Accelerating Excited-State Calculations of Large Systems with Restricted Boltzmann Machines.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate excited-state calculations for large systems are computationally expensive.
• Conventional methods struggle with the vast configuration space.

Purpose of the Study:

• To develop an efficient method for excited-state electronic structure calculations in large systems.
• To accelerate the process using machine learning.

Main Methods:

• Introduced generative machine-learning-accelerated simplified Tamm-Dancoff approximation (gML-sTDA).
• Utilized restricted Boltzmann machines (RBMs) to learn and propose important excited states.
• Iteratively screened Slater determinants to avoid exhaustive evaluation.

Main Results:

• gML-sTDA accurately reproduced excitation energies for carbon nanotubes (MAE of ~0.007-0.011 eV).
• Achieved significant speedups (up to ~40x) compared to sTDA methods.
• Demonstrated consistent performance across various systems including silicon quantum dots and black phosphorus.

Conclusions:

• gML-sTDA offers a practical and scalable solution for large-scale excited-state calculations.
• This method provides an efficient approach for applications needing repeated excited-state evaluations.