Synthetic data-driven deep learning for label-free autonomous atomic force microscopy.

Area of Science:

• Materials Science
• Nanotechnology
• Biophysics

Background:

• Atomic force microscopy (AFM) is crucial for nanoscale characterization but limited in high-throughput studies due to expert dependency and lack of labeled data.
• Data-driven automation in AFM is hindered by the scarcity of annotated experimental datasets.

Purpose of the Study:

• Introduce SimuScan, a synthetic-data-driven framework to enable reliable AFM feature identification, segmentation, and targeted imaging.
• Overcome the need for large manually labeled experimental AFM datasets.
• Facilitate data-driven automation in AFM analysis and materials discovery.

Main Methods:

• SimuScan generates high-fidelity synthetic AFM images with tunable morphologies and realistic experimental artifacts (e.g., tip-sample convolution, noise, debris).
• Utilizes synthetic datasets for scalable, label-free training of deep learning models for AFM analysis.
• Integrates SimuScan-trained models into data-driven AFM workflows for automated analysis and guided imaging.

Main Results:

• Demonstrated scalable, label-free training of deep learning models using synthetic AFM data.
• SimuScan-trained models successfully located and analyzed nanoscale structures across diverse sample types (nanostructured surfaces, DNA assemblies, bacterial cells).
• Achieved robust generalization and minimal operator intervention in AFM analysis.

Conclusions:

• SimuScan provides a reliable method for AFM feature identification and analysis without extensive manual labeling.
• Synthetic data generation is a viable strategy to improve the reliability of downstream models in autonomous microscopy.
• This framework accelerates nanoscale characterization and supports data-driven materials discovery.