Heterogeneous catalysis is one of the core problems in chemistry. Not only has it brought great benefits to human society as a whole (such as the synthesis of ammonia to nitrogen fertilizers), but it has also continued to push the frontiers of chemistry due to its complexity. More and more evidence shows that the real active site is often formed in situ in the catalytic process, and it is difficult to cope with the high temperature and high pressure conditions in catalysis due to the current experimental characterization methods, which makes the theoretical calculation and simulation based on density functional (DFT) a powerful means to understand the catalytic process at the atomic level. At present, with the impact of artificial intelligence on traditional chemistry, the method based on machine Xi potential function far exceeds DFT in terms of computational speed, which promotes the upgrading of heterogeneous catalytic atomic simulation research in the new era.
fig. 1 automated search of silver surface oxides under ethene epoxidation conditions.
Based on the above background, the team of Professor Liu Zhipan from the Department of Chemistry of Fudan University reviewed the technological breakthroughs in the field Xi of heterogeneous catalysis in recent years from the perspectives of structure and reaction. Here is an example from each of them. From a structural point of view, due to the emergence of machine Xi potential function, one can apply the global optimization algorithm to the search for the global minimum value of a complex potential energy surface (i.e., the most stable structure given the initial configuration), and this process can even be nearly the same order of magnitude as the single-step optimization time of DFT.
For complex catalytic reactions with variable atomic numbers (giant canonical ensembles), the thermodynamic stability of several stable structures with different atomic numbers can be finally screened and judged by the method of parallel global optimization calculation given a variety of initial configurations.
fig. 2 the asop algorithm and its application. automated search of silver surface oxides under ethene epoxidation conditions.
Further refinement of this idea, Figure 2 shows an example of the application of such methods to silver surface oxides. It can be seen that without relying on any experimental priori, the potential energy diagram of the silver-oxygen surface structure with different atomic numbers under the given catalytic conditions can be given, from which the corresponding distribution area of the stable structure can be clearly seen, as well as the differences in the configuration and thermodynamic stability of the top stable structures, which undoubtedly has direct guiding significance for the cognition of the surface structure of the catalyst under the actual catalytic process (high temperature and high pressure).
fig. 3 pt clusters for methane activation.
From the perspective of reaction, the emergence of the Xi potential function of machine science makes the search of transition states very cheap, which provides the possibility of automatic exploration of large-scale reaction paths. For example, several chemical reaction databases containing transition state information can be obtained by performing a random potential energy surface search on the initial reactants, sampling possible products to form reaction pairs, and then conducting a massively parallel transition state search.
fig. 4 the ssw-rs method and its application.
By resetting the product with the lowest energy barrier to the reactant and repeating the process, the exploration of the entire reaction network can be continued backwards and the reaction database can be improved. If the reaction database can be machine-Xi, it is possible to realize the common production and energy barrier of a given reactant, and build a reaction network step by step and fully automatically.
fig. 5 the ai-cat method and its application.automated search of reaction pathways for glycerol hydrolysis on cu(111).
The example in Figure 5 realizes this assumption by Xi studying the reaction database of low-carbon compounds on the surface of copper, giving the reaction path of glycerol decomposition without experimental prior conditions, and the proposed new reaction channel explains the mystery of the high selectivity of 1,2-propanediol observed experimentally. Related** was recently published in NPJ Computational Materials
fig. 6 the mmlps method and its application.
Original abstract
machine-learning atomic simulation for heterogeneous catalysisdongxiao chen, cheng shang & zhi-pan liu