Flags containing Lennard-Jones parameters are left unchanged when two Amber PARM topologies are combined using + operator

[gaborb91]

If two Amber PARM structures are combined using the + operator like parm1 += parm2 in Python code, flags related to Lennard-Jones parameters (NONBONDED_PARM_INDEX, LENNARD_JONES_ACOEF, LENNARD_JONES_BCOEF) in parm2 are ignored. The topology will be incorrect if parm2 has extra atom types or the order of nonbonded parameters is different.

[swails]

I think you're right. The simple fix is to not re-use Lennard-Jones types and live with duplicates in the tables. The better fix (since it will impact GPU performance potentially) is to make sure every L-J type is unique.

This will be hard to do in a way that preserves off-diagonal modifications to the L-J matrix, but easy to do otherwise. A reasonable compromise is to look for off-diagonal modifications and take the "easy" way (simply append the Lennard-Jones lists) if they exist and simply recompute the L-J tables if they don't (thereby eliminating duplicates).

I believe as a temporary (and clumsy) workaround you can convert everything to a GROMACS topology, do the addition there, and then convert back (if you need something before I can get to this bug).

[swails]

Also, do you have a test case that demonstrates where this actually fails? That will help me implement a fix faster.