Session:

Machine learning interatomic potentials (MLIPs) trained on quantum mechanical simulation datasets are promising to study the evolution of catalytic processes at extended time and length scales, rendering accuracies comparable to the reference method at a reduced cost.¹ MLIPs approximate the underlying potential energy surface (PES) by learning solely on the correlation between the atomic coordinates and energies, forces in the training dataset. The potentials optimized via appropriate active learning schemes could be applied to identify unexplored regions of the PES, for example to predict novel catalyst morphologies, binding configurations, reaction pathways, and to further model their evolution at specific thermodynamic conditions.² However, strategies to optimize MLIPs through incremental dataset expansion guided by relevant uncertainty metrics in an active learning framework are not well recognized.

In this work, a minimal structured dataset is compiled by accounting for the factors commonly encountered in catalysis to train a base artificial neural network (ANN) potential. The atomic positions and energies from single point calculations using density functional theory, of the smallest model metal surface-adsorbate interface is transformed to training set using Chebyshev descriptor implemented in aenet.³ We propose a strategy to educate the models on forces based on displacements of selective atoms. The trained base potential predicts energies, forces, and binding strengths of both relaxed and unrelaxed larger unit cells with average RMS errors of 4 meV/atom, 0.15 eV/Å, and 0.1 eV/atom respectively.

It gets challenging to apply the model, for example to predict catalyst dynamics at reaction conditions, as its ability to extrapolate to regions beyond the trained (near) equilibrium datapoints of the PES is very limited. We employ deep ensembles⁴ to investigate possible uncertainty driven active learning protocols to upgrade the base model. The spread in predicted energies, forces between the ensemble members correlates with the errors with respect to reference values⁵ when a reliable data splitting strategy is used to train the ensemble, and hence is used as a metric to ascertain uncertainty bounds. An apt solution is offered to generate effective ensemble members trained on extended datasets. Strapping an ANN ensemble with tools such as molecular dynamics allows for effective sampling of the configuration space.⁶ A better approximation of the PES is attained by iterative retraining of the ensemble on highly uncertain datapoints of the explored configuration space. Influence of the choice of uncertainty drivers on the evolution of deep ensemble potentials will be discussed. Approaches to further equip the ensemble, on the levels of both data as well as ANN model will be presented.

1. J. Behler, Angew. Chemie - Int. Ed. 56 (2017) 12828.
2. J.R. Kitchin, Nat. Catal. 1 (2018) 230.
3. N. Artrith, A. Urban, G. Ceder, Phys. Rev. B 96 (2017) 1.
4. K. Tran, W. Neiswanger, J. Yoon, Q. Zhang, E. Xing, Z.W. Ulissi, Mach. Learn. Sci. Technol. 1 (2020) 025006.
5. A.A. Peterson, R. Christensen, A. Khorshidi, Phys. Chem. Chem. Phys. 19 (2017) 10978.
6. M. Kulichenko, K. Barros, N. Lubbers, Y.W. Li, R. Messerly, S. Tretiak, J.S. Smith, B. Nebgen, Nat. Comput. Sci. 3 (2023) 230.

Co-Author(s): Johannes Voss (SLAC National Accelerator Laboratory)

Suman Bhasker Ranganath

Appearances