Difference in reweighting between plumed 2.9.3 and 2.10.0
Domenico Marson
Dear all,
I am performing some OPES_METAD simulations and I am using the driver workflow
REWEIGHT_BIAS → HISTOGRAM → CONVERT_TO_FES → DUMPGRID.
For reasons I will probably describe in a separate thread (they stem from some issues I encountered with the PLANE action), I ran the reweighting with different PLUMED versions (2.9.3, 2.10.0, and git-master).
What surprised me is that I obtain significantly different FES estimates (and errors) depending on the PLUMED version.
Below is the content of my plumed.dat file for the reweighting step (I used either GRID= or ARG= in DUMPGRID depending on the version):
--- plumed.dat ---
UNITS LENGTH=nm TIME=ns ENERGY=kj/mol
dZ: READ FILE=COLVAR_merged.dat IGNORE_TIME VALUES=dT.z
tiltAvg: READ FILE=COLVAR_merged.dat IGNORE_TIME VALUES=tiltAvg
opes: READ FILE=COLVAR_merged.dat IGNORE_TIME VALUES=opes.bias
weights: REWEIGHT_BIAS TEMP=300 ARG=*.bias
hh1: HISTOGRAM LOGWEIGHTS=weights ARG=dZ.z,tiltAvg ...
GRID_MIN=-5,0 GRID_MAX=5,180
GRID_SPACING=0.2,10 BANDWIDTH=0.2,10
...
ff1: CONVERT_TO_FES TEMP=300 GRID=hh1 MINTOZERO
DUMPGRID GRID=ff1 FILE=fes2d.dZ-tiltAvg.dat
h1: HISTOGRAM LOGWEIGHTS=weights ARG=dZ.z ...
GRID_MIN=-4.8 GRID_MAX=4.8
GRID_SPACING=0.1 BANDWIDTH=0.2
...
h2: HISTOGRAM LOGWEIGHTS=weights ARG=tiltAvg ...
GRID_MIN=0 GRID_MAX=180
GRID_SPACING=5 BANDWIDTH=7.5
...
f1: CONVERT_TO_FES TEMP=300 GRID=h1 MINTOZERO
f2: CONVERT_TO_FES TEMP=300 GRID=h2 MINTOZERO
DUMPGRID GRID=f1 FILE=fes.dZ.dat
DUMPGRID GRID=f2 FILE=fes.tiltAvg.dat
---
Observed behaviour
-
PLUMED 2.9.3 → all output files are produced.
-
PLUMED 2.10.0 → the 1D files are produced (but differ from those obtained with 2.9.3).
The 2D FES gives an error about one point lying outside the grid, even though my GRID_MIN/GRID_MAX boundaries are larger than the explored CV range.
-
PLUMED git-master → immediate segmentation fault.
(This is probably not critical, since git-master is not intended for production runs.)
I am attaching the resulting FES files/plots, and at the following link you can find the compressed COLVAR_merged file (this is a concatenated version of 32 multiple walkers simulations, for each of which I skipped the first 10 % of data).
I did not observe any errors or warnings during compilation for any version.
The behavior is identical on both:
-
my local cluster (gcc 13.3, BLAS and other system libraries from Kubuntu 24.04.3 LTS, Linux kernel 6.14.0-35-generic)
-
CINECA Leonardo (PLUMED compiled with gcc 12.2 and the available libraries).
My question:
Which results should one trust? Both PLUMED 2.9 and 2.10 are described as production releases, so I would normally expect consistent reweighting behavior.
PS: I am aware that the current results are not converged, I was just testing the workflow when I noticed the issue. If anyone also spots any problematic feature directly in the COLVAR file, any comment is welcome.
Best regards,
Domenico