Multi-Level Random Telegraph Noise Analysis Protocols based on Machine Learning Algorithms

We present a three-step protocol to analyze multi-level random telegraph noise (RTN) degraded by white noise which we devised to interpret experimental data. Our novel process exploits machine learning techniques of kernel density estimation, Gaussian mixture model, and recurrent neural networks to extract RTN parameters with high accuracy. We successfully decompose multi-level RTN into constituent 2-level signals and quantify their transition amplitude and two switching time constants. In order to evaluate the algorithm performance, we generate 330 RTN signals for one-, two-, and three-trap cases with varying white noise amplitudes. We successfully demonstrate that the accuracy to extract the amplitudes for all generated signals of 20% white noise amplitude is >99% with less than 1 minute of processing by a standard desktop computer. As the white noise level increases, our error for multi-trap RTN rises, but its median still lies below 20%.