Learning Adsorption Patterns on Amorphous Surfaces

Mattia Turchi, Sandra Galmarini, Ivan LunatiJournal of Chemical Theory and Computation 2024 · DOI

TL;DR

This paper introduces a novel Random Forest (RF) based machine learning approach to segment amorphous silica surfaces based on CO2 adsorption patterns. This method effectively distinguishes between weakly and highly heterogeneous regions, revealing distinct adsorption dynamics crucial for understanding and designing materials for applications like CO2 capture.

What Problem Does This Paper Address?

Amorphous materials, unlike their crystalline counterparts, possess inherent structural disorder. This disorder leads to a complex and heterogeneous surface landscape, characterized by topological defects and varying coordination states of atoms. These surface heterogeneities significantly influence how molecules, like CO2, interact with and adsorb onto the surface, impacting adsorption capacity and potential catalytic activity. Traditional methods struggle to analyze these complex, non-uniform adsorption patterns because they are designed for more regular, crystalline structures. The challenge lies in developing tools that can effectively segment these amorphous surfaces into meaningful regions that reflect their unique adsorption behaviors, allowing for a deeper understanding of molecule-surface interactions.

Methods

The core of this work is the application of a Random Forest (RF) classifier for image segmentation. The input data are 2D time-averaged CO2 density maps generated from molecular dynamics (MD) simulations of CO2 adsorption on amorphous silica surfaces. The RF classifier is trained using multiple features extracted from these density maps, including:

  • Pixel intensity: The local CO2 density.
  • Gradient of intensity: How rapidly the density changes.
  • Eigenvalues of the Hessian matrix: These capture the local curvature of the density map, providing information about local minima, maxima, and saddle points.

These features are nonlocal, meaning they consider the context of neighboring pixels, which is crucial for capturing the complex shapes and structures on amorphous surfaces. The process involves:

  1. Data preprocessing: Density maps are smoothed and noise is reduced.
  2. Feature extraction: The intensity, gradient, and Hessian eigenvalues are computed.
  3. Training set definition: Pixels are categorized into three potential classes (LowDens, InterDens, HighDens) based on these features and specific cutoff values. A balanced sampling strategy is employed to ensure all classes are adequately represented.
  4. RF training: An RF model is trained on this labeled data.
  5. Segmentation: The trained RF model is used to classify every pixel in the density map, creating segmented regions.

The authors also explore different parameter sets for the RF segmentation to assess their impact on the resulting segmentation quality and uncertainty.

Datasets

The study utilizes simulated CO2 density maps generated from molecular dynamics (MD) simulations. These simulations were performed on:

  • One crystalline silica pore (as a reference).
  • Twelve amorphous silica pores, categorized into three classes (LowAm, MedAm, HighAm) based on their degree of amorphism, which is controlled by the annealing temperature during their generation.

These simulated density maps serve as the "datasets" for training and evaluating the RF segmentation algorithm. No new experimental datasets were introduced.

Key Insights and Ideas

  • Amorphous surfaces are fundamentally different: The inherent disorder of amorphous silica creates adsorption landscapes that are far more complex than those on crystalline surfaces, requiring specialized analysis tools.
  • RF is a powerful tool for complex segmentation: The Random Forest classifier, by leveraging nonlocal features like gradients and Hessian eigenvalues, can effectively segment these irregular adsorption patterns, outperforming simpler thresholding methods.
  • Segmentation reveals distinct adsorption behaviors: The RF segmentation successfully separates the amorphous surface into distinct regions (LowDens, InterDens, HighDens) that exhibit markedly different CO2 adsorption dynamics, particularly in terms of residence time.
  • Undercoordinated defects are key to heterogeneity: The highly heterogeneous "HighDens" regions, characterized by complex residence-time distributions, are directly linked to the presence of undercoordinated defects on the amorphous surface.
  • Crystalline vs. Amorphous dynamics: While crystalline surfaces show predictable, exponential residence-time distributions, amorphous surfaces exhibit more complex, non-exponential behaviors driven by their heterogeneous regions.

Key Results

  • The RF segmentation algorithm successfully distinguishes between weakly heterogeneous (InterDens) and highly heterogeneous (HighDens) regions on amorphous silica surfaces.
  • InterDens regions on amorphous surfaces behave similarly to crystalline surfaces, showing exponential residence-time distributions and suggesting adsorption primarily governed by hydroxyl groups.
  • HighDens regions are characterized by complex, non-exponential residence-time distributions, indicating a stronger and more varied interaction with CO2 molecules due to undercoordinated defects.
  • The segmentation method allows for the characterization of individual HighDens objects, revealing significant heterogeneity in their area and total residence time (spanning 2-3 orders of magnitude).
  • The choice of segmentation parameters (specifically, the set1 vs. set2 approach) influences the relative areas of the segmented regions, with set2 generally assigning larger areas to InterDens and HighDens, and set1 leading to better-defined patterns and potentially more accurate predictions of target properties.

Limitations and Future Work

  • Parameter sensitivity: The segmentation quality is sensitive to the chosen parameters, requiring careful selection and validation. The paper explores different parameter sets (set1 and set2) and their impact.
  • Generalizability: While the method is presented for CO2 on amorphous silica, its applicability to other adsorbate-adsorbent systems and different types of amorphous materials is suggested but not exhaustively demonstrated.
  • Correlation with atomic structure: The authors propose that the segmented regions could be correlated with the underlying atomic structure of the amorphous surface. Future work could involve using machine learning to directly link specific defect arrangements to the observed adsorption behaviors in the segmented regions.
  • Upscaling: The segmented properties (like residence times) can be used to build more accurate kinetic Monte Carlo (KMC) models, enabling the simulation of adsorption dynamics on much larger scales and for longer times than currently feasible with MD alone. This is a key direction for future research.

Summary updated

Abstract

The physicochemical heterogeneity found on amorphous surfaces leads to a complex interaction of adsorbate molecules with topological and undercoordinated defects, which enhance the adsorption capacity and can participate in catalytic reactions. The identification and analysis of the adsorption structure observed on amorphous surfaces require novel tools that allow the segmentation of the surfaces into complex-shaped regions that contrast with the periodic patterns found on crystalline surfaces. We propose a Random Forest (RF) classifier that segments the surface into regions that can then be further analyzed and classified to reveal the dynamics of the interaction with the adsorbate. The RF segmentation is applied to the surface density map of the adsorbed molecules and employs multiple features (intensity, gradient, and the eigenvalues of the Hessian matrix) which are nonlocal and allow a better identification of the adsorption structures. The segmentation depends on a set of parameters that specify the training set and can be tailored to serve the specific purpose of the segmentation. Here, we consider an example in which we aim to separate highly heterogeneous regions from weakly heterogeneous regions. We demonstrate that the RF segmentation is able to separate the surface into a fully connected weakly heterogeneous region (whose behavior is somehow similar to crystalline surfaces and has an exponential distribution of the residence time) and a very heterogeneous region characterized by a complex residence-time distribution, which is generated by the undercoordinated defects and is responsible for the peculiar characteristics of the amorphous surface.
BibTeX
@article{turchi2024learning,
  author = {Mattia Turchi and Sandra Galmarini and Ivan Lunati},
  title = {{Learning Adsorption Patterns on Amorphous Surfaces}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2024},
  doi = {10.1021/acs.jctc.4c00702},
}
Figure