Reproduce “EquiFiLM: Charge-Conditioned Equivariant Force Fields via Feature-wise Linear Modulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Samuel Sahel-Schackis, Ken-ichi Nomura, Aiichiro Nakano, Matthias F. Kling, Thomas Linker
EquiFiLM introduces a lightweight, E(3)-equivariant extension that enables continuous external conditioning for pre-trained foundation machine learning force fields, achieving high accuracy on charged systems and generalizing efficiently to unseen charge states with minimal data.
Key insights
- A parameter-efficient adapter pattern, previously used in NLP and computer vision, can be effectively extended to equivariant atomistic networks to introduce continuous external conditioning.
- Modulating only the scalar (O(3)-invariant) channels of node features is sufficient to integrate external conditioning while strictly preserving the E(3)-equivariance of the underlying force field.
- Zero-initialization of the FiLM adapter weights allows for efficient fine-tuning, enabling the adapter to start as an identity transformation and leverage the pre-trained knowledge of the foundation model with minimal additional data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.