Thread Content
Paves a new path for the research and development of flexible crystal materials. According to a report from Sinochem New Network on January 5, researchers at the Crystallization Center of Tianjin University’s School of Chemical Engineering have successfully developed an intelligent computing platform named CrystalGAT for the design of flexible crystal materials, thus opening up a data-driven approach for their efficient research and development. This platform innovatively integrates graph attention neural networks with crystal engineering techniques to enable precise prediction and targeted design of the three key mechanical properties of organic molecular crystals: elasticity, plasticity, and brittleness. It accelerates the traditional trial-and-error research approach, which involves months of high-throughput experiments to identify a viable structure, to one that allows for the generation of hundreds of candidate molecules within a single day. Moreover, the model achieves a comprehensive accuracy rate of 90% on the validation set for property assessment. In general perception, crystals are often associated with being brittle and prone to cracking. However, flexible crystals possess properties similar to those of rubber – they can be bent and deformed – while still maintaining their regular crystal structure and special functions. As such, they hold irreplaceable value in advanced fields such as flexible electronics, smart drug delivery systems, and light-driven devices. For a long time, the discovery and development of flexible crystals have relied on chance. The traditional approach involves conducting numerous experimental attempts to identify suitable molecules and prepare crystals, which is not only time-consuming and labor-intensive with high research costs, but also makes it difficult to control the mechanical properties of these crystals, thereby severely hindering technological advancements and the industrialization of related industries. The team at the Crystallization Center of the School of Chemical Engineering at Tianjin University focused on these critical challenges in this industry, and the CrystalGAT platform they developed successfully overcame these obstacles to development. The core innovation of this platform lies in the establishment of a full-chain technical system that encompasses \"data learning—precise prediction—target identification—targeted modification\". By leveraging deep learning on the vast amount of data linking crystal structure to mechanical properties, the platform can rapidly predict the mechanical properties of crystal structures of target organic molecules. More importantly, thanks to the model-based attention mechanism, the platform can quickly identify the key atoms and functional group segments in the crystal that affect mechanical properties, and visualize them in the molecular structure. Therefore, researchers can more accurately identify the molecular modification targets and functional group segments required to characterize the target. Compared to traditional approaches, this platform transforms crystal modification research from \"unguided exploration\" into \"targeted optimization,\" accelerating the research process and improving development efficiency while significantly reducing the costs associated with trial and error. The practical value of the platform has been proven in various fields. In the field of pharmaceutical engineering, the team used a platform to identify two plastic eutectics for the antiepileptic drug gabapentin; the tensile strength of the tablets made from these eutectics increased by 8.5 times and 5.7 times respectively, thereby significantly improving the tablet-forming properties and the quality rate of the active pharmaceutical ingredient. In the field of functional materials, the team successfully transformed the brittle crystal PAPA into a flexible, bendable light-emitting crystal with optical responsiveness, providing an excellent candidate material for the development of light-driven devices and soft robots. The CrystalGAT platform has been made open source on Hugging Face; researchers around the world, without any programming skills required, can simply paste or draw molecular structures to obtain online predictions of their properties as well as visualizations of key fragments. In the future, the team will work to expand the application of this platform in various cutting-edge fields such as flexible electronic sensing, adaptive intraocular lenses, optimization of high-end drug formulations, and flexible display devices. At the same time, they will continue to improve the algorithms behind the platform to enhance prediction accuracy.