The Atomic Energy Network (aenet)

The Atomic Energy Network (aenet) constructs atomic interaction potentials using artificial neural networks (ANNs) and applies machine-learning potentials (MLPs) trained on quantum-mechanics-based first-principles data to enable accurate, lower-cost molecular simulations.

Key Features:

• ANN-based potentials: Uses artificial neural networks (ANNs) to represent atomic interaction potentials.
• Machine-Learning Potentials (MLPs): Trains MLPs on data derived from quantum-mechanics-based first-principles methods.
• Reference-level accuracy with reduced cost: Achieves accuracy comparable to traditional reference methods while reducing computational expense.
• Integration with simulation engines: Interfaces with molecular dynamics and Monte Carlo packages TINKER and LAMMPS.
• Shared-memory optimization: Provides an ænet–TINKER interface optimized for shared-memory systems with near-optimal parallel efficiency.
• Distributed-memory scalability: Provides an ænet–LAMMPS interface tuned for highly parallel distributed-memory systems and leverages LAMMPS neighbor-list optimizations for scalability.

Scientific Applications:

• Diffusion in liquid water: Applied to investigate diffusion phenomena in liquid water.
• Amorphous battery materials: Used to equilibrate nanostructured amorphous battery materials.

Methodology:

Train artificial neural networks (ANNs) to produce machine-learning potentials (MLPs) from quantum-mechanics-based first-principles data and apply these MLPs within TINKER and LAMMPS, including use of LAMMPS neighbor-list optimizations and distinct shared-/distributed-memory interfaces.