EquiFiLM Enhances ML Force Fields with Charge Conditioning

Samuel Sahel-Schackis, Ken-ichi Nomura, Aiichiro Nakano, Matthias F. Kling, Thomas Linker· July 8, 2026 View original

Summary

EquiFiLM is a lightweight extension that adds continuous external conditioning, such as charge, to existing equivariant foundation machine learning force fields (MLFFs) using Feature-wise Linear Modulation (FiLM). This enables MLFFs to accurately model externally induced changes to electronic states with minimal training data.

Foundation machine learning force fields (MLFFs) like MACE-MP-0 and UMA are highly accurate for broad chemical spaces but typically assume equilibrium ground-state physics. This limits their application to dynamic processes involving external influences like charging, applied fields, or electronic excitation. EquiFiLM is introduced as a lightweight, backbone-agnostic extension that integrates continuous external conditioning into any equivariant foundation MLFF. It uses a per-layer Feature-wise Linear Modulation (FiLM) block to learn how external factors modify the potential energy surface, requiring only minimal training data. The method preserves E(3)-equivariance and, when applied to charged liquid water with MACE-MatPES, significantly reduces force and energy RMSE compared to baselines. This allows MLFFs to accurately predict charge-dependent molecular dynamics and responses with far less data than training a charge-aware model from scratch.

Why it matters

For computational chemists, materials scientists, and drug discovery professionals, EquiFiLM provides a powerful tool to extend the applicability of highly accurate MLFFs to complex, driven chemical processes, accelerating research and development.

How to implement this in your domain

  1. 1Investigate current MLFFs used in research for limitations in handling external conditions like charge or applied fields.
  2. 2Experiment with integrating EquiFiLM or similar modulation techniques into existing equivariant MLFFs.
  3. 3Apply charge-conditioned MLFFs to simulate dynamic chemical processes, such as photoexcitation or charge injection.
  4. 4Collaborate with ML researchers to adapt this methodology for other types of external conditioning relevant to specific material science problems.

Who benefits

PharmaceuticalsMaterials ScienceChemical ManufacturingEnergyAcademia

Key takeaways

  • EquiFiLM extends MLFFs to handle external conditions like charge without architectural changes.
  • It uses Feature-wise Linear Modulation (FiLM) to learn changes to the potential energy surface.
  • The method significantly improves accuracy for charged systems with minimal training data.
  • EquiFiLM enables MLFFs to model dynamic, driven chemical processes more effectively.

Original post by Samuel Sahel-Schackis, Ken-ichi Nomura, Aiichiro Nakano, Matthias F. Kling, Thomas Linker

"arXiv:2607.05559v1 Announce Type: new Abstract: Foundation machine learning force fields (MLFFs) such as MACE-MP-0 and UMA cover broad chemical space at near density functional theory (DFT) accuracy. However, they assume equilibrium ground-state physics and do not natively handle…"

View on X

Originally posted by Samuel Sahel-Schackis, Ken-ichi Nomura, Aiichiro Nakano, Matthias F. Kling, Thomas Linker on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses