FFCASP: A Massively Parallel Crystal Structure Prediction Algorithm

A new
algorithm called Fast and Flexible CrystAl Structure Predictor
(FFCASP) was developed to predict the structure of covalent and molecular
crystals. FFCASP is massively parallel and able to handle more than
200 atoms in the unit cell (in other terms, it allows global optimization
around 100 individual parameters). It uses a global optimizer specialized
for Crystal Structure Prediction (CSP) which combines particle swarm
and simulated annealing optimizers. Three different molecular crystals,
including diverse intermolecular interactions, namely, cytosine, coumarin,
and pyrazinamide, have been selected to evaluate the performance of
FFCASP. While cytosine polymorphs have been searched by employing
two different force fields (a DFT-SAPT based intermolecular potential
and generalized amber force field (GAFF)) up to Z = 16, only GAFF has been used both in coumarin and pyrazinamide
polymorph searches up to Z = 4. For these three molecular
crystals, FFCASP generated more than 20 000 crystal structures,
and the unique ones have been further treated by DFT-D3. A combination
of data mining and a machine learning approach was introduced to determine
the unique structures and their distribution into different clusters,
which ultimately gives an opportunity to retrieve the common features
and relations between the resulting structures. There are two known
experimental crystal structures of cytosine, and both were successfully
located with FFCASP. Two of the reported crystal structures of coumarin
have been reproduced. Similarly, in pyrazinamide, three known experimental
structures have been rediscovered. In addition to finding the experimentally
known structures, FFCASP also located other low-energy structures
for each considered molecular crystals. These successes of FFCASP
offer the possibility to discover the polymorphic nature of other
important molecular crystals (e.g., drugs) as well.

本研究开发了一种名为快速灵活晶体结构预测器（Fast and Flexible CrystAl Structure Predictor, FFCASP）的新型算法，用于预测共价晶体与分子晶体的结构。FFCASP具备大规模并行计算能力，可处理晶胞内超过200个原子（换言之，其支持对约100个独立参数开展全局优化）。该算法采用专为晶体结构预测（Crystal Structure Prediction, CSP）设计的全局优化器，融合了粒子群优化与模拟退火两种优化策略。
本研究选取三种具有多样化分子间相互作用的分子晶体——胞嘧啶、香豆素与吡嗪酰胺——作为测试对象以评估FFCASP的性能。针对胞嘧啶多晶型的搜索，研究采用了两种不同的力场：基于DFT-SAPT的分子间势能势函数与广义AMBER力场（GAFF），搜索范围最高至Z=16；而香豆素与吡嗪酰胺的多晶型搜索仅使用了GAFF力场，搜索上限为Z=4。
针对这三种分子晶体，FFCASP共生成了超过20000个晶体结构，其中的独特结构进一步通过DFT-D3方法进行处理。本研究引入数据挖掘与机器学习相结合的方法，用于甄别独特结构并分析其在不同簇中的分布规律，最终得以提取所得晶体结构间的共性特征与内在关联。
胞嘧啶已知存在两种实验晶体结构，二者均被FFCASP成功定位。香豆素的两种已报道晶体结构均被成功复现。类似地，吡嗪酰胺的三种已知实验晶体结构也被重新找到。除成功找到实验已知的晶体结构外，FFCASP还为每一种受试分子晶体找到了其他低能稳定结构。FFCASP的这些成功应用，也为探索其他重要分子晶体（如药物）的多晶型性质提供了可行路径。