Alibek Kaliyev

Student: Alibek Kaliyev

Project: Accelerated Fitting of Band-Excitation Piezorspone Force Microscopy

View: Research Poster (PDF)

Institution: Lehigh University

Major: Computer Science and Engineering

Advisor: Joshua Agar

Abstract

For nearly a decade, band-excitation piezoresponse force-based switching spectroscopy (BEPS) has been used to characterize ferroelectric switching and dynamic electromechanical responses of materials with nanoscale resolution. One of the key outputs of this technique is hyperspectral images of piezoelectric hysteresis loops, wherein there are one or more hysteresis loops at every pixel position. The challenge and dedication required to properly analyze data from these experiments have throttled the impact and widespread use of BEPS. To simplify the extraction of information from these datasets, a common approach involves fitting the piezoelectric hysteresis loops to an empirical function to parameterize the loops. This technique is computationally intensive, requiring more than 24 hours to process a single experiment on a single workstation with parallel processing, and is highly dependent on prior estimates.

My goal was to develop a fully unsupervised approach based on deep recurrent neural networks in the form of an autoencoder which is able to learn a sparse and thus interpretable latent space of piezoelectric hysteresis loops, revealing detailed physical insight that will allude to results from other analysis techniques. We developed generalized pre-trained models which can conduct feature extraction from piezoelectric hysteresis loops with minimal or potentially no training.

Using this approach, it is possible to deploy these methods for real-time analysis of BEPS, thus enabling experimentalists to improve their experimental efficiency and extract more information from these experiments.

While the work focuses on developing models and benchmarking their efficacy in BEPS, this methodology could be adapted to other spectroscopic imaging techniques.