Combining DFTB and Structure-Mapping for the Prediction of Transition Paths in Deactivation of ZnO@Cu Catalysts

Solid–solid transformations are common in nature and in the aging of functional materials. Understanding the origin of these complex phenomena at the atomistic level is a pre-requisite for the design of long-living active materials. For example, catalysts undergo structural and compositional changes during aging, leading to a decline of their catalytic activity and selectivity. Understanding such transformations is necessary to rationally develop mitigation strategies.<br/>From the energetical point of view, the involved activated processes can be modeled as transitions between basins on a complex, high-dimensional free energy surface (FES). Chain-of-states methods optimize presumed pathways between two structural endpoints on FES towards the minimum energy pathway (MEP), yielding transition state estimates. Using the harmonic approximation to transition state theory (hTST), one can estimate reaction rate constants from the location of saddle points on the FES. However, there are two pitfalls that exist in this scheme that typically limit the extent of accessible spatial and time scales. First, the number of possible transition pathways that must be evaluated grows exponentially with system size and, second, ab initio methods, like density functional theory (DFT), are still too computationally expensive to cover the enormously large region of interest in the configurational space.<br/>The first problem can be solved using advanced techniques for the generation of the initial pathway(s) or, in other words, the mapping of atomic structures onto each other. This can be achieved either by purely geometrical methods (mapping of atomic positions and cells) or by topology-based methods, which map the graphs of interatomic bonds. Both approaches are complementary to each other and generate a diverse set of mappings. The combination of mapping algorithms with the chain-of-states method has recently been merged into a generalized workflow.<br/>To tackle the second problem, we harness the power of machine learning (ML) and incorporate it into an approximate, semi-empirical electronic structure model resulting from Density Functional Tight Binding (DFTB) theory [1]. By doing so, we are able to achieve energetics and electronic properties with an accuracy comparable to DFT at a fraction of the cost. Such an approach is realized by means of DFTB parametrization with a Gaussian Process Regression repulsive potential (GPrep-DFTB) [2].<br/>Previous theoretical and experimental studies have shown that ZnO tends to form a graphitic-like overstructure on a Cu surface (catalytically active system) if it is deposited in thin layers. However, the thicker the ZnO phase grows, the more a wurtzitic ZnO structure, which is catalytically less active, becomes favored [3-5]. We will present the combination of mapping algorithms along with GPrep-DFTB to investigate the phase transformations involved in the aging process in ZnO@Cu catalysts.<br/>[1] Hourahine, B, et al., J. Chem. Phys. 2020, 152, 12<br/>[2] Panosetti, C., et al., J. Chem. Theory Comput. 2020, 16, 4<br/>[3] Thang, H. V., et al., Appl. Surf. Sci. 2019, 483<br/>[4] Lunkenbein, T., et al., Angew. Chem., 2015, 127, 15<br/>[5] Lunkenbein, T., et al. Angew. Chem., Int. Ed., 2016, 55, 41