ænet-PyTorch: A GPU-supported implementation for machine learning atomic potentials training.

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Artificial intelligence for materials discovery

Background:

• Artificial neural network (ANN) based machine learning interatomic potentials (MLIPs) are crucial for simulating materials at the atomic scale.
• Existing tools like the atomic energy network (ænet) offer capabilities for MLIPs but may have limitations in training speed and GPU utilization.
• Accelerating the training process of MLIPs is essential for handling larger systems and datasets.

Purpose of the Study:

• To present ænet-PyTorch, a PyTorch-based implementation for training ANN-based MLIPs.
• To leverage GPU acceleration for faster training of MLIPs, including direct training on forces.
• To evaluate the performance and efficiency of ænet-PyTorch on open datasets.

Main Methods:

• Development of ænet-PyTorch as a PyTorch extension of the atomic energy network (ænet).
• Implementation of GPU-accelerated training for direct computation on forces and energies.
• Testing and performance evaluation using open databases for interatomic potentials.

Main Results:

• ænet-PyTorch achieves training time reductions of one to two orders of magnitude compared to CPU implementations.
• Direct training on forces is enabled for systems larger than small molecules.
• Optimal accuracy in interatomic potentials can be achieved using only 10-20% of the available force information, minimizing computational cost.

Conclusions:

• ænet-PyTorch offers a powerful and efficient alternative for training MLIPs, significantly reducing computational time.
• Efficient training strategies for MLIPs can be achieved by judiciously selecting a subset of force data.
• The findings enable faster development and application of accurate MLIPs for materials science research.