In Chemfiles.jl output trajectory file has a compact format and thus loses precision

[Zeki-Zhang]

This is the position matrix:

julia> xyz
3×5 Matrix{Float64}
  0.756778  -0.756778  0.0         3.81242    3.23075
  0.0        0.0       0.0         0.395536   0.109453
 -0.522494  -0.538596  0.0668583  -1.76606   -2.69293

It has full precision:

julia> xyz[:,1]
3-element Vector{Float64}:
  0.7567777232273633
  0.0
 -0.5224940120543102

After the position is written to a frame, the frame still has full precision:

julia> frame
Frame(Chemfiles.CxxPointer{Chemfiles.lib.CHFL_FRAME}(Ptr{Chemfiles.lib.CHFL_FRAME} @0x00000000053ec150, false))
julia> positions(frame)[:,1]
3-element Vector{Float64}:
  0.7567777232273633
  0.0
 -0.5224940120543102

But the file the trajectory is written to loses precision:

5
Properties=species:S:1:pos:R:3
H 0.756778 0 -0.522494
H -0.756778 0 -0.538596
O 0 0 0.0668583
O 3.81242 0.395536 -1.76606
C 3.23075 0.109453 -2.69293

[Luthaf]

Yes, the XYZ format is only written with ~6 digits of precision. Assuming the positions are in Angstroms, this should be more than enough for MD simulation trajectories.

For now, if you need perfect binary reproducibility, you should use a format that enables it, such as the NetCDF restart format (trajectory = Trajectory("filename.ncrst", "w")).

In the long term we could offer the user to select the precision to use when writing free-form text format such as xyz (but this does not guarantee perfect binary reproducibility) using the not yet implemented metadata for formats mechanism.