Segmentation for Learning Adsorption Patterns and Residence-Time Kinetics on Amorphous Surfaces

Mattia Turchi, Ivan LunatiJournal of Chemical Information and Modeling 2025 · DOI

TL;DR

This paper introduces a refined machine learning approach to segment amorphous silica surfaces based on CO2 adsorption patterns. This segmentation allows for a more accurate characterization of adsorption kinetics, revealing distinct residence times that are crucial for developing predictive coarse-grained models of gas adsorption and catalysis.

What Problem Does This Paper Address?

Amorphous materials, like silica, are widely used in catalysis and adsorption applications due to their unique surface properties. However, their inherent structural disorder leads to highly heterogeneous local atomic environments. This heterogeneity makes it challenging to predict how molecules will adsorb and desorb, especially when trying to bridge the gap between atomistic simulations and macroscopic experimental observations. Traditional models often assume simple, uniform adsorption behavior, which fails to capture the complex dynamics arising from these disordered surfaces. Specifically, understanding the different timescales of adsorption and desorption, particularly from highly adsorptive defect sites, is critical for accurate modeling.

Methods

The core of this work is an optimized segmentation protocol using a modified Random Forest (RF) classifier. The authors refined their previous approach by:

  • Feature Reduction: Instead of using 16 features (four feature types, each with four smoothing levels), they reduced it to four features (one smoothing level per feature type: intensity, gradient magnitude, and the two eigenvalues of the Hessian matrix). This was based on feature importance and correlation analysis, aiming to minimize redundancy and improve the classifier's focus.
  • Controlled Smoothing: They systematically investigated the effect of smoothing levels (controlled by the parameter σ) on the Hessian eigenvalues (E¹ and E²). This allowed them to tune the morphology and spatial extent of the segmented regions, suppressing noise and capturing structural heterogeneities at different scales.
  • Standardized Training: They standardized the cutoffs used to define training domains for the RF classifier across all smoothing levels and surfaces, enhancing consistency and robustness.
  • Residence-Time Analysis: They analyzed the residence times of CO2 molecules within segmented regions. To better capture deviations from simple exponential decay, they merged low- and intermediate-density regions into a "background-density" (BD) class. Furthermore, they explored reclassifying portions of high-density (HD) regions into the BD class to improve the fit of residence-time distributions to a single exponential decay for certain transitions.

Datasets

The study utilized adsorption density maps generated from molecular dynamics (MD) simulations of CO2 confined within 12 amorphous silica slit pores. These maps were derived from 100 ns simulations of 100 CO2 molecules per pore, representing 24 distinct amorphous surfaces. The data was preprocessed (logarithmic transformation, denoising) and then segmented. No new benchmark datasets were introduced; rather, existing simulation data was re-analyzed with the refined segmentation method.

Key Insights and Ideas

  • Smoothing is Key for Control: The degree of smoothing applied to the Hessian eigenvalues (E¹ and E²) significantly impacts the size and morphology of the segmented regions (LD, ID, HD). Higher smoothing leads to larger, more coherent regions, while lower smoothing results in more fragmented domains.
  • HD Regions Drive Complexity: High-density (HD) regions, often associated with surface defects like undercoordinated silicon (Si3) and nonbridging oxygen (NBO), are responsible for the complex, non-exponential residence-time distributions. These regions hold adsorbate molecules for significantly longer periods.
  • Two-Step Desorption: Direct desorption from HD regions to the pore is rare. Instead, molecules typically transition to a lower-energy background-density (BD) state before desorbing, suggesting a two-step desorption process.
  • Reclassification Improves Modeling: Reclassifying a portion of HD regions into the BD class can significantly improve the fit of residence-time distributions to a single exponential decay, particularly for the BD → pore transition. This simplification is crucial for building tractable mesoscale models.
  • RF Offers Uncertainty Quantification: The RF classifier inherently provides a measure of classification uncertainty (Shannon entropy), which is absent in simpler thresholding methods. This allows for a more robust evaluation of segmentation quality.

Key Results

  • The optimized RF segmentation protocol, with reduced features and controlled smoothing, provides better control over the morphology and spatial extent of segmented regions compared to previous methods.
  • The analysis revealed that HD regions exhibit residence times that deviate significantly from exponential behavior, indicating multiple time scales associated with distinct surface defects.
  • Reclassifying a small fraction of HD regions into the BD class improved the fit of residence-time distributions to a single exponential decay for the BD → pore transition, with median residence times increasing substantially (e.g., from ~20 ps to ~60 ps).
  • The HD → BD transition was found to be more frequent and better described by a biexponential decay, reflecting the heterogeneity within these high-adsorption regions.

Limitations and Future Work

  • Non-reactive Simulations: The study focused on nonreactive MD simulations, capturing physisorption. To model chemisorption or catalytic reactions, more advanced methods (DFT, reactive MD) would be needed to obtain accurate kinetic constants.
  • Rate Inference Deferred: While the study extracts residence-time statistics, the full inference of time-dependent transition rates for mesoscale models is deferred to future work.
  • Mesoscale Model Construction: The authors suggest that the extracted residence-time statistics can inform upscaled microkinetic or kinetic Monte Carlo models, enabling simulations over larger domains and longer times. This is a clear direction for future research.
  • Generalizability: While the framework is presented as flexible, its direct application to other amorphous materials or different adsorbate molecules would require validation.

Summary updated

Abstract

Heterogeneous surfaces such as amorphous silica are characterized by highly heterogeneous local atomic environments that govern the adsorption of gas molecules through spatial arrangements. These surfaces exhibit properties that are particularly relevant for adsorption and catalytic applications. Here, we investigate CO<sub>2</sub> adsorption landscapes, captured by CO<sub>2</sub> density maps, which display complex patterns requiring machine learning (ML) segmentation for systematic analysis. We present an optimized segmentation protocol based on a modified Random Forest (RF) classifier designed to control the morphology and spatial extent of the segmented regions via feature smoothing and standardized training parameters. While broadly applicable for specific modeling goals and properties of interest, here, the method is tailored to identify high-density regions that dominate heterogeneous adsorption dynamics. For these regions, we extract residence-time statistics that deviate from exponential behavior, revealing multiple time scales associated with distinct surface defects on amorphous surfaces. The extracted kinetics provide essential information for coarse-grained models of adsorption on disordered surfaces. Such models, parametrized using atomistic simulations, enable the prediction of macroscopically measurable adsorption and desorption rates, which can be directly compared with experiments also under conditions not limited by mass transfer.
BibTeX
@article{turchi2025segmentation,
  author = {Mattia Turchi and Ivan Lunati},
  title = {{Segmentation for Learning Adsorption Patterns and Residence-Time Kinetics on Amorphous Surfaces}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.5c01463},
}
Figure