Warning from loch

[mark-mackey-cresset]

/data/cresset/Flare_12.0.0.somd2/third-party/sire/lib/python3.11/site-packages/loch/_sampler.py:1410: RuntimeWarning: overflow encountered in exp
acc_prob = _np.exp(
/data/cresset/Flare_12.0.0.somd2/third-party/sire/lib/python3.11/site-packages/loch/_sampler.py:1326: RuntimeWarning: overflow encountered in exp
acc_prob = _np.exp(

[lohedges]

Thanks. I think grand has some handling for this like this, i.e accepting if there's an inf, rather than doing the exponential. I haven't seen this for a long time, though, so would be useful to know how frequently you are seeing this, whether it's for a specific system, etc.

[lohedges]

It would also imply that the PME energy change for an accepted RF candidate is infinite, which is strange, i.e. both energy calculations are using the same soft-core, the only difference is to electrostatics in the regular nonbonded force for PME.

[lohedges]

Well, not infinite, but large. It would be good to figure out how big the difference is. I can either catch values that would overflow, or suppress the warning if needed.

[lohedges]

Just to check. Are you using a coulomb_power of zero (the default). This is the only valid setting at present and would be needed for GCMC to work at intermediate lambda.

[mark-mackey-cresset]

I'm not specifying coulomb_power, so it's whatever the default is

[lohedges]

I've created a fix branch where I clamp the exponent to avoid the overflow warning. I'd still like to understand whether this is just a rare occurrence, or a problem for specific systems, which could point to a deeper issue.