The 2024 Peking University "Boya M-Talents" International Doctoral Students Academic Forum has been held at Peking University (referred as PKU) from October 25 to 27, 2024. With the topic of "AI for Materials", through expert reports, sub-forums, and poster exchanges, this forum solicits the latest academic achievements from outstanding doctoral students worldwide. It aims to discuss academic frontiers, exchange scientific research results, and look forward to the future development trend of artificial intelligence (AI) driven the materials discipline.
Solid-state hydrogen storage technology has the advantages of high hydrogen storage density, low working pressure and good safety, and is an important development direction of hydrogen storage and transportation in the future. Among them, metal-organic framework (MOF) is one of the key materials in solid hydrogen storage technology because of its outstanding advantages such as high energy efficiency, fast response to hydrogen adsorption and desorption, and easy control. However, at present, MOF-type hydrogen storage materials still have some problems, such as lack of strong adsorption sites, low volumetric hydrogen storage density, unclear hydrogen adsorption and desorption mechanism, and high material cost, which limit their promotion and application in hydrogen energy. Storing hydrogen using MOFs under ambient temperature remains a grant challenge. To discover new MOFs and reveal how material properties decide the hydrogen storage capacity, machine learning assisted high-throughput screening and corresponding experiments are carried out. MOF database used in this research is based on CoRE MOF 2019 which consists of more than 10,000 clean different MOF structures. Structures with heavy elements (heavier than iodine) or more than 2000 atoms are first removed. The topology structures of remained MOFs are then calculated through CrystalNets using four different algorithms and ones with the same results from at least three algorithms have been chosen. The filtered database contains 1688 MOF structures whose physical properties, such as surface area, void fraction and so on, are then calculated by applying Zeo++. Since the structure data of MOFs do not contain charge information, the charge of each structure has been calculated through machine learning assisted tool PACMOF that applied pre-trained random forest model. The hydrogen storage capacities at 300 K and pressure from 100 bar to 5 bar for charge-calculated structures are calculated using grand canonical monte carlo (GCMC) simulation through RASPA. The hydrogen working capacity is therefore calculated. After screening, physical properties with and without topology information that is encoded with one-hot encoding are set as two sets of descriptors. To reveal the influence of topology structures, random forest and neural network have been applied to fit hydrogen storage capacity for both sets using 70 % of the data and 30 % are used to test.