(569dm) Improving Thermodynamic Accuracy of Small Adsorbed Molecules with Adtherm

Here we present AdTherm, an open source software which provides accurate thermodynamic properties for adsorbates. It uses a sampling based approach to explore a high-dimensional potential energy surface, where a set of rigid body displacements are generated and the DFT energy at these points are calculated. These energy values are used as training data for a minima preserving neural network (MPNN). MPNN's are designed to emphasize the accuracy of fitted energy surfaces near minima, which are the the regions in potential energy surfaces which have the greatest contribution to partition functions. The training of an MPNN results in a 6D surrogate potential energy surface which can be evaluated at any number of points without needing to run any more DFT calculations. From this surrogate energy surface 10⁵-10⁷ samples are drawn. These samples are used in the evaluation of a set of phase space integrals. The evaluation of these integrals is done via importance sampling and results in the relevant thermodynamic properties of the adsorbate. Importance sampling is chosen over conventional Monte-Carlo because it reduces the number of samples required by two orders of magnitude. So far, AdTherm has been used to generate thermodynamic data for H/Cu(111), CO/Pt(111), CH₃OH/Cu(111), and CH₃/Ni(111). It is seen that the results from AdTherm are in excellent agreement with benchmarked DVR values, while results from other methods such as the: free translator/rotor, harmonic oscillator, and hindered translator/rotor are in poor agreement with the benchmarked data. The harmonic oscillator model tends underestimate partition functions, whereas the free translator/rotor overestimates partition functions.