Zeolite-confined subnanometric PtSn mimicking mortise-and-tenon joinery for catalytic propane dehydrogenation

Heterogeneous catalysts are often composite materials synthesized via several steps of chemical transformation, and thus the atomic structure in composite is a black-box. Herein with machine-learning-based atomic simulation we explore millions of structures for MFI zeolite encapsulated PtSn catalyst, demonstrating that the machine-learning enhanced large-scale potential energy surface scan offers a unique route to connect the thermodynamics and kinetics within catalysts’ preparation procedure. The functionalities of the two stages in catalyst preparation are now clarified, namely, the oxidative clustering and the reductive transformation, which form separated Sn4O4 and PtSn alloy clusters in MFI. These confined clusters have high thermal stability at the intersection voids of MFI because of the formation of “Mortise-and-tenon Joinery”. Among, the PtSn clusters with high Pt:Sn ratios (>1:1) are active for propane dehydrogenation to propene, ∼103 in turnover-of-frequency greater than conventional Pt3Sn metal. Key recipes to optimize zeolite-confined metal catalysts are predicted.


Computation for propene yield
Our calculated the PDH reaction rate is ~ 1.1 * 10 5 s -1 at 773 K . To estimate the propane in mol C3 per mol Pt s -1 to compare with experiment, we make the following derivation.
Considering the existence of Pt x Sn y alloys with varied Pt:Sn ratio and particle sizes, we assume that the concentrate of Pt 6 Sn 2 cluster is 0.1 ‰.

Self-learning for NN potential construction
The neutral network (NN) potential is generated by iterative self-learning of the plane wave density functional theory (DFT) global potential energy surface (PES) dataset. At first, we need to prepare an initial global PES dataset which covers all the likely compositions of Pt-Sn-Si-O systems. Then the NN potential is generated using the method as introduced in our previous work (J. Chem. Phys. 2019, 151, 050901). It starts from generating a first-generation NN potential using the initial dataset which contains ~ 73,000 structures. This first-generation NN potential is then used to carry out long-time SSW/MD-NN simulation. A small additional dataset is thus obtained from the SSW/MD sampling trajectories, containing the structures on PES either randomly selected or exhibiting new atomic environment (e.g., out-of-bounds in structural descriptor, unrealistic energy/force/curvature). After calculating these additional data by DFT, they are added into the training dataset and the whole self-learning procedure returns back to the previous stage.
Typically, after ∼100 iterations, a robust and accurate NN potential can be obtained with a compact training set that contains the most representative structures. It is worth noting that we would add a small amount of the structures that we are concerned about (e.g. PtSnO x @MFI) to the dataset and then retrain to obtain the final NN potential function.

Transition state search methods: DESW
The method operates two images starting from the initial and the final states, respectively, to walk in a stepwise manner toward each other until they meet. Once the pathway building is complete, we select the highest energy image from the chain and utilize the constrained Broyden dimer (CBD) method to locate the transition state exactly. The CBD method contains two independent modules, namely, the dimer rotation and the translation. The dimer rotation is to identify the reaction coordinate, an associated eigenvector of Hessian matrix with the negative eigenvalue, using a numerical finite difference method. Then the structure is translated gradually toward the TS along the reaction coordinate using a Quasi-Newton Broyden method. Finally, the identified transition states will be verified by further vibrational frequency analysis, which should have one and only one imaginary frequency along the reaction coordinate.