Machine learning in a data-limited regime: Augmenting experiments with synthetic data uncovers order in crumpled...

Area of Science:

• Physics
• Materials Science
• Data Science

Background:

• Machine learning (ML) excels with complex, high-dimensional data.
• ML application is limited in experimental systems with scarce or costly data.
• Crumpled thin sheets exhibit complex spatial order in crease networks.

Purpose of the Study:

• To develop a strategy for applying machine learning in data-limited experimental settings.
• To investigate the effectiveness of augmenting scarce experimental data with simulated data.
• To improve the predictive capabilities of machine learning models in experimental research.

Main Methods:

• Augmenting scarce experimental data with synthetically generated data from a simpler sister system.
• Simulating rigid flat-folded sheets to generate inexhaustible training data.
• Utilizing machine learning techniques to analyze local order in crease networks of crumpled sheets.
• Testing the improved predictive power on a pattern completion task.

Main Results:

• Machine learning techniques proved effective even in a data-limited regime for crumpled sheet analysis.
• Augmenting experimental data with simulated data significantly improved predictive power.
• The strategy demonstrated the utility of ML in bench-top experiments with limited data.
• Common statistical properties between systems facilitated successful data augmentation.

Conclusions:

• A novel strategy effectively overcomes data scarcity limitations in experimental machine learning.
• Simulated data from simpler systems can successfully augment experimental datasets.
• This approach enhances the applicability of machine learning in resource-constrained experimental research.
• The findings pave the way for broader ML adoption in experimental science.