Reproduce “TriForces: Augmenting Atomistic GNNs for Transferable Representations” on OpenResearch

About this paper

Ali Ramlaoui, Alexandre Duval, Hannah Bull, Victor Schmidt, Hugues Talbot, Fragkiskos D. Malliaros, Joseph Musielewicz

TriForces introduces a model-agnostic framework that augments atomistic Graph Neural Networks by disentangling compositional and structural information through a three-stream architecture and multi-objective self-supervised pretraining. This approach enhances transferability, improves data efficiency in low-data regimes, and enables interpretable similarity retrieval, reducing Orb-v3 energy per atom MAE from 107 to 19.4 meV/atom on OMat24.

Key insights

• Explicitly disentangling compositional and structural information at the representation level significantly improves the transferability and interpretability of atomistic GNNs.
• Self-supervised learning, particularly when combining objectives that stabilize geometry, reinforce chemical patterns, and prevent representation collapse, is effective for pretraining MLIPs for diverse downstream tasks.
• A modular, stream-based approach allows for nuanced and interpretable similarity searches, enabling users to retrieve materials based on specific chemical or geometric criteria, or their combination.

Summary by alphaXiv.