Variational autoencoders understand knot topology.

Area of Science:

• Polymer Physics
• Computational Chemistry
• Machine Learning

Background:

• Standard mathematical methods struggle with identifying knots in complex polymer structures.
• Machine learning offers promising alternatives for polymer topology analysis.

Purpose of the Study:

• To develop a hybrid supervised/unsupervised machine learning approach for polymer knot classification.
• To assess the model's ability to understand and utilize topological concepts.

Main Methods:

• Introduction of a variational autoencoder enhanced with a knot type classifier (VAEC).
• Training the VAEC on labeled 3D polymer configurations.
• Evaluating the VAEC's performance on knot classification and chirality detection.

Main Results:

• The VAEC successfully organized knots in its latent representation, capturing topological features like chirality and knot families.
• The model accurately distinguished the chirality of previously undetected knots (9_42 and 10_71).
• The VAEC's latent space enabled faithful generation of knotted polymer configurations.

Conclusions:

• Hybrid machine learning models can effectively capture complex topological features of entangled filaments.
• The VAEC demonstrates the potential for advanced knot analysis and generative modeling in polymer science.
• This approach offers a simulation-free method for reconstructing and producing knotted polymer configurations.