Molecular Identification from AFM Images Using the IUPAC Nomenclature and Attribute Multimodal Recurrent Neural...

Area of Science:

• Chemical Physics
• Materials Science
• Computational Chemistry

Background:

• Traditional molecular identification relies on spectroscopic methods like NMR and mass spectrometry.
• These methods are often applied to molecular ensembles in solution.
• Limitations exist in resolving individual molecule structures and compositions with current techniques.

Purpose of the Study:

• To introduce a novel chemical characterization approach using high-resolution atomic force microscopy (HR-AFM).
• To demonstrate HR-AFM's capability in resolving the internal structure of individual quasiplanar organic molecules.
• To leverage machine learning for complete molecular identification from HR-AFM data.

Main Methods:

• Utilizing noncontact atomic force microscopy with CO-functionalized tips (HR-AFM) for molecular imaging.
• Acquiring stacks of constant-height HR-AFM images to capture chemical information.
• Employing multimodal recurrent neural networks (M-RNN) for image analysis and molecular identification.
• Training the M-RNN model on a large dataset of theoretical molecules and AFM images.

Main Results:

• HR-AFM imaging successfully resolves the internal structure, composition, and bond topology of individual molecules.
• Machine learning algorithms can accurately disentangle contributions to AFM contrast from molecular properties.
• The M-RNN model achieved high accuracy in identifying molecules by generating IUPAC names from AFM images.
• Validation performed using extensive theoretical and some experimental AFM images.

Conclusions:

• HR-AFM combined with deep learning offers a powerful new paradigm for molecular identification.
• This approach overcomes limitations of traditional spectroscopic methods, especially for on-surface synthesis.
• The study highlights the potential of AI in automating chemical compound identification using AFM imaging.