Relative Backmapping Neighborhood Model

This version contains a set of 53 models that each fit exactly one atomistic position of the DOPC lipid.

Results

We have already some good results but we are not quite there yet to past the model predictions into MD simulation to do the last energy minimization. The minimum positional loss we currently face is about 0.07Å. For the MD simulations to work properly (especially Gromacs), we have to further decrease the positional loss. For now, the network structure as it is in this version will not be further developed, and we will aim at a new strategy.

Currently, we have big errors in the angle and dihedral angle predictions. We assume this is largely an error due to our set of 3N coordinates. In the future, we will develop an algorithm to find all degrees of freedom f to decrease the dimension R=3N of the variable space down to 3N-f, in theory, this should reduce the errors in any angles dramatically as we also change the coordinates to Spherical (or even Euler) coordinates.

Achievements

For some radial distribution functions, we already have some promising results:

Also the backmapping of a full lipid already assembles an atomistic representation (as discussed before, it is not quite good enough to simulate on) that kind of already looks like a proper lipid:

(Green: the true atomistic representation, Blue: the predicted atomistic representation, Red: the coarse grained structure)

V1 Absolute Positions Model

Das aktuelle Model fittet aus den Beads die 54 Atompositionen relativ zur Position des N-Atoms.

Full Changelog: https://github.com/NoxelS/backmapping/commits/v0.1.0