Platonic representation of foundation machine learning interatomic potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Foundation machine learning interatomic potentials (MLIPs) are crucial for atomistic simulations.
• Disparate models create incompatible latent spaces, hindering direct comparison and interoperability.
• The platonic representation hypothesis posits that capable models share a common statistical reality representation.

Purpose of the Study:

• To demonstrate that independently developed MLIPs exhibit statistically consistent geometric organization of atomic environments.
• To unify the latent spaces of diverse MLIP architectures into a common framework.
• To enable cross-model analysis and identify representational biases and prediction failures.

Main Methods:

• Projecting embeddings relative to atomic anchors to unify latent spaces.
• Utilizing a common latent space that preserves chemical periodicity and structural invariants.
• Applying the unified framework for cross-model optimal transport and interpretable embedding arithmetic.

Main Results:

• Independently developed MLIPs show consistent geometric organization of atomic environments.
• A unified latent space was created for seven diverse MLIPs, preserving key chemical and structural properties.
• The framework facilitates interpretable embedding arithmetic, bias detection, and identification of atypical structures.

Conclusions:

• The platonic representation hypothesis provides a practical pathway for developing interoperable and comparable foundation models in materials science.
• Deviation within the unified space serves as a ground-truth-free metric for identifying unusual structures and physical prediction errors.
• This work paves the way for more interpretable and reliable MLIPs in scientific discovery.