(Keynote) Machine Learning and High-Speed Circuitry in Thin Film Transistors for Sensor Interfacing in Hybrid Large-Area Electronic Systems

Abstract

Large-Area (and flexible) Electronics (LAE) on substrates such as plastic is a technologically attractive platform for deploying large numbers (potentially millions) of sensors over surfaces, from strain sensors on bridges and airplanes to flexible neural sensing caps worn on the head. Eventually the data has to end up in conventional IC’s for external communication and system integration, resulting in a “hybrid electronic” system. With large numbers, each sensor clearly cannot be directly connected to an IC, and a classic approach is active-matrix addressing enabled by low-performance thin film transistors (TFTs) fabricated directly on the plastic, architecturally similar to many conventional imaging arrays. Our work at Princeton has focused on using novel circuit, algorithmic, and machine-learning approaches implemented in the TFT/LAE domain, to greatly further reduce the number of hard physical connections between the TFT/LAE domain and conventional IC’s. The work has been demonstrated by prototype hybrid electronic sensor systems ranging from structural health monitoring, remote gesture sensing, brain wave (EEG) sensing, handwriting recognition from images, pattern recognition of physical objects by weight distribution, to isolation of individual speakers (when many are speaking) via microphone arrays. A first set of work is based on the analog properties of TFT’s, especially via high-frequency TFT oscillators. While large feature sizes and low process temperatures typically lead to “low performance” TFT’s, with simple self-alignment methods we have extended ZnO TFT performance [2,3] to currently well over 2 GHz. Along with HF thin film diodes, this enables near field coupling of signals and power between adjacent large area sheets [4] and sending sensor data “off-sheet” by wireless methods [5]. With sufficient bandwidth, frequency-hopping (spread spectrum) methods can be used to send far more data for a fixed number of connections than with a conventional active matrix approach [6]. Second, in many real-world problems, the goal of sensor arrays is often some kind of pattern recognition, either the in time domain (e.g. seizure detection via neural sensors) or in the spatial domain (e.g. initial detection of a crack on a wing). Machine-learning (ML) is a powerful tool for this, conventionally done in software. This requires sending all of the raw data from the sensors to the CMOS ICs and perhaps then to the cloud. Instead, we have implemented various levels of machine learning on the sensor data directly in TFTs in the large-area domain, so that only the results are sent to the IC domain portion of the system. ML conceptually consists of two steps: (i) feature extraction/data compression, and (ii) classification/inference. Data compression can be implemented in several ways using digital TFT circuits. Classification requires “memory” to be integrated into TFTs to save the “learning.” We have directly implemented both steps in TFT circuits for problems such as handwriting recognition from images in amorphous silicon sensors arrays, seizure detection, and tactile sensing [6-10]. Finally, we address the number of TFT’s required for different approaches and yield dependence. N Verma, Y Hu; L Huang, W Rieutort-Louis; J. Sanz Robinson; T Moy; B. Glisic; S. Wagner; J. Sturm, Proc. IEEE (2015)Y. Mehlman, N. Verma, and J.C. Sturm, Dig. Device Research Conf. (2017)Y. Mehlman, C. Wu, S. Wagner, N. Verma, and J.C. Sturm, Dig. Device Research Conf. (2019)Y. Hu; W. Rieutort-Louis; J. Sanz-Robinson; K. Song; J.C. Sturm; S. Wagner; N. Verma, 2012 Symp. VLSI CircuitsY. Hu, T Moy, L. Huang, W Rieutort-Louis, J Robinson, S Wagner, J Sturm, N Verma, Proc. 2014 IEEE Custom IC Conf.Y. Afsar, T. Moy, N. Brady, S. Wagner; J. Sturm, N. Verma, IEEE J. Solid-State Circuits (2018)T. Moy, W. Rieutort-Louis, S. Wagner, J.C. Sturm, N. Verma, IEEE Trans. Circuits and Systems I (2016)W. Rieutort-Louis; T.Moy, Z. Wang; S. Wagner; J.C. Sturm; N. Verma, IEEE J. Solid State Circuits (2016)T. Moy, L. Huang; W. Rieutort-Louis; C. Wu; P. Cuff; S. Wagner; J Sturm; N Verma, IEEE J. Solid-State Circuits (2017)L.E. Aygun, P. Kumar, Z. Zheng, T-S. Chen, S. Wagner, J.C. Sturm, and N. Verma, Dig. IEEE Inter. Solid State Circ. Conf. (2019)