[General]: Optimizing RAD neighbor search performance

[harryswift01]

Feedback or Inquiry

Running RAD neighbor search on larger systems / longer trajectories is getting pretty slow, and runtime scales significantly with the number of molecules.

It seems most of the cost is coming from:

• Repeated COM calculations for fragments inside loops
• Nested Python loops in the RAD blocking step
• Repeated distance/angle calculations with PBC handling
• Limited vectorization / compiled execution in the inner parts of the algorithm

I’d like to explore optimising this while keeping the RAD method exactly the same (i.e. no changes to the neighbour definition or results).

Main goals:

• Avoid redundant calculations (especially COMs)
• Reduce Python overhead in the inner loops
• Keep results identical to the current implementation