Using an autoencoder to reduce experimental noise

Experimental noise is ubiquitous in the sciences. Some of this noise is random, however some is systematic. We use machine learning to try to remove the systematic noise in experimentally found scans of crumpled sheets of paper in order to be able to study this classical disordered system. Properties that were previously uncovered using other denoising mechanisms were that crease mileage is a state variable of crumpling, is independent of crumpling history, and has a logarithmic scaling property. Through the use of an autoencoder, we were able to recover the scaling properties of crease mileage while improving the automation of the denoising process. This denoiser is able to outperform other methods on extremely noisey sheets. The removal of noise from our data is essential in order to be able to further explore the characteristics of the crease networks in a crumpled sheet.