A Deep Learning Approach for Molecular Classification Based on AFM Images

Area of Science:

• Surface Science
• Computational Chemistry
• Machine Learning

Background:

• Non-contact atomic force microscopy (AFM) offers high resolution for molecular imaging.
• Unambiguous identification of molecules from AFM images alone is challenging.
• Current interpretation methods require prior knowledge or extensive analysis.

Purpose of the Study:

• To develop an automated method for classifying AFM experimental images.
• To address the limitations of standard pattern recognition models in AFM image analysis.
• To improve the accuracy and generalizability of AFM image classification.

Main Methods:

• A deep learning model, specifically a variational autoencoder (VAE), was developed and trained.
• The model was primarily trained on a theoretically generated dataset.
• Few experimental AFM images were used to incorporate characteristic features into the training set.

Main Results:

• The developed deep learning model demonstrated optimal depth for accurate AFM image classification.
• The model showed a high ability to generalize across both theoretical and experimental datasets.
• A variational autoencoder (VAE) effectively integrated experimental features for enhanced classification.

Conclusions:

• The study presents a significant step towards automated AFM image classification.
• The VAE-based approach enhances the accuracy of identifying molecular systems from AFM data.
• This method facilitates more reliable and efficient analysis of AFM experimental images.