Reported by Noritaka Inoue.
When the simulation is performed with non-cuboidal periodicity, the interaction energy ramps up and apparently inconsistent.
使用ツール:
MDサンプリング:Gromacs 5.1
自由エネルギー:ERmod 0.3.2
症状:
・ refsの計算が、dodecahedron、octahedronだとεmin < -3000kcal/molとなってしまう
・ solnの計算が、octahedronだとεmax > 400kcal/molとなってしまう
(trjconv -pbc mol or nojumpで修正してもダメでした。一方、dodecahedronではsolnの計算は上手くいっています)
・ 同等の計算設定でcubicの計算を実行すると上手くいく
Confirmed with gfortran 4.6.3 w/ ATLAS3.
The source of problem is the process to remove periodicity around realcal.F90:862. Current version of the source is only correct on cuboids and is thus apparently wrong. The catch is that prior to GROMACS 5.1, the atoms in trajectory are within the periodic box, and the problem has not came surface.
a25a162c1aaa to fix this issue (testing atm)
Seems the fix (a25a162c1aaa) is not enough.
Real-space interaction grid might also have the issue, I'll investigate into them later this week.
changest 37ab44c502ae implements new real-space kernel. This seems to fix the problem.
Related: issue #28
I believe that it is finally properly implemented in 1.0.