Buffer overflow in writecmp when trying to converge geometry optimization to extremely tight tolerances

[tibo...@gmail.com]

Dear developers,

Recently I have been trying to converge some CCSD(T)-F12b geometry optimizations to extremely tight tolerances. To get the best possible numerical gradients and hessians, I have tightened both gthresh,twoint and gthresh,compress to 1e-20, resulting in 100% of the integrals being kept in the compression step. With the VDZ-F12 and VTZ-F12 basis sets, I had no notable issues, but with VQZ-F12 I am getting these buffer overflow errors more or less randomly during the optimization:

 lb=    133125  lenbfi=     65536  lenbf2=     67588  lenbfi+lenbf2=    133124
 Buffer overflow in writecmp
 This is a serious error. Please contact Molpro support
 You can avoid this error using the directive "NOCOMPRESS"
 ? Error
 ? Buffer overflow
 ? The problem occurs in writecmp

Notably, lb is always larger than lenbfi+lenbf2 by 1, suggesting that there might be a classic off-by-one error somewhere in the code. The error message suggests using "NOCOMPRESS", but that leads to the following error after the Opt program starts:

 Generated   8 symmetrical displacements
 ifil= 1  impl= 2
 ? Error
 ? Invalid implementation
 ? The problem occurs in truncatew

Loosening the tolerances to 1e-16, so that a few integrals are neglected in the integral compression algorithm seems to be a workaround for both of these problems.

These errors happen with both Molpro 2023.2 and 2025.1

Best regards,

Tibor Győri