ForceBalance is taking very long time over 1 day

[okbckim]

I am running openff-bespoke job for custom torsion fitting. Most of molecules completed pretty quickly but this particular molecule takes very long time in ForceBalance.py. Since optimization config file is probably same for all molecules, I wonder what is special for this molecule. Any suggestions will be very helpful. Thank you!

My molecule is O=C(N1CCN(CC1)C(=O)[C@@h]1NCCCC1)N(C)C in neutral form.

Here is the command I used in openff-beskpoke
openff-bespoke executor run --n-fragmenter-workers 2 --n-optimizer-workers 2 --n-qc-compute-workers 2 --qc-compute-n-cores 1
--file Z1450000036.sdf --output Z1450000036.json --output-force-field Z1450000036.offxml --workflow "default" --default-qc-spec xtb gfn2xtb none

[okbckim]

To update the case, the job is still running. The most up-to-date file in job directory is stage_0/log.txt. Here is the last few lines from the log file.

0.210 s (L-BFGS 1) 0.237 s (L-BFGS 2) Hessian diagonal search: H+0.3002I, length 1.0235e-01, result -4.8901e-02
0.195 s (L-BFGS 1) 0.237 s (L-BFGS 2) Hessian diagonal search: H+0.0002I, length 1.0455e-01, result -4.9804e-02
0.253 s (L-BFGS 1) 0.286 s (L-BFGS 2) Hessian diagonal search: H+0.0000I, length 1.0454e-01, result -4.9714e-02
0.210 s (L-BFGS 1) 0.252 s (L-BFGS 2) Hessian diagonal search: H+0.0472I, length 1.0410e-01, result -4.9703e-02
0.238 s (L-BFGS 1) 0.390 s (L-BFGS 2) Hessian diagonal search: H+0.0122I, length 1.0470e-01, result -4.9801e-02
0.233 s (L-BFGS 1) 0.272 s (L-BFGS 2) Hessian diagonal search: H+0.0038I, length 1.0445e-01, result -5.0017e-02
0.298 s (L-BFGS 1) 0.337 s (L-BFGS 2) Hessian diagonal search: H+0.0038I, length 1.0445e-01, result -4.9840e-02
0.231 s (L-BFGS 1) 0.256 s (L-BFGS 2) Hessian diagonal search: H+0.0064I, length 1.0452e-01, result -4.9926e-02
0.191 s (L-BFGS 1) 0.223 s (L-BFGS 2) Hessian diagonal search: H+0.0050*I, length 1.0470e-01, result -5.0034e-02
0.272 s (L-BFGS 1) 0.303 s (L-BFGS 2)

Thank you.

[leeping]

Hello there,

Thanks for your message. I can see that the optimizer is taking too many iterations in the Hessian diagonal search. Could you try setting search_tolerance 0.1 in your input file (in the general options section) and try again?

The default value of search_tolerance is 1e-4, and you might be able to resolve it with a less aggressive fix than 0.1. If you can get it to work with 0.1, try 0.01 next.

• Lee-Ping

[okbckim]

Thank you for suggestion!

I am running ForceBalance.py through bespokefit workflow. I don't know how to set search_tolerance value in bespokefit workflow. Maybe I can ask how to do it in bespokefit channel unless you have any idea.