Quantum Perturbation Theory Using Tensor Cores and a Deep Neural Network.

Area of Science:

• Computational Physics
• Quantum Chemistry
• Artificial Intelligence

Background:

• Quantum response calculations are computationally intensive.
• Traditional methods face performance bottlenecks.
• Deep learning offers a novel approach to accelerate scientific computing.

Purpose of the Study:

• To implement time-independent quantum response calculations using Tensor cores.
• To map density matrix perturbation theory onto a deep neural network architecture.
• To analyze the performance and accuracy of this novel computational approach.

Main Methods:

• Utilized Tensor cores for high-performance computing.
• Mapped density matrix perturbation theory to deep neural network structures.
• Employed mixed-precision arithmetic for tensor contractions.
• Developed a parameter-free convergence criterion for linear response calculations.

Main Results:

• Achieved close to peak performance in tensor contractions using mixed-precision.
• Demonstrated quantum response calculations with self-consistent charge density-functional tight-binding and coupled-perturbed Hartree-Fock theory.
• Presented a novel parameter-free convergence criterion suitable for low-precision operations.
• Showcased a peak performance of nearly 200 Tflops on two Nvidia A100 GPUs.

Conclusions:

• Deep neural networks provide an efficient framework for quantum response calculations.
• The developed method leverages Tensor cores for significant performance gains.
• The novel convergence criterion enhances the robustness of calculations in low-precision environments.