Correction of AFM data artifacts using a convolutional neural network trained with synthetically generated data.

Area of Science:

• Surface Science
• Microscopy Techniques
• Computational Imaging

Background:

• Atomic Force Microscopy (AFM) inherently produces images with distortions due to its physical principles.
• Existing post-processing methods for AFM image correction often require specific knowledge of distortions and operator input.
• Developing autonomous and accurate artifact removal techniques is crucial for reliable AFM data interpretation.

Purpose of the Study:

• To develop and validate a machine learning-based approach for autonomous artifact removal in AFM microscopy images.
• To train a convolutional neural network (CNN) using purely synthetic data that mimics AFM distortions.
• To demonstrate the autonomous capability of the trained CNN in recognizing and correcting artifacts in unseen AFM images.

Main Methods:

• A CNN was trained on pairs of distorted AFM images and their corresponding ground truth images, generated using a simulator based on analytical descriptions of physical distortion phenomena.
• The CNN was trained exclusively on synthetic data, encompassing various combinations of AFM-induced distortions.
• The trained model was tested for its ability to autonomously identify and correct artifacts in new AFM images without prior knowledge of the specific distortions present.

Main Results:

• The developed machine learning approach successfully removed or suppressed artifacts in AFM images.
• The CNN demonstrated autonomous recognition and correction of distortions without requiring any operator input.
• Experimental results indicate the new method performs comparably to or better than conventional post-processing techniques.

Conclusions:

• The proposed machine learning method offers an autonomous and effective solution for correcting AFM image artifacts.
• Training CNNs on purely synthetic data generated from physical models is a viable strategy for artifact removal.
• The autonomous nature of this approach simplifies AFM data analysis and enhances reliability, with source code and datasets made publicly available.