From interatomic potentials to phase diagrams: Integrated tools for validation and fitting

基于经典原子间势的模拟的预测能力取决于基本相互作用模型的可转移性和质量。目前产生和拟合原子间势的方法集中在简单、易于计算的目标量上，例如晶格常数、弹性常数或点缺陷的能量。还使用了各种结构构型的总能量和原子力，其参考值可以从静态或动态从头计算中获得。验证电势性能的更严格的测试将是单组分和多组分体系的固相和液态相图。这样的测试将立即揭示给定的势能是否显示出伪像，如热力学稳定相的缺失，非物理“幽灵”状态的出现，或不准确的转变温度。然而，目前可用的计算相图的方法在数值上太昂贵，不能常规应用，特别是当目标精度可与实验得出的相图相媲美时。基于两个相关小组在采样热力学组态空间方面的互补专业知识，我们计划开发一种方法，允许以目前必要计算工作量的一小部分来计算准确的相图。在第一步中，我们将分析和测试现有的计算相图的原子方法，并从计算效率、精度和自动化程度方面对它们进行评估。基于这一认识，我们将开发自适应协议和计算工具，用于具有内置收敛检查的高吞吐量、自动化相图计算。使用这些工具，我们建立了一个热力学参考数据的数据库，并用它来验证和基准大量的潜在模型。我们将公开这些工具，使开发人员和用户能够验证他们在准确复制实验相图方面的潜力。这项研究将为开发具有准确热力学行为的下一代势能奠定基础，这些势能直接与实验相图数据和从头算自由能相匹配。

The predictive power of simulations based on classical interatomic potentials depends on the transferability and quality of the underlying interaction models. Present day approaches to generate and fit interatomic potentials focus on simple, easy-to-calculate target quantities such as e.g. lattice constants, elastic constants, or energies of point defects. Also total energies and atomic forces of various structural configurations are used, for which reference values from static or dynamic ab initio calculations are available. A much more stringent test to validate the performance of potentials would be phase diagrams for the solid and liquid phases of single- and multi-component systems. Such a test would immediately reveal whether a given potential shows artifacts such as the absence of thermodynamically stable phases, the occurrence of unphysical "ghost" states, or inaccurate transition temperatures. However, presently available approaches to compute phase diagrams are numerically too expensive to be applied routinely, particularly when targeting an accuracy of a few K comparable to experimentally derived phase diagrams. Based on the complementary expertise of the two involved groups in sampling thermodynamic configuration space we plan to develop approaches that allow to compute accurate phase diagrams at a fraction of the presently necessary computational effort. In a first step we will analyse and test existing atomistic methods for phase diagram calculation and asses them with regard to computational efficiency, accuracy and degree of automation. Based on this insight we will develop adaptive protocols and computational tools for high-throughput, automated phase diagram calculations with built-in convergence checking. Using these tools we build up a database of thermodynamic reference data and use it to validate and benchmark a large number of potential models. We will make these tools publicly accessible to allow developers as well as users to validate their potentials with respect to the accurate reproduction of the experimental phase diagrams. This research will lay the groundwork for the development of next-generation potentials with accurate thermodynamic behavior, which are directly fitted to experimental phase diagram data and ab initio free energies.