem is with Nr, as Nr=1E4, Nrho=1E6 shows no instability. The potential itself is smooth. Actually I saw the problem with two different potentials.

So now my question, how come that a lot of points in EAM table causes numeric instabilities? And what would be the "sweet spot" for number of points in the EAM table to minimize the interpolation errors and simultaneously avoid the numerical instabilities?

Or is there a problem with the spline interpolation in Lammps, when the tabled values are too close to each other?

The tungsten potential in question has a rcut=4.403A, so for Nr=1E6 dr=4.403E-6A, should not be a problem in double, right? The F, rho and r*phi values are written with 20 digits precision. Not sure, how precise Lammps can read them.

Best regards,