E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Accurate interatomic potentials are crucial for molecular dynamics simulations.
• Existing symmetry-aware models often use invariant convolutions, limiting their representational capacity.
• Deep neural networks typically require extensive training datasets.

Purpose of the Study:

• Introduce Neural Equivariant Interatomic Potentials (NequIP), an E(3)-equivariant neural network for learning interatomic potentials.
• Demonstrate NequIP's ability to achieve high accuracy with remarkable data efficiency.
• Challenge the paradigm that deep learning necessitates massive training data.

Main Methods:

• Developed NequIP, an E(3)-equivariant neural network architecture.
• Utilized E(3)-equivariant convolutions for geometric tensor interactions, capturing richer atomic environment information.
• Trained and evaluated NequIP on diverse molecular and material datasets.

Main Results:

• NequIP achieved state-of-the-art accuracy across various challenging datasets.
• The model demonstrated exceptional data efficiency, outperforming existing methods with up to 1000x fewer training data points.
• NequIP enables high-fidelity molecular dynamics simulations over extended timescales.

Conclusions:

• NequIP offers a powerful and data-efficient approach for learning interatomic potentials.
• The method's success challenges assumptions about deep learning data requirements in scientific applications.
• NequIP facilitates the use of high-level quantum chemical references for accurate simulations.