This paper introduces a novel molecular dynamics (MD) methodology that combines the accuracy of ab initio MD (AIMD) with the efficiency of classical MD (CMD). This is achieved by using deep neural networks (DNNs) to accurately fit the potential energy surface (PES) and deploying a multiplication-less DNN on a non von Neumann (NvN) architecture to overcome the memory wall bottleneck. Testing across various molecules and bulk systems demonstrates the general applicability of this approach, which is implemented on an FPGA-based server.

Pinghui Mo, Chang Li, Dan Zhao, Yujia Zhang, Mengchao Shi, Junhua Li, Jie Liu