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

Current quantum chemistry and materials science are dominated by first-principles methodologies such as density functional theory. However, these approaches face substantial computational costs as system scales up. In addition, the von Neumann bottleneck of digital computers imposes energy efficiency limitations. Here we propose a software-hardware co-design: the resistive memory-based reservoir graph neural network for efficient modeling of ionic and electronic interactions. Software-wise, the reservoir graph neural network is evaluated for computational tasks, including atomic force, Hamiltonian and wavefunction prediction, achieving comparable accuracy while reducing computational costs by approximately 10⁴-, 10⁶- and 10³-fold, respectively, compared with traditional first-principles methods. Moreover, it reduces training costs by approximately 90% due to reservoir computing. Hardware-wise, validated on a 40-nm 256-kb in-memory computing macro, our co-design achieves improvements in area-normalized inference speed by approximately 2.5-, 2.5- and 2.7-fold, and inference energy efficiency by approximately 2.7, 1.9 and 4.4 times, compared with state-of-the-art digital hardware, respectively.