Linear-Scaling and Memory-Efficient Implementation of van-der-Waals Interaction (DFT-D3) for Large Systems.

Area of Science:

• Computational materials science
• Condensed matter physics
• Physical chemistry

Background:

• Van der Waals (vdW) interactions are fundamental to material properties but computationally intensive for large systems.
• Accurate atomic simulations require efficient methods to include long-range vdW forces.

Purpose of the Study:

• To develop a highly efficient GPU implementation of the DFT-D3 method for vdW corrections in large-scale atomic simulations.
• To enable fast vdW calculations compatible with machine-learning potentials.

Main Methods:

• Implemented the DFT-D3 method using CUDA for GPU acceleration (LASP-D3).
• Achieved linear-scaling time complexity O(N) for periodic systems.
• Combined LASP-D3 with a generalized global neural network potential.

Main Results:

• LASP-D3 offers up to 100x speedup for systems >100,000 atoms compared to existing methods.
• Significantly reduced GPU memory consumption.
• Accurately reproduced experimental conductivity of the solid electrolyte LiTaCl6, highlighting the role of vdW interactions in Li-ion diffusion.

Conclusions:

• LASP-D3 provides a computationally efficient approach for large-scale atomic simulations incorporating vdW interactions.
• The method facilitates accurate modeling of material properties, such as ionic conductivity in solid electrolytes.
• This work enables faster discovery and design of advanced materials.