Inconsistent energies when extending simulations

GROMACS version: 2018.3
GROMACS modification: No

Dear GROMACS users,
I had a question about extending simulations. I was running Hamiltonian replica exchange of 140 states. PLCpep7_dhdl_103.xvg was the dhdl file of replica 103 and it was ended by timeout. Subsequently, I extended the simulation with the -cpi option, which generated PLCpep7_dhdl_103_part2.xvg. As shown in the figure in the bottom, the first simulation ends at 1592 ps (the data at 1592 ps is not complete though) and the extended simulation starts from 1562 ps because the data from 1562 to 1592 ps was not saved in the last cpt file. However, I found that the energies of the overlapped time frames of the two files are not the same. For example, the difference in the total energy (the second column) is about 80 kJ/mol. I thought the velocities should be preserved when I used the -cpi option so I assumed that the energies of the overlapped time frames should be the same. Is there anything else that I missed? I was using GROMACS 2018.3 for these simulations. I’m not sure if this is a parallelization artifact (so that the extended simulation is still valid since the velocity was randomly drawn from the same ensemble) or a worse problem such that any of my extended simulations were actually wrong. Is this some known bug in GROMACS or not? For you information, the command that I was using for extending the simulation was:

mpirun -np 140 gmx_mpi mdrun -deffnm PLCpep7 -dhdl PLCpep7_dhdl.xvg -replex 100 -nex 2744000 -multidir state_0 state_1 state_2 state_3 state_4 state_5 state_6 state_7 state_8 state_9 state_10 state_11 state_12 state_13 state_14 state_15 state_16 state_17 state_18 state_19 state_20 state_21 state_22 state_23 state_24 state_25 state_26 state_27 state_28 state_29 state_30 state_31 state_32 state_33 state_34 state_35 state_36 state_37 state_38 state_39 state_40 state_41 state_42 state_43 state_44 state_45 state_46 state_47 state_48 state_49 state_50 state_51 state_52 state_53 state_54 state_55 state_56 state_57 state_58 state_59 state_60 state_61 state_62 state_63 state_64 state_65 state_66 state_67 state_68 state_69 state_70 state_71 state_72 state_73 state_74 state_75 state_76 state_77 state_78 state_79 state_80 state_81 state_82 state_83 state_84 state_85 state_86
state_87 state_88 state_89 state_90 state_91 state_92 state_93 state_94 state_95 state_96 state_97 state_98 state_99 state_100 state_101 state_102 state_103 state_104 state_105 state_106 state_107 state_108 state_109 state_110 state_111 state_112 state_113 state_114 state_115 state_116 state_117 state_118 state_119 state_120 state_121 state_122 state_123 state_124 state_125 state_126 state_127 state_128 state_129 state_130 state_131 state_132 state_133 state_134 state_135 state_136 state_137 state_138 state_139 -cpi PLCpep7.cpt -noappend

And here is the image:

Any experience shared will be much appreciated! Thank you in advanced!

Best,
Wei-Tse

Hi Wei-Tse,

even if the velocities are kept, simulations diverge extremely quickly due to numerical noise. It is a bit surprising that you get divergence so quickly, but I would not be too worried here, because the dH/dL measure is very sensitive to quite small changes in the system, which is reflected in the fluctuations that you will observe throughout your simulation.

While I don’t have hands-on experience with the replica exchange simulations, I would assume that your setup and continuation is valid on first sight.