Efficient modelling of ionic and electronic interactions by resistive memory-based reservoir graph neural network
Zhongrui Wang,Meng Xu,Shaocong Wang,Yangu He,Yi Li,Woyu Zhang,Ming Yang,Xiaojuan Qi,Ming Xu,Dashan Shang,Qi Liu,Xiangshui Miao,Ming Liu
DOI: https://doi.org/10.21203/rs.3.rs-3967300/v1
2024-01-01
Abstract:Abstract Current quantum chemistry and material science are dominated by first-principles methodologies (e.g., density function theory, DFT) running on digital computing platforms, which are widely used for modelling electronic and ionic interactions. However, this computational paradigm faces significant challenges as the system scales up, due to the computational complexity of DFT alongside the energy efficiency limit due to the von Neumann bottleneck and slowdown of Moore’s law in digital computers. To address these obstacles, we propose a software-hardware co-design: the resistive memory-based reservoir graph neural network (RGNN) for modelling systems with ionic and electronic interactions. Software-wise, RGNN eliminates the need for tedious self-consistent field calculation of DFT, significantly reducing computational complexity. In addition, RGNN features small training complexity thanks to reservoir computing. Hardware-wise, resistive in-memory computing collocates data storage and processing, improving energy efficiency associated with data shuttling. Moreover, stochasticity of resistive memory is leveraged to implement random weights of the reservoir. This co-designed system has been experimentally validated on a 40 nm 256 K in-memory computing macro and evaluated in computational tasks such as atomic forces, Hamiltonian, and ground-state many-body wavefunction, demonstrating comparable accuracy and computational complexity reduction of approximately 104, 106 and 102 times compared to DFT. Furthermore, our co-design achieves approximately 90.9%, 93.5%, 95.7% reductions in training costs compared to traditional GNNs, and approximately 72.1%, 64.7% and 82.5% improvements in energy efficiency relative to state-of-the-art digital hardware. Our co-design sets the stage for future large-scale, efficient, and affordable material system design.