S4: Self-Supervised Learning of Spatiotemporal Similarity

Area of Science:

• Scientific Visualization
• Machine Learning
• Data Science

Background:

• Visual exploration of large, unlabeled scientific datasets lacks effective similarity measures.
• Spatiotemporal data from simulations and experiments often exhibits complex behaviors.
• Interactive querying for similar behaviors is crucial for data analysis.

Purpose of the Study:

• To introduce a machine learning-driven approach for interactive, example-based similarity queries in spatiotemporal scientific data ensembles.
• To enable efficient visual exploration of large, unlabeled scientific datasets.
• To develop a method that does not require pre-defined similarity measures.

Main Methods:

• Utilized a self-supervised Siamese Neural Network to learn an expressive latent space for spatiotemporal behavior.
• Exploited the principle that nearby spatial locations often exhibit similar behaviors.
• Implemented an interactive, example-based querying system leveraging the learned latent space.

Main Results:

• The developed method effectively identifies similar spatiotemporal behaviors using user-provided examples.
• Qualitative and quantitative evaluations demonstrate superior performance compared to existing methods.
• The learned latent space provides a meaningful representation for similarity searches.

Conclusions:

• The ML-driven approach offers a powerful tool for interactive visual exploration of scientific data ensembles.
• This method addresses the challenge of finding similar behaviors in large, unlabeled spatiotemporal datasets.
• The self-supervised learning strategy effectively captures complex spatiotemporal patterns for similarity retrieval.