Exploring the configuration space of elemental carbon with empirical and machine learned interatomic potentials 

George A. Marchant, Miguel A. Caro, Bora Karasulu & Livia B. Pártay 

npj Computational Materials 9: 131 (2023).

doi.org/10.1038/s41524-023-01081-w

Published online: 23 July, 2023

编辑概述

探寻碳材料家族的新物种：基于经验势的势能面搜索

碳是宇宙中第四丰富的元素，是构成有机生命体最重要的元素之一。同时，碳具有极为丰富的同素异形体，可以形成石墨、金刚石、碳纳米管、石墨烯、富勒烯等一系列极具特色的结构。对于碳的新结构探索至今仍是一个活跃且充满神秘色彩的研究领域。基于原子间相互作用势的模拟是预测寻找新型碳结构的重要手段，尤其是在高温高压等极端条件下。发展和评估碳原子间作用势对于预测新型碳结构具有重要意义。

目前，来自英国和芬兰的研究者将嵌套算法与主流的碳原子势相结合，系统计算了碳的温度-压力相图，通过与实验的比较评估了这些作用势的可靠性。他们主要比较了经典Tersoff、EDIP势以及近年来兴起的机器学习势GAP。结果表明，几种势都可以给出正确的定性结果，即低压下易形成石墨，高压下易形成金刚石，但对于不同压力下的熔点等参数的定量预测并不可靠。相比之下，机器学习势GAP是一种更为先进的解决方案，其可以通过选取特定的结构来构建训练集从而获得更为准确的势模型。本研究中，作者发现训练的势模型甚至可以超越训练集，预测出高达200 GPa的正确结构。尽管目前GAP还不能足够准确的描述熔点线等参数，但通过进一步合理的训练，以其为代表的机器学习势模型将有望成为完善碳材料相图的有利工具。

Editorial Summary

New species in the carbon material family: Potential energy landscape mapping

Carbon is the fourth most abundant element in the universe and one of the most important elements in organisms. Meanwhile, carbon forms extremely rich allotropes, such as graphite, diamonds, carbon nanotubes, graphene, and fullerenes. The exploration of new structures of carbon is still an active and mysterious research field. Simulations based on the interatomic potential are important means to predict new carbon structures, especially at high temperatures and high pressures. Therefore, developing and evaluating the interatomic potential between carbon atoms is of great significance for predicting new carbon structures.

Currently, researchers from the UK and Finland combine the nested algorithm with the mainstream interatomic potentials to systematically calculate the temperature-pressure phase diagram of carbon, aiming at evaluating the reliability of these potentials by comparisons with experiments. They mainly compared the classic Tersoff, EDIP potential and the machine learning potential GAP that has emerged in recent years. The results show that all potentials considered can give correct qualitative results, i.e. graphite is easy to form under low pressure, and diamond is easy to form under high pressure. However, the quantitative predictions of parameters such as melting point under different pressures are not reliable. By comparison, machine learning potential GAP is a more advanced solution that can obtain a more accurate description by selecting specific structures to construct a training set. In this study, the authors found that the trained potential model can even predict the correct structure up to 200 GPa beyond the training set. Through further reasonable training, the potential field of machine learning is expected to become a beneficial tool for improving the phase diagram of carbon materials.