Abnormal dVremain/dl values in GROMACS 2025.0 with vdw-type=pme and free-energy

User discussions
,
1

GROMACS version: 2025.0 (bug) / 2024.4 (normal)
GROMACS modification: No

I found a potential bug in GROMACS 2025.0 related to
free energy calculations with LJ-PME (vdw-type=pme).

Symptom:

  • GROMACS 2024.4: dVremain/dl ~ -700 kJ/mol (normal)
  • GROMACS 2025.0: dVremain/dl ~ 5×10²² kJ/mol (abnormal)

Key MDP settings:

  • vdw-type = pme
  • free-energy = yes
  • couple-lambda0 = none
  • couple-lambda1 = vdw-q
  • couple-intramol = no
  • sc-alpha = 0.5

Full MDP:

integrator = md
nsteps = 200000
dt = 0.002
cutoff-scheme = Verlet
rlist = 2.0
verlet-buffer-tolerance = -1
coulombtype = pme
rcoulomb = 1.2
vdw-type = pme
rvdw = 1.2
dispcorr = no
constraints = all-bonds
lincs-order = 12
tcoupl = v-rescale
tc-grps = system
tau-t = 0.2
ref-t = 310
pcoupl = parrinello-rahman
ref-p = 1
compressibility = 4.5e-5
tau-p = 5
free-energy = yes
couple-moltype = MOLTYPE
couple-lambda0 = none
couple-lambda1 = vdw-q
couple-intramol = no
init-lambda-state = LAMBDA
fep-lambdas = 0.0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85
0.9 0.95 1.0
sc-alpha = 0.5
sc-power = 1
sc-sigma = 0.3
sc-coul = yes
calc-lambda-neighbors = 1
nstdhdl = 50

Is this a known bug in 2025.0?
Does anyone have a workaround?

Thank you.

2

No, this is not a known issue.

At which lambda(s) does this occur? Are you using GPUs?

Have you tried 2025.4 and/or 2026.1?

3

Thank you for your response.

This occurs at all lambda windows (lambda 0 to 20).

Regarding GPU: Yes, we are using NVIDIA GPUs.

Regarding newer versions:

  • GROMACS 2024.4: dVremain/dl ~ -700 kJ/mol (normal)
  • GROMACS 2025.0: dVremain/dl ~ 5×10²² kJ/mol (abnormal)
  • GROMACS 2025.4: dVremain/dl ~ 5×10²² kJ/mol (abnormal)

We have not tested 2026.1 yet.

Is this bug present in 2026.1 as well?
Any workaround would be appreciated.

Thank you.

4

I have no clue what could be wrong here. I don’t see any changes in the free-energy code between the 2024 and 2025 releases. Could you run one lambda value simulation with -nb cpu to check if this is GPU related?