Nov 3 – 4, 2022
Europe/Prague timezone

Reactive Neural Network Potentials for Zeolites with Density Functional Theory Accuracy

Nov 4, 2022, 11:20 AM
atrium (IT4Innovations)



Studentská 6231/1B 708 00 Ostrava-Poruba
User's talk Users' talks User's Talks V


Lukáš Grajciar (Charles University in Prague)


Neural network potentials (NNPs) are becoming increasingly popular in multiple areas of material science and chemistry thanks to their ability to keep ab initio accuracy at the cost of the standard reactive force fields such as ReaxFF [1,2]. However, the vast majority of the studies focused on systems with rather low-dimensional configurational space. This apparent curse of dimensionality could be behind the fact that NNPs for microporous solids such as zeolites are, to the best of our knowledge, non-existent [1,2].
In this work, we present the development of linear-scaling, reactive NNPs using the SchNet architecture [3] for various zeolite classes by using robust training and data curation procedures. The resulting NNPs retain DFT accuracy across the complex zeolitic classes considered, outperforming specialized ReaxFF by order(s) of magnitude in accuracy, while speeding-up calculations by at least three orders of magnitude compared to DFT. Using the developed NNPs we have been witnessing intriguing results such as: i) large-scale simulations of zeolite databases (>330k hypothetical zeolites) revealing more than 20k additional hypothetical frameworks in the thermodynamically accessible range for zeolite synthesis [4]; ii) multi-nanosecond molecular dynamics (MD) simulations of reactive diffusion of sub-nanometer Pt clusters in zeolites showing intermittent bond breaking of the zeolite framework [5]; iii) effects of minor topology variations on the germanium distribution in germanosilicate zeolites with profound effects on their delamination propensity [6,7]; iv) MD simulations of aluminosilicate zeolites quantifying the effects of water loadings and Si/Al ratios on proton solvation and water diffusivity [8]. Lastly, our recent work [9] suggests that the learned NNP representations of atomic environments, a by-product of our NNP generation, can be reused to construct (using variational autoencoders) robust collective variables (CVs) that are "aware" of the underlying potential energy surfaces - such CVs will facilitate any subsequent biased MD simulations that generate realistic free energy surfaces of reactive processes in zeolitic systems. Therefore, the herein developed, reactive zeolite NNPs enable calculations of complex zeolitic frameworks under experimentally relevant - operando - conditions making them a new standard in the field.

[1] N. Artrith, J. Phys. Energy 2019, 1, 032002.
[2] O. T. Unke, S. Chmiela, H. E. Sauceda, M. Gastegger, I. Poltavsky, K. T. Schütt, A. Tkatchenko, K.-R. Müller, Chem. Rev. 2021, 121, 10142–10186.
[3] K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, K.-R. Müller, J. Chem. Phys. 2018, 148, 241722.
[4] A. Erlebach, P. Nachtigall, L. Grajciar, npj Comput. Mater. 2022, 8, 1–12.
[5] D. Hou, L. Grajciar, P. Nachtigall, C. J. Heard, ACS Catal. 2020, 10, 11057–11068.
[6] M. Opanasenko, M. Shamzhy, Y. Wang, W. Yan, P. Nachtigall, J. Čejka, Angew. Chem. 2020, 132, 19548–19557.
[7] M. Jin, O. Veselý, C. J. Heard, M. Kubů, P. Nachtigall, J. Čejka, L. Grajciar, J. Phys. Chem. C 2021, 125, 23744–23757.
[8] I. Saha, A. Erlebach, P. Nachtigall, C. J. Heard, L. Grajciar, preprint: ChemRxiv. Cambridge: Cambridge Open Engage 2022,
[9] M. Šípka, A. Erlebach, L. Grajciar, J. Chem. Theory Comput. (under revision) 2022, preprint:

Primary authors

Andreas Erlebach (Charles University in Prague) Martin Šípka (Charles University in Prague) Dr Christopher J. Heard (Charles University in Prague) Prof. Petr Nachtigall (Charles University in Prague) Lukáš Grajciar (Charles University in Prague)

Presentation materials

There are no materials yet.