数据驱动的原子间势：超主动学习

数据驱动的原子间势已经成为近似从头算势能面的有力工具，但创建适合的训练数据库仍然是一项充满挑战性和耗时的任务。如何有效地生成合适的训练数据库以准确插值势能面，同时避免在训练模型中出现非物理的能量最小值，已成为材料科学亟待解决的关键问题。该研究提出了超主动学习（ HAL ）框架，改进了训练数据库的创建方法，加速了数据驱动的原子间势的开发。

来自英国剑桥大学的 Cas van der Oord 研究团队，提出一种名为“超主动学习（ HAL ）”的加速主动学习方法，它结合分子动力学（ MD ）与贝叶斯优化（ BO ），加速了训练数据库的组装，用于生成数据驱动的原子间势。

HAL 框架使用包含少量原子构型的初始数据库，通过引入偏置项将分子动力学模拟引向不确定性较高的构型，从而更快地生成看不见的或有价值的训练构型。他们成功应用 HAL 框架开发了 AlSi10 合金和聚乙二醇（ PEG ）聚合物的原子簇展开（ ACE ）原子间势。

作者研究表明， HAL 生成的 ACE 原子间势能够准确预测宏观性质，如 AlSi10 的熔化温度和 PEG 的密度，其预测精度与实验接近。该研究提供了一种新的加速主动学习方法，促进了数据驱动的原子间势的开发展，提高了模拟精度和效率，为解决材料科学相关问题提供技术支持。 相关论文近期发布于 npj Computational Material s   9: 168 (2023) 。 手机阅读原文，请点击本文底部左下角 “ 阅读原文 ” ，进入后亦可下载全文 PDF 文件。

Fig. 4 HAL dynamics for several iterations for the AlSi10 random alloy showing maximum softmax normalised relative force error  estimate max sðf iÞ, temperature and pressure.

Editorial Summary

Data-driven interatomic potentials have become a powerful tool for approximating ab initio potential energy surfaces. However, creating an appropriate training database remains a challenging and time-consuming task. Effectively generating such a database for accurate potential energy surface interpolation while avoiding non-physical energy minima in the training model has become a pressing issue in materials science.

Fig. 5 Coefficient magnitude ∣ci∣ for the 723 basis functions grouped per correlation order and element interaction for various ARD  tolerances α 0 .

A team led by Prof. Cas van der Oord from the University of Cambridge, UK, has presented an accelerated active learning approach called "Hyperactive Learning (HAL)," which combines molecular dynamics (MD) with Bayesian optimization (BO) to expedite the assembly of training databases for data-driven interatomic potentials. 

Fig. 6 Pair interaction energy at close approach.

The HAL framework employs an initial database with a limited number of atomic configurations and introduces biasing terms to guide MD simulations toward configurations with higher uncertainty, thus accelerating the generation of unseen or valuable training configurations. They successfully applied the HAL framework to develop atomic cluster expansion (ACE) interatomic potentials for the AlSi10 alloy and polyethylene glycol (PEG) polymer.

The research findings demonstrate that the ACE potentials generated using HAL can accurately predict macroscopic properties, such as the melting temperature of AlSi10 and the density of PEG, with predictions closely matching experimental results. This study provides a novel accelerated active learning method that advances the development of data-driven interatomic potentials, enhancing simulation accuracy and efficiency, and offering technical support for addressing materials science-related challenges.  Thisarticle was recently published in  npj Computational Materials   9: 168 (2023) .

原文Abstract及其翻译

Abstract Data-driven interatomic potentials have emerged as a powerful tool for approximating ab initio potential energy surfaces. The most time-consuming step in creating these interatomic potentials is typically the generation of a suitable training database. To aid this process hyperactive learning (HAL), an accelerated active learning scheme, is presented as a method for rapid automated training database assembly. HAL adds a biasing term to a physically motivated sampler (e.g. molecular dynamics) driving atomic structures towards uncertainty in turn generating unseen or valuable training configurations. The proposed HAL framework is used to develop atomic cluster expansion (ACE) interatomic potentials for the AlSi10 alloy and polyethylene glycol (PEG) polymer starting from roughly a dozen initial configurations. The HAL generated ACE potentials are shown to be able to determine macroscopic properties, such as melting temperature and density, with close to experimental accuracy.

Fig. 9 Energy scatter plots for the 500 K (left), 500 K + 800 K (middle) and 500 K + HAL (right) ACE models.

摘要 数据驱动的原子间势已经成为近似从头算势能面的有力工具。创建这些原子间势最耗时的步骤通常是生成一个合适的训练数据库。为了帮助这一过程，超主动学习（ HAL ，一种加速主动学习方案）作为一种快速自动训练数据库组装的方法被提出。 HAL 在物理驱动的采样器（例如分子动力学）中增加了一个偏置项，将原子结构引向不确定性，从而生成看不见的或有价值的训练构型。提出的 HAL 框架被用于开发 AlSi10 合金和聚乙二醇（ PEG ）聚合物的原子簇展开（ ACE ）原子间势，从大约十几种初始构型开始。 HAL 产生的 ACE 势被证明能够确定熔化温度和密度等宏观性质，结果接近实验精度。

Fig. 10 HAL protocol for building linear ACE PEG model accurately determining PEG(n = 200) density within experimental accuracy of  1.2 g/cm3 at 297 K (shaded area).