王涵研究员做学术报告Modeling the interatomic potential by deep learning
2020年11月19日，我中心邀请到北京应用物理与数学研究所的王涵研究员做学术报告，介绍使用深度学习的方法来拟合高精度第一性原理势函数，通过高性能计算、数学和物理方法的结合，最终在Summit超级计算机上达到双精度91PFLOPS，混合单精度164PFLOPS，混合半精度275PFLOPS，进一步增加网络维度，该工作可以推到1.1EFLOPS。这一工作也于20日揭晓获得今年的高性能计算最高奖项-戈登贝尔奖（Gordon Bell Prize）。
Modeling the interatomic potential by deep learning
An accurate description of the interatomic potential energy surface (PES) is one of the central problems in molecular simulations. For a long time, one has to choose between the first principle PESs that are accurate but computationally expensive and the empirical PESs (force fields) that are efficient but of limited accuracy. We discuss the solution to this dilemma in two aspects: PES construction and data generation. In terms of PES construction, we introduce the Deep Potential (DP) method, which faithfully represents the first principle PES by a symmetry-preserving deep neural network. In terms of data generation, we present a new concurrent learning scheme named Deep Potential Generator (DP-GEN). This approach automatically generates the most compact training dataset that enables the training of DP with uniform accuracy. By contrast to the empirical PESs, the DP-GEN opens the opportunity of continuously improving the quality of DP by exploring the chemical and configurational space of the system. After a few examples of DP and DP-GEN, we introduce the open-source implementations of DP named DeePMD-kit, and a recent GPU optimization of DeePMD-kit for the world’s fastest supercomputer, which makes possible nanosecond simulation of 100 million atoms with ab initio accuracy in a day.