Learning Inter-Atomic Potentials without Explicit Equivariance

2025-09-25

AI Summary

This work addresses the challenge of implicitly enforcing rotational and translational symmetries (SO(3) × ℝ³) in machine-learned interatomic potentials (MLIPs) without resorting to explicit equivariant architectures. We propose TransIP—the first Transformer-based paradigm for implicitly equivariant potential energy modeling. By jointly optimizing embedding spaces and representation learning, TransIP enables non-equivariant models to spontaneously acquire physical symmetries during training, circumventing the architectural rigidity and computational overhead inherent in explicit equivariant models. Evaluated on the large-scale OMol25 dataset, TransIP achieves 40–60% higher force prediction accuracy than data-augmentation baselines and matches the performance of state-of-the-art explicit equivariant methods. Consequently, TransIP significantly enhances the flexibility, computational efficiency, and scalability of MLIPs while preserving physical consistency.

Technology Category

• Machine Learning: Transparent, Interpretable, Explainable ML
• Intelligent Robots: Embodied AI
• Computer Vision: Interpretability, Explainability, and Transparency

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Security and Privacy: Data transparency and provenance
• User Modeling, Personalization and Recommendation: Accountability, Transparency, and Ethics for personalization

Abstract

Accurate and scalable machine-learned inter-atomic potentials (MLIPs) are essential for molecular simulations ranging from drug discovery to new material design. Current state-of-the-art models enforce roto-translational symmetries through equivariant neural network architectures, a hard-wired inductive bias that can often lead to reduced flexibility, computational efficiency, and scalability. In this work, we introduce TransIP: Transformer-based Inter-Atomic Potentials, a novel training paradigm for interatomic potentials achieving symmetry compliance without explicit architectural constraints. Our approach guides a generic non-equivariant Transformer-based model to learn SO(3)-equivariance by optimizing its representations in the embedding space. Trained on the recent Open Molecules (OMol25) collection, a large and diverse molecular dataset built specifically for MLIPs and covering different types of molecules (including small organics, biomolecular fragments, and electrolyte-like species), TransIP attains comparable performance in machine-learning force fields versus state-of-the-art equivariant baselines. Further, compared to a data augmentation baseline, TransIP achieves 40% to 60% improvement in performance across varying OMol25 dataset sizes. More broadly, our work shows that learned equivariance can be a powerful and efficient alternative to equivariant or augmentation-based MLIP models.

Problem

• Learning inter-atomic potentials without architectural equivariance constraints
• Achieving symmetry compliance through embedding space optimization
• Improving computational flexibility and scalability in molecular simulations

Innovation

• Transformer-based model learns equivariance without architectural constraints
• Optimizes representations in embedding space for SO(3)-equivariance
• Achieves performance comparable to equivariant baselines without hard-wired symmetry