Problems with LNO-CCSD(T)

Dear MRCC developers,
I would like to report some problems I have encountered using core-valence LNO-CCSD(T) for organometallic systems. Please take a look at the attached logs.

The first computation is an all-electron LNO-CCSD(T) for a doublet system. The computation fails with an Intel MKL ERROR.

As a workaround, I attempted to correlate all but two electrons (core equal to 1). This worked for Loose LNO settings. For Normal settings (the second attached log), the computation formally finished; however, please note the (T) energy contribution of approximately -1030 Ha for MO 48. This problem can also be reproduced for an analogue closed-shell system and Tight settings.

MRCC version 25.1.2, precompiled binaries.

Best regards, Alex.

Not sure whether logs were attached correctly.

 
 

Some general words about these computations. DFT reference orbitals are needed because it is very difficult to obtain a stable HF solution for such systems. I also believe that UKS (QRO) works better than ROKS here. The appropriate (for core-valence correlation) pwCVnZ orbital basis sets were chosen. The choice of suitable auxiliary basis sets is a pain, but in my experience, the current setup provides reasonable numbers. I also experienced convergence issues with Jacobi boysalg in some rare cases, so I switched to Newton. Similar computations (RKS and UKS, TZ and QZ, from Loose to Tight LNO) with core eq 15 (or frozen) can be run without any issues.