Probing Nanoscale Structural Perturbation in a WS2 Monolayer via Explainable Artificial Intelligence

Nanoscience Artificial Intelligence Chemistry Information Science Materials Science Semiconductors and Information Devices
nanoscale structural perturbation explainable artificial intelligence Raman spectroscopy machine learning

Background Introduction

Two-dimensional (2D) materials exhibit significant potential in fields such as nanoelectronics and optoelectronics due to their unique physical and chemical properties. However, structural perturbations at the nanoscale have a profound impact on their performance. Traditional characterization methods like Raman spectroscopy, while capable of providing structural information, are often limited by the diffraction limit, making it challenging to precisely detect structural changes at the nanoscale. To address this issue, researchers have begun exploring the integration of machine learning (ML) with spectroscopic techniques to enhance spatial resolution and reveal nanoscale structural perturbations.

This study, conducted by a collaborative research team from Hanyang University, Sungkyunkwan University, Korea Advanced Institute of Science and Technology, and other institutions, was published in Applied Physics Reviews on April 16, 2025. By developing a convolutional neural network (CNN)-based machine learning model, combined with data from Kelvin probe force microscopy (KPFM) and atomic force microscopy (AFM), the researchers successfully improved the spatial resolution of Raman spectroscopy to 50 nanometers and revealed the strain distribution in nanoscale wrinkles of WS2 monolayers.

Research Process

1. Data Collection and Preprocessing

The research team first prepared WS2 monolayer samples using chemical vapor deposition (CVD) and performed measurements using a confocal Raman spectrometer and an AFM system. To train the machine learning model, a sliding window approach was employed to convert KPFM and AFM data into 15×15 input tensors, which were then paired with corresponding Raman spectra. The Raman spectra underwent background noise removal and baseline correction to emphasize peak signals.

2. Machine Learning Model Training

The team designed a CNN-based machine learning model to predict Raman spectra from KPFM and AFM data. During the training process, the data were randomly divided into training, validation, and test sets in an 8:1:1 ratio. Training utilized the Adam optimizer and mean squared error (MSE) loss function, with 75 training epochs to prevent overfitting. After training, the model demonstrated high prediction accuracy on the test set, with an MSE of 0.15 and a mean absolute percentage error (MAPE) of no more than 15%.

3. Spatial Resolution Enhancement and Nanoscale Analysis

Using the trained model, the team generated Raman images with a spatial resolution of up to 50 nanometers, far exceeding the diffraction limit of traditional confocal Raman spectroscopy. By analyzing the E0 Raman mode in nanoscale wrinkle regions, the researchers observed the coexistence of compressive strain and tensile strain within the same wrinkle. Additionally, local strain calculations based on AFM images further validated the physical reasonableness of the model’s predictions.

4. Explainable AI Analysis

To understand the model’s decision-making process, the team employed explainable AI (XAI) techniques, observing changes in Raman spectra by perturbing input data (e.g., AFM height and KPFM work function). The analysis revealed that AFM data primarily influence local behaviors in nanoscale wrinkle regions, while KPFM data significantly contribute to the overall features of Raman spectra. This synergistic effect enables the model to effectively capture nanoscale structural perturbations.

5. Quantum Mechanical Calculation Validation

To further validate the findings, the team used density functional theory (DFT) to calculate changes in Raman activity and work function under different strains in WS2 monolayers. The results showed that tensile strain causes a red shift and increased activity in the E0 Raman mode, while compressive strain leads to a blue shift and decreased activity. These findings align with the XAI analysis, further confirming the reliability of the model.

Key Results

  1. Spatial Resolution Enhancement: The machine learning model improved the spatial resolution of Raman spectroscopy from several micrometers to 50 nanometers, successfully revealing the strain distribution in nanoscale wrinkles of WS2 monolayers.
  2. Strain Distribution Analysis: The study discovered the coexistence of compressive and tensile strains within nanoscale wrinkles, consistent with local strain calculations from AFM images.
  3. XAI Reveals Model Decisions: XAI analysis demonstrated that AFM data dominate local strain effects, while KPFM data influence the overall features of Raman spectra.
  4. Quantum Mechanical Validation: DFT calculations confirmed the impact of strain on Raman activity and work function, validating the physical reasonableness of the model’s predictions.

Conclusions and Significance

By integrating machine learning with spectroscopic techniques, this study successfully achieved nanoscale resolution enhancement in Raman spectroscopy and revealed the physical mechanisms of nanoscale structural perturbations in WS2 monolayers. This approach not only provides a new tool for nanoscale characterization of 2D materials but also offers important insights for defect engineering and the development of high-performance semiconductor materials in materials science.

Research Highlights

  1. High-Resolution Raman Imaging: The machine learning model enabled Raman imaging with a spatial resolution of 50 nanometers, breaking the diffraction limit of traditional techniques.
  2. Strain Distribution Revelation: The coexistence of compressive and tensile strains in nanoscale wrinkles was observed for the first time, providing new perspectives on the mechanical behavior of 2D materials.
  3. Integration of XAI and DFT: The combination of XAI and DFT calculations revealed the physical mechanisms of the machine learning model, enhancing the credibility of the research findings.
  4. Broad Applicability: This method can be extended to other material systems, offering a universal framework for nanoscale spectroscopic analysis.

Additional Valuable Information

The research team also provided detailed supplementary materials, including data processing methods, model architecture diagrams, and DFT calculation details, serving as a reference for other researchers. Furthermore, the study received funding from multiple institutions, including the National Research Foundation of Korea, highlighting its significant academic value and application prospects.