Crystal structures offer crital insights into the physical and chemical properties of a substance. Given their significance, experimental methods are commonly employed for structure determination.[1] However, experimental techniques face limitations, particularly with compounds like drugs that have flexible and large molecules. In such cases, crystal structure prediction becomes valuable as it offers a preview of potential crystal structures. Despite its usefulness, crystal structure prediction encounters several challenges.[2] The primary obstacle lies in the optimization of generated structures, balancing time and accuracy. Optimizing structures for organic compounds, especially drugs with flexible and bulky molecules, using Density Functional Theory (DFT) methods can be excessively time-consuming, albeit providing good accuracy.[3] With the advancement of computational methods in material science, the use of machine learning techniques to predict crystal structures has become important. This combination provides a balance between reasonable accuracy and computational efficiency.[4] This research focuses on a pharmaceutical molecule with a quinazoline skeleton renowned for its anticancer properties. For two coformers, structures were generated with one molecule in the asymmetric unit (Z´ = 1), utilizing 13 of the most commonly occurring racemic space groups (PĪ, Pc, P2/c, P21/c, Cc, Cm, C2/c, Pca21, Pna21, Pbca, Pbcn, Aba2, and Pccn). The evolutionary algorithm, implemented in the USPEX 10.5 code [5], was employed to explore low-energy crystal structures. Structural optimization was carried out using the DFTB3-TS method. molecular structure, maintaining rigidity throughout the generation process. Post-generation, structures with similarities based on an RMSD15 < 0.1 criterion were discarded. Following these steps, approximately 4000 structures were obtained and then input into machine learning algorithms, with a cutoff of 15 kJ/mol. Subsequently, these structures were analyzed from the point of view of crystal engineering. The final optimization step involved employing the MTP potentials within the MLIP (Machine Learning Interatomic Potentials) code [6], coupled with an active learning methodology. For the optimization of training sets, the Vienna Ab initio Simulation Package (VASP) code was utilized alongside the single-point method. During this optimization process, specific parameters were carefully tuned to ensure accurate and efficient calculations.

Keywords