Quantum-Based Molecular Dynamics Simulations Using Tensor Cores.

Area of Science:

• Computational Chemistry
• Artificial Intelligence Hardware

Background:

• Tensor cores offer high speed and efficiency for AI, but are limited to low-precision operations.
• Quantum-based Born-Oppenheimer molecular dynamics requires high accuracy for electronic structure and forces.

Purpose of the Study:

• To demonstrate the efficient application of tensor cores to Born-Oppenheimer molecular dynamics.
• To achieve high performance in quantum molecular simulations using AI hardware.

Main Methods:

• Utilized generalized deep neural networks for on-the-fly electronic structure calculations.
• Employed extended Lagrangian Born-Oppenheimer molecular dynamics for stable trajectories.
• Implemented a canonical ensemble simulation with Langevin-like dynamics to manage numerical noise.

Main Results:

• Achieved computational performance exceeding 100 Tflops on a single Nvidia A100 GPU.
• Generated stable molecular dynamics trajectories despite low-precision computations.
• Successfully absorbed numerical noise from low-precision operations into the simulation dynamics.

Conclusions:

• Tensor cores can be effectively applied to numerically sensitive quantum molecular dynamics.
• AI hardware acceleration enables efficient and stable quantum simulations with potential for broader applications.