(415d) Prediction of Chemisorption Energies By Gaussian Processes | AIChE

(415d) Prediction of Chemisorption Energies By Gaussian Processes


Hansen, M. H. - Presenter, Stanford University
Bligaard, T., SLAC National Accelerator Laboratory
Design of heterogenous catalysts is going through a revolution due to the development of quantum chemical computations and descriptor based approaches, which reduce the dimensionality of the design problem based on a detailed physical understanding.[1]

Chemisorption energies of reaction intermediates are most often the descriptors of choice. Meanwhile, a bottleneck is currently caused by the CPU time needed to calculate energetics of these intermediates and transition states from atomic structures. It is therefore feasible to use machine learning methods to predict chemisorption energies, particularly as the amount of data grows.[2]

For catalyst screening purposes, a Gaussian Process (GP) has advantageous properties, the biggest of which is the capability of error estimates using Bayesian statistics.[3] Simple linear models are already sufficient for describing correlation between adsorbate energies, when they bind through the same atom to various surfaces. Simple GP models may therefore be made with the advantage of being highly interpretable, which provides physical insight. Typical numbers of datapoints in calculated chemisorption energy databases has yet to reach the regime of "big data" and GP's may therefore also prove to be sufficiently efficient.

We present our framework for modeling chemisorption energies using a Gaussian Process and we present applications in elucidating trends and designing metal catalyst surfaces for relevant problems in modern thermal catalysis.

  1. Jens K. Nørskov et al. “Towards the computational design of solid catalysts”. In: Nature chemistry 1.1 (2009), pp. 37–46.

  2. Zachary W. Ulissi et al. “To address surface reaction network complexity using scaling relations machine learning and DFT calculations”. In: Nature Communications 8 (2017).

  3. Carl E. Rasmussen and Christopher K. I. Williams. “Gaussian processes for machine learning. 2006”. In: The MIT Press, Cambridge, MA, USA 38 (2006), pp. 715–719.