[Paper Review] A reactive neural network framework for water-loaded acidic zeolites

Andreas Erlebach, Martin Šípka+4 more|arXiv (Cornell University)|Jul 3, 2023

Machine Learning in Materials Science83 references4 citations

TL;DR

This paper introduces a general reactive neural network potential (NNP) for protonic aluminosilicate zeolites under water-loaded conditions, enabling accurate, scalable ab initio-level dynamics simulations across diverse Si/Al ratios and water concentrations. The framework supports discovery of collective defect formation mechanisms and enables data-efficient $oldsymbol{ riangle}$-learning and machine-learned collective variables for rare event sampling.

ABSTRACT

<strong>Content (Creative Commons Attribution 4.0 International</strong><strong>):</strong> Energy and force data (ASE trajectory files) calculated at the DFT (SCAN+D3(BJ)), NNP and ReaxFF level for error statistics of the NNP generalization tests Trajectories of NNP level MD simulations (ASE trajectory files) for all generalization tests Trajectories for all NEB calculations Trained Neural Network Potentials NNPs including the Δ-learned model (SchNetPack, version 1.0) are available in another Zenodo repository version under CC BY-NC-SA 4.0 license (https://doi.org/10.5281/zenodo.8139369) for non-commercial use only.

Motivation & Objective

• To develop a general-purpose reactive neural network potential (NNP) for the entire class of Brønsted acidic zeolites under experimentally relevant conditions.
• To overcome the limitations of fixed-functional-form force fields like ReaxFF by enabling broad chemical and configurational transferability.
• To enable large-scale, accurate dynamics simulations of zeolite-water systems at meta-GGA DFT accuracy with minimal computational cost.
• To facilitate the discovery of novel reaction mechanisms, such as collective hydrolysis at zeolite surfaces, through extended sampling.
• To provide a foundation for data-efficient refinement using higher-level DFT corrections ($oldsymbol{ riangle}$-learning) and accelerated rare event sampling via learned collective variables.

Proposed method

• Trained a global reactive NNP on a diverse dataset spanning dense silica/alumina polymorphs, zeolites with varying Si/Al ratios, water clusters, and bulk water, using meta-GGA DFT (SCAN+D3(BJ)) as reference.
• Employed SchNetPack and ASE for NNP training and simulation, with initial structures sourced from the IZA database and solvated using Materials Studio’s Solvate module.
• Used variational autoencoders (VAEs) trained on NNP-generated representation vectors to automatically identify collective variables (CVs) for proton transfer and Al–O(H) bond dissociation.
• Applied well-tempered metadynamics in PLUMED to bias dynamics along the learned CVs, accelerating rare event sampling for reaction pathways.
• Implemented $oldsymbol{ riangle}$-learning to correct NNP energies using higher-level (hybrid) DFT reference data, enabling data-efficient refinement of accuracy.
• Restricted degrees of freedom in biased dynamics (e.g., fixing H–O distances and preventing H-bond permutations) to stabilize reaction pathways and improve sampling efficiency.

Experimental results

Research questions

• RQ1 Can a single reactive neural network potential accurately describe the full configurational and chemical space of water-loaded acidic zeolites, including varying Si/Al ratios and water loadings?
• RQ2 Can the NNP framework discover unexpected reaction mechanisms, such as collective surface hydrolysis, that are inaccessible to conventional DFT or ReaxFF simulations?
• RQ3 To what extent can NNP representations be used to construct effective, machine-learned collective variables that accelerate rare event sampling in complex zeolite systems?
• RQ4 How efficiently can higher-level DFT corrections be learned from limited reference data using $oldsymbol{ riangle}$-learning, while maintaining accuracy for specific reaction pathways?
• RQ5 Can the NNP model generalize to unseen zeolitic frameworks and chemical species not present in the training set, indicating robust transferability?

Key findings

• The trained NNP achieves consistent performance across diverse systems, including zeolites with Si/Al ratios from 1 to 12, various water loadings, and bulk water, with force errors comparable to meta-GGA DFT.
• The NNP successfully predicted and revealed a collective hydrolysis mechanism at the surface of a zeolite nanosheet, a previously unobserved defect formation pathway.
• The variational autoencoder-based collective variables enabled efficient sampling of proton transfer and Al–O(H) bond dissociation, with well-tempered metadynamics converging in 1.5–1.8 million steps.
• The $oldsymbol{ riangle}$-learning approach allowed data-efficient correction of NNP energies using only a few hundred high-level DFT reference points, improving accuracy without retraining.
• Generalization tests confirmed the model’s ability to predict energies and forces for previously unseen zeolite frameworks and water cluster configurations with high fidelity.
• All trained models, simulation data, and code—including the VAE-based CV generator and modified PLUMED library—were made publicly available on Zenodo and GitHub under a CC-BY-NC-SA 4.0 license.