Transferable neural wavefunctions for solids.

Area of Science:

• Quantum Chemistry
• Computational Materials Science
• Artificial Intelligence in Science

Background:

• Deep-learning-based variational Monte Carlo (DL-VMC) offers high accuracy for the many-electron Schrödinger equation.
• DL-VMC methods face challenges due to the high computational cost of optimizing neural network weights for each new system.

Purpose of the Study:

• To extend the approach of optimizing a single neural network across multiple systems to solid-state materials.
• To reduce the computational cost associated with simulating solids, which involve diverse geometries and conditions.

Main Methods:

• Implemented a single neural network ansatz optimized across various geometries, boundary conditions, and supercell sizes for solid-state calculations.
• Transferred a pre-trained neural network from smaller (2x2x2) to larger (3x3x3) LiH supercells.

Main Results:

• Optimization of a single ansatz across different solid-state variations significantly reduced the number of optimization steps required.
• Simulating larger 3x3x3 LiH supercells using a transferred network required 50 times fewer optimization steps compared to previous methods.

Conclusions:

• Optimizing a single neural network across multiple solid-state systems is a viable strategy to decrease computational overhead.
• This transfer learning approach shows promise for efficient high-throughput materials simulations.