Deep neural network analysis models for complex random telegraph signals

Marcel Robitaille, HeeBong Yang, Lu Wang, Bowen Deng, Na Young KimScientific 2023 · DOI

TL;DR

Researchers have developed a robust three-step deep learning protocol to decode complex, multi-level Random Telegraph Signals (RTS) buried in heavy background noise. By using a "progressive knowledge-transfer" approach, the system automatically identifies the number of signal sources and their amplitudes to train a tailored neural network that outperforms traditional statistical models in high-noise environments.

What Problem Does This Paper Address?

In the world of nano-electronics, quantum computing, and even cellular biology, signals often "flicker" between discrete states. These are Random Telegraph Signals (RTS), typically caused by a single particle (like an electron) getting caught and released by a "trap."

While a single trap is easy to analyze, real-world devices often have multiple traps, creating a "multi-level" signal that looks like a chaotic staircase. When you add background white noise or "pink" (1/f) noise, the actual signal becomes nearly invisible. Traditional methods like Hidden Markov Models (HMMs) struggle with these complex levels and often require tedious manual initialization. This paper aims to automate the extraction of these signals directly from raw, noisy temporal data.

Methods

The authors propose a three-step "Progressive Knowledge-Transfer" protocol:

  1. Peak Estimation: Uses Kernel Density Estimation (KDE) to analyze the probability density of the raw signal. This identifies the number of active traps ($N_{traps}$).
  2. Amplitude Extraction: Fits a Gaussian Mixture Model (GMM) to the KDE results to determine the specific switching amplitudes ($\Delta_{RTS}$) for each trap.
  3. Pure RTS Reconstruction: This is the core innovation. The parameters from Steps 1 and 2 are used to generate tailored synthetic training data that mimics the specific noise and signal characteristics of the sample. This data trains a Deep Neural Network (DNN) to "digitize" the raw signal.

The researchers tested three DNN architectures:

  • Bi-directional Gated Recurrent Units (Bi-GRU)
  • Bi-directional Long Short-Term Memory (Bi-LSTM)
  • WaveNet (a convolutional architecture originally designed for audio)

Datasets

  • Synthetic Data: A massive library of 720 simulated RTS examples, each 1,000,000 time-steps long. These covered 1, 2, and 3-trap configurations with varying levels of white and pink noise (from 0% to 100% of the signal amplitude).
  • Real-World Data: To prove the model's utility, they tested it on experimental data from carbon nanotube (CNT) film devices and short-channel NMOS transistors.

Key Insights and Ideas

  • Tailored Training is Key: Instead of training one "universal" model, the authors use the first two steps to create a "digital twin" of the signal's environment. Training the DNN on this specific synthetic data allows it to learn the exact noise-to-signal ratio it needs to overcome.
  • Breaking the HMM Barrier: Traditional HMMs suffer from "flip errors" (swapping high/low states) and struggle as the number of levels increases. The DNN approach treats the problem as a sequence-labeling task, making it much more resilient to complexity.
  • Handling Anomalies: The model successfully processed "anomalous" RTS types—such as metastable signals (where states change behavior over time) and coupled traps (where one trap only activates if another is already full)—which typically break standard analysis tools.

Key Results

  • Trap Detection: The protocol achieved 100% success in identifying the number of traps for 1 and 2-trap systems, and 97.3% success for 3-trap systems.
  • Noise Resilience: Even with background noise equal to 100% of the signal amplitude, the amplitude error remained remarkably low (the third quartile of error stayed below 10%).
  • Performance: All three DNN models outperformed HMMs in high-noise scenarios. Among them, Bi-LSTM generally provided the highest accuracy for time-constant extraction, while WaveNet was the fastest to train (roughly 40% faster than Bi-LSTM).
  • Real-World Validation: The model successfully digitized complex, multi-trap signals from carbon nanotube devices that were previously difficult to interpret.

Limitations and Future Work

  • The Pink Noise Challenge: While the model handles white noise excellently, "pink noise" (1/f noise) remains difficult. When pink noise exceeds 30% of the signal amplitude, the accuracy of time-constant extraction drops significantly because the noise's low-frequency "wiggles" mimic actual signal transitions.
  • Filtering Dependencies: The authors suggest that future work should focus on specialized filtering techniques to remove low-frequency pink noise components without accidentally altering the RTS statistics.
  • Anomalous Optimization: While the current model handles anomalous RTS, the authors plan to develop designated steps to further optimize the algorithm for specific types of trap interactions.

Summary updated

Abstract

Time-fluctuating signals are ubiquitous and diverse in many physical, chemical, and biological systems, among which random telegraph signals (RTSs) refer to a series of instantaneous switching events between two discrete levels from single-particle movements. A reliable RTS analysis is a crucial prerequisite to identify underlying mechanisms related to device performance and sensitivity. When numerous levels are involved, complex patterns of multilevel RTSs occur and make their quantitative analysis exponentially difficult, hereby systematic approaches are often elusive. In this work, we present a three-step analysis protocol via progressive knowledge-transfer, where the outputs of the early step are passed onto a subsequent step. Especially, to quantify complex RTSs, we resort to three deep neural network architectures whose trained models can process raw temporal data directly. We furthermore demonstrate the model accuracy extensively with a large dataset of different RTS types in terms of additional background noise types and amplitude size. Our protocol offers structured schemes to extract the parameter values of complex RTSs as imperative information with which researchers can draw meaningful and relevant interpretations and inferences of given devices and systems.
BibTeX
@inproceedings{robitaille2023deep,
  author = {Marcel Robitaille and HeeBong Yang and Lu Wang and Bowen Deng and Na Young Kim},
  title = {{Deep neural network analysis models for complex random telegraph signals}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-37142-9},
}
Figure