No HF exchange in 2c-rdgrad(GA)

[januszc]

Hello,

I'm encountering a fatal error during a two-component X2C geometry optimization using jobex -ri with Turbomole v7.9. The initial SCF step (using ridft) converges successfully. However, the subsequent gradient calculation with rdgrad fails.
Key settings in control file:

Code: [Select]

$dft
    functional   b3-lyp
    gridsize   4a
$rij
$marij
$disp3 bj
$rx2c
$rlocal
$finnuc
$gdiis
$soghf
$snso
$snsopara 7
Key cluster run settings:

Code: [Select]

ulimit -s unlimited
export PARA_ARCH=MPI
export MPI_USESRUN=1
The job.1 file shows the following error:

Code: [Select]

X2C 1e gradient contribution done.
 
 setting up bound for integral derivative estimation
 increment for numerical differentiation : 0.00050000
 biggest AO integral is expected to be     16.285438153
 biggest cartesian 1st derivative AO integral is expected to be     55.098967153

========================
 internal module stack:
------------------------
    rdgrad
========================

 fatal error for MPI rank     0

 No HF exchange in 2c-rdgrad(GA)
 rdgrad ended abnormally
error in gradient step (1)
The manual generally indicates that rdgrad supports two-component calculations. I could not find an explicit general limitation stating that HF exchange is unsupported in rdgrad for two-component calculations.
Could the error "No HF exchange in 2c-rdgrad(GA)" point to a limitation in a specific parallel code path (perhaps related to "GA") when using hybrid functionals in 2c gradient calculations? Or is there a known incompatibility with my combination of keywords?

Any advice on how to resolve this or alternative settings would be greatly appreciated.

All the best!

[uwe]

GA is the Global Array library used for the MPI parallelization which was introduced long time ago. Meanwhile this library is not used any more, but the parallelization using MPI still lacks support of some of the newer features of Turbomole. So the 'GA' term is a bit misleading here as it refers to the MPI parallel version.

Please use the parallel SMP version instead of MPI (export PARA_ARCH=SMP) for 2c calculations. It runs on a single node only, but with all cores available there.

[januszc]

Thank you for the clarification. The method works as intended now.