I have a question regarding the suitability of LAMMPS for simulating small simulation boxes, i.e., boxes with a handful of particles. I need to conduct these simulations for unit cell optimization, i.e., obtaining the most stable crystalline arrangement for a unit cell comprised from a small number of particles (say 2-8 particles in the entire box). The problem though is that atoms might interact with not only their immediate periodic images, but also by second, third, etc. periodic images. This is of course not physical for actual MD simulations (leading to strong finite size effects), and is only meaningful for unit cell optimization. My question is whether LAMMPS is capable of handling multiple layers of periodic images (without using a long-range solver)?
LAMMPS is not subject to minimum image conventions, so you can have very small boxes with a cutoff larger than half the box diameter. However, the computational cost is similar to that of a supercell, so there is really no benefit from using a small cell.
Instead, since the force computation is very fast (this is not quantum chemistry), it is not a significant burden to use a supercell, i.e. use the replicate command to create an 8x or 27x the size cell with replicate 2 2 2 or replicate 3 3 3, respectively, and then compute the optimal cell dimension. This will also lead to more accurate results, since the stress tensor is particularly noisy for small cells and thus the cell optimization will have more difficulties to converge.