Accelerated Atomistic Modeling of Phase Change Memory Using Deep Neural Network and Specialized Hardware

Atomistic simulations offer valuable insights into phase change memory (PCM) device research and development. Current methods, such as density functional theory (DFT) and machine learning interatomic potential (ML-IAP), face limitations in device-scale modeling. DFT achieves a time-to-solution (ηt) ≈ 10-2 s/step/atom and an energy-to-solution (ηp) ≈ 10-1 J/step/atom. State-of-the-art ML-IAPs, accelerated by GPUs, achieve (ηt) ≈ 10-6 s/step/atom and (ηp) ≈ 10-3 J/step/atom. We present a more efficient method that integrates deep neural network (DNN) with a special-purpose hardware accelerator, achieving (ηt) ≈ 10-7 s/step/atom and (ηp) ≈ 10-5 J/step/atom. Our method offers a powerful tool for high-throughput screening of PCM design possibilities (e.g., doping, interfaces), within practical time and energy consumption.