NequIP Depth Allocation Benchmark

Charge Modeling Is Not a Substitute for Geometric Depth in Equivariant Interatomic Potentials

Introduction

Equivariant interatomic potentials like NequIP combine two different ideas:

• geometric message passing (local, symmetry-aware)
• charge-based modeling (long-range electrostatics)

Both improve performance, but they do very different things.
So the question is: if total model depth is fixed, where should that capacity go?

What this repo is

This repo contains the experiments and analysis used to study that question.

The main setup is simple:

• keep total number of layers fixed
• vary how many are geometric vs charge layers

This way, changes in performance come from where depth is used, not just increasing model size.

Main observations

• more geometric layers → consistently helps
• charge modeling → helps initially, then saturates
• too many charge layers → hurts performance

So charge modeling is useful, but it doesn’t replace geometric depth.

Also, this depends on the dataset:

• 3BPA benefits more from charge modeling
• GeTe is unstable with QEq
• MAPI works best with a mix

Models

• NequIP (baseline)
• NequIP-LR (with charge equilibration)
• NequIP-NoQeQ (direct charge prediction)

Datasets

• 3BPA
• GeTe
• MAPI-1000K

Built on

This work builds on the NequIP-charge implementations from AMERG:

• https://github.com/ahmad-research-group/nequip-charge
• https://github.com/ahmad-research-group/nequip-charge/tree/charge-encoding
• https://github.com/ahmad-research-group/nequip-charge/tree/no-Qeq

Note

This is not a fully reproducible training repo.
Some parts depend on upstream code and HPC setup.

Author

Chirag Sindhwani
IIT (BHU), Varanasi
chirag.sindhwani.cd.eee23@itbhu.ac.in