Gromacs: What can be reason of getting final potential energy greater than the starting one?

Question

I wanted to optimize potential energy of a molecule using Gromacs software (v. 5.0.4, single precision). I was really surprised when I discovered that after running Conjugate Gradients integrator, the final potential energy was greater than the starting one. Does anyone know what can be a reason of that?

Perhaps I don't understand what is optimized because potential energy went down and up during simulation.

Fragment of logs:

Step 0, Epot=4.840216e+06, Fnorm=2.899e+05, Fmax=4.849e+07 (atom 672)
....
Step 56, Epot=8.949906e+03, Fnorm=2.368e+03, Fmax=4.351e+05 (atom 783)
...
Step 244, Epot=8.550481e+07, Fnorm=2.406e+05, Fmax=4.658e+07 (atom 54660)

Energy minimization has stopped, but the forces have not converged to the
requested precision Fmax < 10 (which may not be possible for your system). It
stopped because the algorithm tried to make a new step whose size was too
small, or there was no change in the energy since last step. Either way, we
regard the minimization as converged to within the available machine
precision, given your starting configuration and EM parameters.

Polak-Ribiere Conjugate Gradients converged to machine precision in 244 steps,
but did not reach the requested Fmax < 10.
Potential Energy  =  8.5504808e+07
Maximum force     =  4.6581140e+07 on atom 54660
Norm of force     =  2.4062566e+05

Answer 1

You are right, in a normal simulation using conjugate gradient (i.e. an energy minimization) the total energy should gradually decrease towards a plateau.

Based on the numbers in the log I guess that there are some atom clashes. The potential energy of the system is too high and the forces also. I would check atom 54660 (as indicated in the log) to see if you have it too close to other atoms or in a strange configuration. If you don't see anything wrong, double check the dimension of the simulation box, maybe it is too small and it causes problem with the periodic boundary conditions.

Anyway, energy minimization using conjugate gradients (CG) should be able to solve these kind of problems. You say you run gromacs in single precision. Try to use gromacs compiled in double precision, at least at the beginning of the energy minimization. If the initial conformation of the system is very bad, single precision can be unstable and unable to minimize the energy.