Reconfigurable Mixed-Kernel Heterojunction Transistors for Personalized Support Vector Machine Classification

The continual increase in the volume of data generated by sensors, robots, and mobile devices poses a challenge for cloud computing centers in terms of power costs, efficiency, accessibility, and scalability. In particular, artificial intelligence (AI) based classification has shown tremendous potential in a variety of applications such as signal processing and image recognition but often has huge data and energy costs due to the underlying von Neumann architecture of complementary metal-oxide-semiconductor (CMOS) chips. In contrast, the use of edge computing to locally analyze and process data has the potential to alleviate the burden on large data centers, while also improving latency, security, and privacy. Because edge computing resources are often limited, edge-based hardware needs to be more power-efficient than conventional graphics processing units (GPUs), application-specific integrated circuits (ASICs), or field programmable gate arrays (FPGAs). <br/>This presentation will discuss edge computing at the device level using a highly reconfigurable mixed-dimensional dual-gated MoS₂/carbon nanotube (CNT) van der Waals heterojunction device achieved in a self-aligned manner. By tailoring the degree of electric-field screening through control over CNT density and overlapping heterojunction area, the heterojunction device is shown to generate tunable Gaussian, sigmoid, and mixed Gaussian/sigmoid functions. These heterojunction-generated functions are then used as kernels for support vector machine (SVM) based arrhythmia detection from electrocardiogram (ECG) signals. For a single device, mixed kernels were found to outperform purely Gaussian and purely sigmoid kernels. Bayesian optimization was further used to determine the optimal Gaussian/sigmoid hyperparameters and mixing ratio for a full mixed-kernel library that was generated from the integration of two devices, resulting in exceptionally high classification accuracies. Compared to existing literature on Gaussian transistors and CMOS circuits, our mixed-kernel heterojunction (MKH) transistors show unprecedented tunability in a compact four-terminal geometry with close to two orders of magnitude fewer circuit elements. In this manner, CNT-MoS₂ MKHs have the potential to be widely applicable to SVM classification in diverse wearable and edge applications.&lt;!--![endif]----&gt;