Reconstruction of subsampled Landau Fan Measurement using Compressed sensing and Deep Learning

A flurry of recent developments has established 2D material and van der Waals heterostructure as an ideal solid-state platform for studying novel emergent phenomena as it offers versatile experimental controls for characterizing and manipulating the underlying order. However, experimental efforts to explore the multi-dimensional phase space, e.g. Landau fan measurement in graphene/hexagonal boron nitride structure, often require prolonged measurements. Here we explore a method to accelerate data acquisition by reconstructing an undersampled data set. Landau fan maps measured in graphene/hexagonal boron nitride structure will serve as an example to demonstrate the method’s viability and efficiency. To reconstruct undersampled data, we explore a traditional method as compressive sensing and two deep learning techniques: an enhanced deep residual network for a single image super-resolution model (EDSR), and a Noise2Noise neural network. For the sampling ratio 0.11, the EDSR and Noise2Noise methods indicate comparable performance, while both of the deep learning techniques demonstrate better performance compared to compressed sensing.