2026-06-26 Long-Range Electrostatics for Short-Range MLFFs

Goal

Give short-range MLFFs (MACE first) a physically correct long-range electrostatic tail, using a range-separation / delta-learning scheme: predict per-atom charges, compute the long-range Coulomb energy and forces from them, subtract that baseline from the DFT training data, fit the MLFF to the short-range remainder, and add the identical baseline back at inference. The motivating application is reliable condensed-phase and interfacial MD where the bare short-range cutoff of MACE is not enough.

This is decomposed into a sequence of subcampaigns, each a discrete, separately landable piece (reference charges via DDEC6/Bader, the charge model, the baseline, the MACE charge head, the MDI-engine electrostatics, and validation). They are intended to be tackled one at a time, in roughly dependency order, and each can be validated standalone before the next begins.

This campaign builds directly on the MDI engine work from the 2026-06-19 and 2026-06-22 campaigns: production MD runs MACE as an MDI engine driven by LAMMPS, which is what makes the electrostatics implementation tractable (see the plan).

Contents:

• Plan: Long-Range Electrostatics for Short-Range MLFFs
  • Overview
  • Design decision: electrostatics live in the MDI engine, not in LAMMPS
  • The subcampaigns
  • SEAMM pipeline (how the subcampaigns compose)
  • Open questions carried into this plan
  • References
• Notes: Atomic Charge Methods and MLFF Long-Range Prior Art
  • The one distinction that matters: what the scheme consumes
  • Method landscape
  • Recommendations
  • Bader vs DDEC6 (why both, and how they differ)
  • MLFF long-range prior art
  • References