Cutoff frequency determination algorithm for ferroelectric device pulse measurements with application to machine learning based prediction

Abstract

The accurate characterization of the ferroelectric switching dynamics is essential but challenging. Square-pulse measurements are commonly used, yet achieving both an accurate and precise level of measurement leads to weak current responses and poor signal-to-noise ratios. Since polarization is obtained by time-integrating these noisy currents, reliable denoising is critical. However, typical Conventional fast Fourier transform (FFT)-based noise filtering relies on manual determination by a person of the cutoff frequency in the Fourier domain, without a rigorous or reproducible criterion, which makes it difficult to automate the process. To address this issue, we propose a cutoff frequency determination algorithm that finds the frequency at which the inverse gradient of the logarithmic mean squared error (log MSE) curve is minimized. Furthermore, we developed a machine learning model combining 1-D dilated convolution and gated recurrent unit (GRU) layers to predict the cutoff frequency labeled by the proposed algorithm. This model performed particularly well in cases with a limited number of data points, where the algorithm tended to fail, effectively capturing both local and global patterns in signals. The post-processing following noise filtering with the determined cutoff frequency was implemented as well, enabling efficient handling of large measurement datasets.