(474d) Bayesian Chemisorption Theory of Catalysis with Uncertainty Quantification

The adsorption of molecules or their fragments at surfaces is the fundamental process for many technological applications, such as self-assembled molecular coatings, corrosion protection, chemical sensing, and catalysis [1]. With electron microscopy and spectroscopy techniques which promise to reveal the structure of adsorption complexes, a number of chemisorption models have been developed [2-3]. But our understanding of molecule-surface interactions is still far from being complete.

In recent years, there has been a rapid rise in the development and application of machine learning algorithms in catalysis. Instead of taking “black box” machine learning models which provide limited physicochemical insights into a particular system, we choose another area of machine learning in materials research, Bayesian approach[4], which is open-box and utilizes available physical models and learns model parameters from data. In this talk, we demonstrate that by marrying the Newns-Anderson model with ab initio data in Bayes’s rule [5], the Bayesian model of chemisorption can be developed for probing orbitalwise nature of adsorbate-surface interactions and (electro)-catalytic processes with uncertainty quantification.

We use the Bayesian approach to infer unknown parameters of Newns-Anderson model from evidence, i.e., projected density of states. After a sufficiently large number of MCMC iterations with accepted and rejected sampling, the distribution of the fitting parameters will eventually converge to the desired posterior distribution. I will discuss two applications of this Bayesian approach in this talk. First, we calibrated the Newns-Anderson chemisorption model to DFT simulations for H adsorption on Pt alloy surfaces. By extending the parameters we fitted from pure Pt systems to alloys, we can capture reactivity trends of Pt alloys with quantified uncertainties. Second, we revisited the Schmickler model [6] for understanding mechanism of charge transfer step using model Hamiltonians and quantified the uncertainty of this model using our fitted parameters.

References

1. K. Nørskov, Rep. Prog. Phys. 53, 1253 (1999).
2. M. Newns, Phys. Bl. 32, 611 (1976).
3. Hammer and J. K. Nørskov, Nature 376, 238 (1995).
4. C. Kennedy and A. O’Hagan, J. R. Stat. Soc. Series B Stat.
5. M. Lee, Bayesian Statistics: An Introduction, 4th ed. (Wiley, 2012).
6. Santos, P. Quaino, and W. Schmickler, Phys. Chem. Chem. Phys. 14, 11224 (2012).