Multiscale Modeling Combined with Active Learning for Microstructure Optimization of Bifunctional Catalysts

A number of studies
have recently demonstrated that catalyst microstructure
and defect engineering are important to enhance reaction rates, but
rigorous microstructure optimization studies are lacking. Kinetic
rate prediction requires models that resolve catalyst sites and their
coupling arising from surface diffusion, spatial arrangement of multifunctional
sites, and lateral interactions. Modeling these effects requires kinetic
Monte Carlo (KMC) simulation. The computational demand of KMC simulation
makes direct microstructure optimization infeasible. To overcome this
challenge, we parametrize the KMC data (reaction rate) using an active
learning approach to capture complex structural dependencies among
sites at negligible computational cost. We apply our method to a prototype
chemistry over bifunctional materials, a case study reminiscent of
the ammonia decomposition reaction on defected NiPt (core/shell) structures.
We demonstrate that machine learning can effectively develop surrogate
models for system tasks in multiscale modeling.