Negative information entropy in free energy calculations

GROMACS version: 2016.1
GROMACS modification: No

Dear all,
we are calculating free energy of solvation (in water and octanol) by FEP method on several solutes (e.g. diphenyl esther and Pd(P-Ph3)4).

We used the approach described in the tutorial at the following link

Sometimes we get negative information entropy when decoupling electrostatics interactions. As an example:
lam_A lam_B DG +/- s_A +/- s_B +/- stdev +/-
0 1 0.5958 0.0643 0.2816 0.0495 0.2802 0.0487 0.4204 0.0228
1 2 0.3046 0.0312 0.0109 0.0631 0.0067 0.0631 0.2831 0.0134
2 3 0.3344 0.0699 -0.0364 0.0444 -0.0403 0.0441 0.2928 0.0309
3 4 0.3028 0.0461 0.0719 0.0631 0.0679 0.0631 0.2799 0.0269
4 5 0.2141 0.0285 0.0209 0.0133 0.0189 0.0132 0.2011 0.0168
5 6 0.1668 0.0167 0.0283 0.0203 0.0272 0.0201 0.1695 0.0126
6 7 0.1436 0.0158 -0.0040 0.0178 -0.0046 0.0176 0.1459 0.0121
7 8 0.1289 0.0106 0.0193 0.0181 0.0189 0.0182 0.1313 0.0042

We use lambda=0.05 and simulation time is 1ns for each lambda, for decoupling both electrostatics and van der Waals interactions (separately).

We tried to reduce the lambda steps, to increase the simulation time and also to apply position restraints to the solute atoms but the problem persists.

We would appreciate any comment or suggestion.

Thank you!