Carbon materials and their unique properties have been extensively studied by molecular dynamics, thanks to the wide range of available carbon bond order potentials (CBOPs). Recently, with the increase in popularity of machine learning (ML), potentials such as Gaussian approximation potential (GAP), trained using ML, can accurately predict results for carbon. However, selecting the right potential is crucial as each performs differently for different carbon allotropes, and these differences can lead to inaccurate results. This work compares the widely used CBOPs and the GAP-20 ML potential with density functional theory results, including lattice constants, cohesive energies, defect formation energies, van der Waals interactions, thermal stabilities, and mechanical properties for different carbon allotropes. We find that GAP-20 can more accurately predict the structure, defect properties, and formation energies for a variety of crystalline phase carbon compared to CBOPs. Importantly, GAP-20 can simulate the thermal stability of C₆₀ and the fracture of carbon nanotubes and graphene accurately, where CBOPs struggle. However, similar to CBOPs, GAP-20 is unable to accurately account for van der Waals interactions. Despite this, we find that GAP-20 outperforms all CBOPs assessed here and is at present the most suitable potential for studying thermal and mechanical properties for pristine and defective carbon.

中文翻译：

---

碳材料经验潜力的综合评估

由于可用的碳键序势（CBOP）范围广泛，碳材料及其独特的性质已通过分子动力学进行了广泛的研究。最近，随着机器学习 (ML) 的普及，使用 ML 训练的高斯近似电位 (GAP) 等电位可以准确预测碳的结果。然而，选择正确的电位至关重要，因为对于不同的碳同素异形体，每种电位的表现不同，这些差异可能导致结果不准确。这项工作将广泛使用的 CBOP 和 GAP-20 ML 潜力与密度泛函理论结果进行了比较，包括晶格常数、内聚能、缺陷形成能、范德华相互作用、热稳定性和不同碳同素异形体的机械性能。我们发现，与 CBOP 相比，GAP-20 可以更准确地预测各种晶相碳的结构、缺陷特性和形成能。重要的是，GAP-20 可以模拟 C 的热稳定性₆₀以及碳纳米管和石墨烯的准确断裂，而 CBOP 则在此挣扎。然而，与 CBOP 类似，GAP-20 无法准确解释范德瓦尔斯相互作用。尽管如此，我们发现 GAP-20 优于此处评估的所有 CBOP，并且是目前研究原始和有缺陷碳的热和机械性能的最合适的潜力。