CoChat

A reactive neural network framework for water-loaded acidic zeolites

Open access

The DOI matches its Crossref record; 109 of the 112 references it cites were found in a registry or an index; 23 of the 64 sources its claims cite were checked for support.

Abstract

Under operating conditions, the dynamics of water and ions confined within protonic aluminosilicate zeolite micropores are responsible for many of their properties, including hydrothermal stability, acidity and catalytic activity. However, due to high computational cost, operando studies of acidic zeolites are currently rare and limited to specific cases and simplified models. In this work, we have developed a reactive neural network potential (NNP) attempting to cover the entire class of acidic zeolites, including the full range of experimentally relevant water concentrations and Si/Al ratios. This NNP has the potential to dramatically improve sampling, retaining the (meta)GGA DFT level accuracy, with the capacity for discovery of new chemistry, such as collective defect formation mechanisms at the zeolite surface. Furthermore, we exemplify how the NNP can be used as a basis for further extensions/improvements which include data-efficient adoption of higher-level (hybrid) references via Δ-learning and the acceleration of rare event sampling via automatic construction of collective variables. These developments represent a significant step towards accurate simulations of realistic catalysts under operando conditions.

Overview

CoChat Summary

## TL;DR

The paper develops a reactive neural network potential (NNP) for water-loaded Brønsted acidic zeolites (BAS) that aims to cover broad chemical and configurational space, including varied water loadings and Si/Al ratios. It reports that the model retains roughly meta-GGA DFT-level accuracy across diverse tests and can also support Δ-learning and machine-learned collective variables for rare-event sampling.

## Key findings

  • The final SCAN+D3(BJ) training database contained 248,439 structures; an ensemble of six SchNet NNPs trained on it achieved average test RMSEs of 5.3 meV atom−1 for energies and 186 meV Å−1 for forces.
  • Across 2700 test configurations, the NNP gave an average reaction-energy error of about 6 kJ mol−1 for FAU reaction-path tests, and the authors say this is within DFT accuracy.
  • For the in-domain Δ-learning example (proton jump in CHA), 150 training structures were sufficient to reach test-set RMSEs of 1.3 meV atom−1 and 69 meV Å−1, with the ΔNNP deviating by less than 3 kJ/mol from the ωB97X-D3(BJ) reference.
  • In out-of-domain tests, performance remained good but worsened modestly: proton jump in FAU showed about 5 kJ mol−1 relative-energy error, and Al–O(H) bond dissociation in FAU reached up to 8 kJ mol−1 error.
  • The model captured multiple reactive behaviors not in the training set, including a collective hydrolysis mechanism at an external MFI surface and a high-temperature GIS liquid-state run with thousands of bond-breaking events.

## Methods

  • The authors built a broad training set from 500 short (10 ps) AIMD trajectories plus lattice deformations, bulk-water simulations, water clusters, ice/alumina polymorphs, and prior CHA trajectories; the resulting reference dataset used SCAN+D3(BJ) single-point labels.
  • Generalization was tested with MD and NEB on unseen systems, including FAU, MFI, GIS, bulk-water-solvated fragments (Si(OH)4 and Al(OH)3), and high-temperature simulations at 1600 K and 3000 K.
  • They trained an ensemble of six SchNet-based NNPs and also built two extensions: Δ-learning to ωB97X-D3(BJ) and machine-learned collective variables via a variational autoencoder on NNP representations.

## Limitations

  • The paper states that finite-size effects hinder the interpretation of confined-water dynamics in FAU because a small primitive cell was used primarily for benchmarking.
  • The authors note that the very high-temperature conditions used in some tests (1600 K and 3000 K) are not applicable to standard applications of water-zeolite systems; this is an evident limitation of the stress tests, not the target use case.
  • For the biased-dynamics free-energy calculations, the authors explicitly say their restraints may bias the reported free-energy profiles and that the goal was to demonstrate capability rather than maximal accuracy.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41467-024-48609-2

Cites

109 of 109 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Training and testing of general BAS zeolite NNPs. a Computational workflow for the creation of the SCAN+D3(BJ) database using subsets of ab initio molecular dynamics (MD) trajectories selected by Farthest Point Sampling (FPS) with the smooth overlap of atomic positions (SOAP) descriptor. The generality of the NNPs was tested by (biased) MD and nudged-elastic band (NEB) calculations. The end-to-end learned representations are used for Δ-learning and construction of ML collective variables (ML CV). b t-distributed stochastic neighbor embedding (t-SNE) plot of the average representation vectors of all configurations in the training database (color codes shown on the left). Generalization tests are highlighted in red (Si: yellow, Al: gray, O: red, H: white). c Reaction energy error distribution ΔEr (see Eq 1) of the NNPs in comparison with ReaxFF58.
Fig. 1 Training and testing of general BAS zeolite NNPs. a Computational workflow for the creation of the SCAN+D3(BJ) database using subsets of ab initio molecular dynamics (MD) trajectories selected by Farthest Point Sampling (FPS) with the smooth overlap of atomic positions (SOAP) descriptor. The generality of the NNPs was tested by (biased) MD and nudged-elastic band (NEB) calculations. The end-to-end learned representations are used for Δ-learning and construction of ML collective variables (ML CV). b t-distributed stochastic neighbor embedding (t-SNE) plot of the average representation vectors of all configurations in the training database (color codes shown on the left). Generalization tests are highlighted in red (Si: yellow, Al: gray, O: red, H: white). c Reaction energy error distribution ΔEr (see Eq 1) of the NNPs in comparison with ReaxFF58.CC BY · PubMed Central
Fig. 2 Surface defect creation in an 2D-MFI nanosheet. a Snapshot of 2D-MFI from an exploratory 1 ns MD run at 1600 K to sample reactive events (Si: yellow, O: red, H: white). b–e Reaction steps of silanol defect creation at the 2D-MFI-water interface: b water adsorption on a surface Si and water autoprotolysis; c proton transfer from the adsorbed water to the hydroxide ion; d migration of the hydroxonium ion and adsorption on a framework oxygen; e Si--O bond hydrolysis, creating a silanol defect in axial position to the formed surface silanol.
Fig. 2 Surface defect creation in an 2D-MFI nanosheet. a Snapshot of 2D-MFI from an exploratory 1 ns MD run at 1600 K to sample reactive events (Si: yellow, O: red, H: white). b–e Reaction steps of silanol defect creation at the 2D-MFI-water interface: b water adsorption on a surface Si and water autoprotolysis; c proton transfer from the adsorbed water to the hydroxide ion; d migration of the hydroxonium ion and adsorption on a framework oxygen; e Si--O bond hydrolysis, creating a silanol defect in axial position to the formed surface silanol.CC BY · PubMed Central
Fig. 3 Reaction path modeling using Δ-learning and ML collective variables. Reaction path modeling for an (Δ)NNP in-domain (ID) case: a proton jump in CHA (a–c); and two out-of-domain (OOD) cases: a proton jump (d, f, h) and an Al–O(H) bond dissociation (e, g, i) in FAU. a, d, e Atomic structures along the reaction path (Si: yellow, Al: gray, O: red, H: white). b, f, g Static (Δ)NNP simulations and corresponding DFT energies for the NNP baseline level (SCAN+D3(BJ)) and the ΔNNP level (ωB97X-D3(BJ)). c, e, f Estimated free energy profiles using a standard collective variable CVd (see “Methods” section) for CHA and using ML collective variables for FAU.
Fig. 3 Reaction path modeling using Δ-learning and ML collective variables. Reaction path modeling for an (Δ)NNP in-domain (ID) case: a proton jump in CHA (a–c); and two out-of-domain (OOD) cases: a proton jump (d, f, h) and an Al–O(H) bond dissociation (e, g, i) in FAU. a, d, e Atomic structures along the reaction path (Si: yellow, Al: gray, O: red, H: white). b, f, g Static (Δ)NNP simulations and corresponding DFT energies for the NNP baseline level (SCAN+D3(BJ)) and the ΔNNP level (ωB97X-D3(BJ)). c, e, f Estimated free energy profiles using a standard collective variable CVd (see “Methods” section) for CHA and using ML collective variables for FAU.CC BY · PubMed Central

Verification

Claims the paper makes

17 Valid · 2 Not valid · 15 Not checked · 66 Own results and interpretations

49 more claims · 4 with their cited sources checked

Create a free account to see all 100 claims and to verify the 23 references still open.

See all claims — free account
  • Probably not a substantial claim. 14
  • It may state two separate claims as one. 38

Create a free account to see these claims.

Citations

Cites

Show all 109 references

Research works printed but not linked: 3

  1. Schütt, K., Unke, O. & Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In Proceedings of the 38th International Conference on Machine Learning 9377–9388 (PMLR, 2021).
  2. van der Maaten, L. & Hinton, G. Visualizing Data using t-SNE. J. Mach. Learn Res. 9, 2579–2605 (2008).
  3. Kingma, D. P. & Ba, J. Adam: a method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings (eds Bengio, Y. & LeCun, Y.) (ICLR, 2015).

Cited by

See all 47

Metadata

Authors
Erlebach, A., Šípka, M., Saha, I., Nachtigall, P., Heard, C. J., & Grajciar, L.
Affiliations
Charles University (CZ)
Corresponding author
  • Erlebach, Andreas
  • Grajciar, Lukáš
Venue
Nature Communications
Published
2024-05-17
Volume
15
Issue
1
Pages
4215-4215
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4396999021
Cited by
47
References
112
Citations by year
2026: 13 · 2025: 20 · 2024: 13
FWCI
4.47
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by-nc-sa)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-sa)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Univerzita Karlova v Praze (PRIMUS/20/SCI/004, UNCE/SCI/014)
  • Ministerstvo Školství, Mládeže a Tělovýchovy (ID:90254, CZ.02.1.01/0.0/0.0/15 003/0000417)
  • Grantová Agentura České Republiky (23-07616S)
Topics
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Physical Sciences › Chemistry › Inorganic Chemistry › Zeolite Catalysis and Synthesis
  • Physical Sciences › Physics and Astronomy › Atomic and Molecular Physics, and Optics › Advanced Chemical Physics Studies
Keywords
acid zeolites, zeolite acidity, hydrothermal stability, Si/Al ratio, delta learning, rare event sampling, collective variables, machine learning interatomic potentials

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: A reactive neural network framework for water-loaded acidic zeolites. https://app.cochat.ai/papers/10.1038/s41467-024-48609-2
The paper
Erlebach, A., Šípka, M., Saha, I., Nachtigall, P., Heard, C. J., & Grajciar, L. (2024). A reactive neural network framework for water-loaded acidic zeolites. Nature Communications. https://doi.org/10.1038/s41467-024-48609-2

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.