Efficient modeling of ionic and electronic interactions by a resistive memory-based reservoir graph neural network

Area of Science:

• Quantum Chemistry
• Materials Science
• Computational Science
• In-memory Computing

Background:

• First-principles methodologies like density functional theory dominate quantum chemistry and materials science.
• These methods incur substantial computational costs and face energy efficiency limitations due to the von Neumann bottleneck in digital computers.

Purpose of the Study:

• To propose a software-hardware co-design for efficient modeling of ionic and electronic interactions.
• To leverage reservoir graph neural networks (RGNNs) on resistive memory for enhanced computational performance.

Main Methods:

• Development of a reservoir graph neural network (RGNN) model.
• Implementation on resistive memory-based in-memory computing hardware.
• Evaluation of RGNN for atomic force, Hamiltonian, and wavefunction prediction.
• Comparison with traditional first-principles methods and state-of-the-art digital hardware.

Main Results:

• RGNN achieved comparable accuracy while reducing computational costs by 10^4-, 10^6-, and 10^3-fold for atomic force, Hamiltonian, and wavefunction prediction, respectively.
• Training costs were reduced by approximately 90% due to reservoir computing.
• The co-design demonstrated improved area-normalized inference speed (2.5-2.7x) and energy efficiency (1.9-4.4x) on a 40-nm 256-kb in-memory computing macro compared to digital hardware.

Conclusions:

• The proposed software-hardware co-design offers a significant advancement in computational efficiency for quantum chemistry and materials science.
• Resistive memory-based RGNNs provide a promising pathway to overcome the limitations of current digital computing architectures for complex simulations.