用于快速且准确建模铁电钙钛矿的交换关联泛函
An Exchange-Correlation Functional for Fast and Accurate Modeling of Ferroelectric Perovskites
AI总结：
该研究提出新型交换关联泛函C09x-PBEc，兼具准确性与计算效率，用其训练的机器学习原子间势可准确复现钛酸铅的铁电-顺电相变温度，性能优于基于GGA训练的模型。
AI中文摘要：
我们提出了一种新型交换关联泛函C09x-PBEc，它将C09交换与PBE关联相结合，旨在准确建模钙钛矿的铁电性质，同时保留GGA泛函的计算效率。随着人们对开发机器学习原子间势（MLIPs）以建模具有技术相关性的大规模铁电系统的兴趣日益增长，仔细审查用于计算MLIPs训练所依赖的力、能量和应力的密度泛函理论交换关联泛函变得十分重要。以典型铁电材料钛酸铅（PbTiO₃）和钛酸钡（BaTiO₃）为例，我们发现许多广泛使用的泛函往往会高估它们的晶格常数和自发极化。相反，带有C09交换的非局域范德华泛函能准确捕捉这些与实验相符的性质，但计算开销比GGA等泛函更大。我们证明，C09x-PBEc兼具C09交换的准确性与GGA的计算经济性，是用于铁电钙钛矿MLIPs训练的极佳候选。我们还展示，使用C09x-PBEc训练的MLIPs能准确复现PbTiO₃的铁电-顺电相变温度，与实验结果一致，相比使用GGA训练的MLIPs有显著提升。
英文摘要：
We present a novel exchange correlation functional, C09x-PBEc, which combines C09 exchange with PBE correlation, to accurately model the ferroelectric properties of perovskites while retaining the computational efficiency of GGA functionals. With a growing interest in developing machine learning interatomic potentials (MLIPs) to model large-scale ferroelectric systems of technological relevance, it is important to scrutinise the density functional theory exchange-correlation functionals which are used to compute the forces, energies and stresses the MLIP is trained on. Using the example of the prototypical ferroelectrics lead titanate, PbTiO3, and barium titanate, BaTiO3 we show that many widely used functionals tend to overestimate their lattice constants and spontaneous polarization. Conversely, non-local van der Waals functionals with C09 exchange accurately capture these properties compared to experiment, but with a larger computational overhead than, for example, GGA. We show that C09x-PBEc combines the accuracy provided by the C09 exchange with the computational affordability of GGA, making it an excellent candidate to be used in the training of MLIPs for ferroelectric perovskites. We also demonstrate that an MLIP trained using C09x-PBEc accurately reproduces the ferroelectric-to-paraelectric phase transition temperature of PbTiO3 with respect to experiment, showing a marked improvement on MLIPs trained using GGA.
