High-performance training and inference for deep equivariant interatomic potentials

Key Points

• The aim is to improve the scalability and speed of training and inference for molecular dynamics simulations using the NequIP framework.
• Redesigned the NequIP framework for scalable distributed training.
• Implemented PyTorch 2.0 compilation techniques.
• Utilized AOT (Ahead Of Time) Inference for optimized performance.
• Developed optimized Allegro kernels to enhance calculation speed.
• Achieved performance acceleration in molecular dynamics by factors of 5–18.
• Demonstrated efficiency improvements across practical system sizes.