Applications of Atomistic Machine Learning for Estimating Adsorbate Free Energy and Entropy on Late-Transition Metal Surfaces

Free energies of adsorption are fundamental quantities in microkinetic modeling of chemical reactions on catalyst surfaces. Density functional theory (DFT) based approaches are well suited for calculating the electronic components of these free energies, but are too expensive to perform the sampling necessary to evaluate the finite temperature contributions. Consequently, these entropic contributions to the free energy are usually approximated, e.g. by treating the adsorbate as a harmonic oscillator or as a two-dimensional ideal gas. In this work, we test an alternative approach for computing the non-electronic contributions to the adsorption free energies by using neural network potentials [1] to sample the potential energy landscape. The neural network potentials are trained on a set of DFT calculations of simple adsorbates on transition metal surfaces. The adsorption free energy and entropy are then computed using the canonical partition function. We compare our free energy estimates to those obtained using standard approximations, and discuss tradeoffs in computational cost and accuracy.

References

[1] Khorshidi, A. and Peterson, A.A., 2016. Computer Physics Communications, 207, pp.310-324.