转载自公众平台：npj计算材料学

在物理、化学、材料科学和生物学领域，分子动力学（MD）是模拟多体系统动力学演化的一个重要工具。该方法通过求解涉及复杂原子间势的牛顿运动方程，确定相互作用粒子的运动轨迹。然而，经典MD方法在任务间可转移性较差，而从头算分子动力学（AIMD）方法严格处理电子自由度的成本较高，不适用于大规模计算。

一个解决方案是使用机器学习方法来促进MD模拟，这在很多情况下都取得了优越的性能。大多数机器学习方法都以齐次图对分子系统进行建模和处理，但在实际应用中，具有多节点和边缘类型的图数据是普遍存在的，这便严重限制了代表不同相互作用的表达能力。因此，改进机器学习方法对材料的MD模拟非常重要。

来自清华大学低维量子物理国家重点实验室与物理系的段文晖院士团队，提出了一种端到端的异构关系消息传递网络（HermNet），可在一个具有从头计算精度的单一模型中有效地描述多个相互作用。该模型将分子图或晶体图分割成几个子图，并对不同的子图使用不同的消息传递网络。

在每个子图中，选择了一个改进版本的可极化原子相互作用神经网络作为子网络。对分子体系和块材体系进行了模拟，取得了令人满意结果。

具体来说，在修正的分子动力学17（rMD17）、量子机器9（QM9）和扩展系统数据集上，HermNet分别在近75%、83%和69%的任务中优于其他测试模型。最后，作者从密度泛函理论的角度阐明了HermNet的设计如何与量子力学兼容。该研究提出了用于MD模拟的通用机器学习框架，对材料特性预测有巨大价值。相关论文发表于npj Computational Materials  8: 53 (2022).

Editorial Summary

New Development of Molecular Dynamics: Heterogeneous relational message passing networks

In the realm of physics, chemistry, material science and biology, molecular dynamics (MD) simulation is an essential tool for modeling dynamical evolution of a many-body system. This method determines the trajectories of interacting particles by solving Newton’s equations of motion involving complex interatomic potentials. However, classical MD possesses poor transferability across tasks, while ab initio molecular dynamics (AIMD) is not applicable to large-scale calculations due to the high cost of rigorously treating the electronic degrees of freedom. One solution is to use machine learning (ML) methodsto facilitate MD simulations, which have yielded superior performances in many occasions. Most of ML model and process molecular systems in terms of homogeneous graph, however, in practice, graph data with multiple node and edge types is ubiquitous and more appropriate for molecular systems, which severely limits the expressive power for representing diverse interactions. Therefore, improving machine learning method is very important for MD simulations of material properties. 

A team led by Prof. Wenhui Duan from State Key Laboratory of Low Dimensional Quantum Physics and Department of Physics, Tsinghua University, presented an end-to-end heterogeneous relational message passing network (HermNet), to efficiently express multiple interactions in a single model with ab initio accuracy. This model splits the molecular or crystal graph into several subgraphs and use different message passing networks for different subgraphs. Within each subgraph, we choose a modified version of polarizable atom interaction neural network (PAINN) as the sub-network. Experiments on molecular and extended systems were performed and the results were satisfactory. Specifically, HermNet outperforms other tested models in nearly 75%, 83% and 69% of tasks on revised Molecular Dynamics 17 (rMD17), Quantum Machines 9 (QM9) and extended systems datasets, respectively. Finally, authors elucidate how the design of HermNet is compatible with quantum mechanics from the perspective of the density functional theory. This study proposes a universal machine learning framework for MD simulations, highly valuable for future materials properties prediction. This article was recently published innpj Computational Materials  8: 53 (2022).

原文Abstract及其翻译

Abstract With many frameworks based on message passing neural networks proposed to predict molecular and bulk properties, machine learning methods have tremendously shifted the paradigms of computational sciences underpinning physics, material science, chemistry, and biology. While existing machine learning models have yielded superior performances in many occasions, most of them model and process molecular systems in terms of homogeneous graph, which severely limits the expressive power for representing diverse interactions. In practice, graph data with multiple node and edge types is ubiquitous and more appropriate for molecular systems. Thus, we propose the heterogeneous relational message passing network (HermNet), an end-to-end heterogeneous graph neural networks, to efficiently express multiple interactions in a single model with ab initio accuracy. HermNet performs impressively against many top-performing models on both molecular and extended systems. Specifically, HermNet outperforms other tested models in nearly 75%, 83% and 69% of tasks on revised Molecular Dynamics 17 (rMD17), Quantum Machines 9 (QM9) and extended systems datasets, respectively. In addition, molecular dynamics simulations and material property calculations are performed with HermNet to demonstrate its performance. Finally, we elucidate how the design of HermNet is compatible with quantum mechanics from the perspective of the density functional theory. Besides, HermNet is a universal framework, whose sub-networks could be replaced by other advanced models.

摘要随着许多基于消息传递神经网络的框架被提出来预测分子和固体特性，机器学习方法已经极大地改变了支撑物理学、材料科学、化学和生物学的计算科学的范式。虽然现有的机器学习模型在许多情况下都取得了优越的性能，但大多数都以齐次图对分子系统进行建模和处理，这就严重限制了代表不同相互作用的表达能力。在实际应用中，具有多节点和边缘类型的图数据是普遍存在的，更适用于分子系统。因此，我们提出了异构关系消息传递网络（HermNet），一种端到端的异构图神经网络，可以在一个具有从头计算精度的单一模型中有效地描述多个相互作用。HermNet在分子系统和块材系统上对许多高性能模型上表现得令人印象深刻。具体来说，HermNet在修正的分子动力学17（rMD17）、量子机器9（QM9）和扩展系统数据集上，分别在近75%、83%和69%的任务中优于其他测试模型。此外，利用HermNet进行了分子动力学模拟和材料性能计算，以证明其性能。最后，我们从密度泛函理论的角度阐明了HermNet的设计如何与量子力学兼容。此外，HermNet是一个通用的框架，其子网络可以被其他高级模型所取代。