Machine-Learned Corrections to Mean-Field Microkinetic Models at the Fast Diffusion Limit

Abstract

We present a machine learning-based formalism to correct the mean-field assumption in microkinetic models to incorporate adsorbate interactions and surface inhomogeneity at the fast diffusion limit. Lattice Monte Carlo simulations are used to compute the macroscopic reaction rate in the presence of adsorbate interaction at different values of surface coverage. This dataset is then used to train an artificial neural network to compute precise reaction rates as a function of surface coverage of intermediates, and the underlying microkinetic model of the reaction system is modified by correcting the typical mean-field rate terms with these data-driven functions. An example of CO oxidation on the square ordered lattice is used to illustrate the speed, accuracy, and robustness of this approach, vis-à-vis a full-fledged kinetic Monte Carlo simulation. In particular, we show that while the traditional mean-field model completely misses the bistability of this system under certain conditions, the neural network-modified formalism correctly captures this phenomenon. We posit that this method scales well to larger reaction systems and is a cost-effective means to improve the accuracy of differential equation-based microkinetic models.

Unraveling Multiscale Kinetics over Subnanometer Cluster Catalysts: H2 Desorption from Pt3(-H)2/γ-Al2O3(110)

“…If the time scales of the various processes are many orders of magnitude apart, the catalytic reaction network can be treated as a singularly perturbed system, where the quasi-equilibrium-based scheme outlined above can be used. In fact, such concepts have been used to calculate ensemble-averaged transition probabilities in KMC simulations where there is a large time scale disparity between adsorbate diffusion and reaction. ,, …”

“…In fact, such concepts have been used to calculate ensemble-averaged transition probabilities in KMC simulations where there is a large time scale disparity between adsorbate diffusion and reaction. 52,57,58…”

Simple Model and Spectral Analysis for a Fluxional Catalyst: Intermediate Abundances, Pathway Fluxes, Rates, and Transients

“…The spectral analysis for our simple fluxional catalyst model also illustrates that transient relaxation of the intermediate abundances may take an extremely long time. Despite the simplicity of the model, these considerations will likely remain important for more complex systems with interacting adsorbates, 90 Langmuir−Hinshelwood steps, 90 and transient association/dissociation of the active centers. 91,92 The spectral analysis approach should be applicable to more complicated mechanisms and large ensembles of intermediates for two reasons.…”

Three Grand Challenges for the Computational Design of Heterogeneous Catalysts

“…25 A different approach was chosen by Tian and Rangarajan, who combined kinetic Monte Carlo studies with an artificial neural network to modify mean-field microkinetic models for the impact of surface coverage. 26 They showed that using this correction leads to a significantly better agreement between results for kinetic Monte Carlo simulation and microkinetic modeling.…”

“…Over the past years, methods were developed that use machine learning to predict favorable coverage patterns of CO on Pt surfaces or NO and propyne on mixed metallic surfaces . A different approach was chosen by Tian and Rangarajan, who combined kinetic Monte Carlo studies with an artificial neural network to modify mean-field microkinetic models for the impact of surface coverage . They showed that using this correction leads to a significantly better agreement between results for kinetic Monte Carlo simulation and microkinetic modeling.…”