Speaker
Description
Lattice kinetic Monte Carlo (kMC) simulations allow for a most accurate microkinetic modeling of (catalytic) processes at crystalline surfaces that fully resolves the spatial distributions. Next to the exploitation of the lattice translational symmetry, it is in practice particularly the reliance on a complete prior-understanding of the possible elementary processes and their corresponding energy barriers that renders lattice kMC tractable. However, neither of these two efficiency drivers may be available for complex reactions and working catalysts. Incomplete knowledge of the reaction network can then be paired with an operando evolving change or loss of crystalline order.
Adaptive kMC (akMC) aims to overcome these limitations by working on the minima of the potential energy surface instead of fixed lattice sites and by determining possible elementary processes during the actual simulation. The price for the gained flexibility is an excessive number of transition state searches (TSSs) that needs to be conducted in every kMC step. To date, this prevents the practical application of akMC beyond simplest model systems, especially when based on predictive-quality first-principles energetics.
To improve its efficiency, we augment akMC with machine learning on local atomic environments. The framework employs the Smooth Overlap of Atomic Positions (SOAP) descriptor to fuzzily categorize the atoms in the system. Results of individual TSSs are then directly linked to all atoms in the corresponding group, on-the-fly establishing the elementary process database. Initial guesses for TSSs in a kMC step are proposed based on the proximity of the current environments to the environment groups in the database. The catalog is filled by encountering new local environments and discovering new kinetics through TSSs during the simulation. The accumulated TSS statistics and new environments are then fed back to the database and the machine learning cycles.
As the database fills, the recognition performance improves and the proposed guesses become close to the true transition states. These high-quality TSS guesses improve the simulation efficiency in two ways. First, the number of TSSs per kMC step gets significantly reduced by avoiding unsuccessful or repeating random TSSs. Second, damped Newton-Raphson algorithms become practical, which complete a proposed TSS in only a handful iterations. In a first application of the new framework to surface diffusion at a Pd surface, the increased akMC flexibility immediately unveils an intriguing variety of non-intuitive and hitherto not considered multi-atom processes. With their low barriers, these processes lead to qualitative changes of the surface evolution as compared to traditional kMC with a fixed list of common single-atom diffusion processes.
| Abstract Number (department-wise) | TH 12 |
|---|---|
| Department | TH (Reuter) |
