wrong dhdl with lj-pme
In solvation free energy simulations with gmx 2020.1 the vdW component of dG and DHDL is ~2X overestimated at the lambda region in which vdw-lambdas are changing. Additional testing has shown that this issue is related to "vdw-type = PME".
I have compared (for lambda-state=4) <dVvdw/dl> results of gmx2020 with vdw-type: PME and CutOff, with the result of old gmx5.1 with vdw-type = PME. All was run on CPU for 5 nanosecs with PME electrostatics. It turned out that gmx2020 with vdw-type=CutOff reproduces (within the accuracy) the <dVvdw/dl> result of gmx5.1 with vdw-type = PME, which is expected. In the uniform environment both vdw-types should give accurate solvation FE. The result of gmx2020 with vdw-type=PME, however, is around 2X exaggerated, which seems to be a bug.
Here are the numbers:
Note, the problem is related to DHDL only, all potential energy components (including vdW) in all runs are consistent. Test files are in the attached archive test_solvation.tar.gz. Compare log in lambda.04 folder.