검색 상세

Development of artificial synapses based on organic/inorganic ferroelectric transistor devices for intelligent neuromorphic computing

초록/요약

The implementation of the brain-inspired neuromorphic computing system is an emerging field of research in the era of the 4th industrial revolution due to the inefficiency of von Neumann computing architectures in processing complex and unstructured information. Emulating the essential synaptic functions is widely recognized as the requisite first step in hardware building brain-like computers and artificial intelligent system. Heretofore, diverse types of artificial synapse have been proposed as a building block for realization of neuromorphic computing. Among them, ferroelectric transistor-based artificial synapses have the advantages of good stability, relatively easy controllability of key parameters, clear operation mechanisms. In particular, synaptic weight update can be performed without interrupting the signal transmission process, enabling concurrent learning unlike a two-terminal device. In addition, the analog modulation of synaptic weight with low power consumption can be emulated by the virtue of the excellent multidomain polarization switching capability of ferroelectric materials. Such unique capabilities can offer endless opportunities for numerous applications in various fields. Nevertheless, the development of ferroelectric transistor-based artificial synapses based on different materials is still in its infancy. Herein, we introduce the novel artificial synapses based on organic/inorganic ferroelectric transistor devices. Firstly, we developed 1D fiber-shaped multi-synapses comprising ferroelectric organic transistors fabricated on a 100-μm Ag wire and used them as multi-synaptic channels in an e-textile neural network for wearable neuromorphic applications. The device mimics diverse synaptic functions with excellent reliability even under 6000 repeated input stimuli and mechanical bending stress. Notably, the 1D multi-synapses achieved up to ~90 and ~70% recognition accuracy for MNIST and electrocardiogram patterns, respectively, even in a single-layer neural network, and almost maintained regardless of the bending conditions. Secondly, an integrated artificially intelligent tactile learning electronic skin (e-skin) based on arrays of ferroelectric-gate field-effect transistors with dome-shape tactile top-gates, which can simultaneously sense and learn from a variety of tactile information, is introduced. To test the e-skin, tactile pressure is applied to a dome-shaped top-gate that measures ferroelectric remnant polarization in a gate insulator. This results in analog conductance modulation that is dependent upon both the number and magnitude of input pressure-spikes, thus mimicking diverse tactile and essential synaptic functions. Additionally, it has a low variability of only 3.18%, resulting in high-performance and robust tactile perception learning. The 4 × 4 device array is also able to recognize different handwritten patterns using 2-dimensional spatial learning and recognition, and this is successfully demonstrated with a high degree accuracy of 99.66%, even after considering 10% noise. Thirdly, we designed a new class of artificial synaptic architecture, a graphene barristor with an CMOS-compatible inorganic ferroelectric ultra-thin Hf0.5Zr0.5O2 film embedded as a gate dielectric layer to allow the multilevel modulation of the non-volatile channel conductance. Noticeably, the unique vertical architecture of ferroelectric synaptic barristor facilitates high-density large-scale integration with crossbar-like structure while maintaining the advantages of multiterminal transistors: operation stability, nondestructive read-out, simultaneous weight updating, and unnecessariness of additional selector components. We demonstrated the essential synaptic behaviors with relatively low power consumption, high operating speed and excellent reliability even under 5000 repeated input stimuli. In addition, the recognition accuracy of 91 % was achieved for MNIST clothes patterns while reducing energy consumption by 86 % thanks to diagonal-gated array structure utilizing the facile structural tunability. We believe that the research introduced in this paper will bring widespread interest and have a strong impact on the development of next-generation synaptic devices for the efficient neuromorphic system. Novel approaches to the ferroelectric transistor-based synapses with various applications presented in this thesis will provide new insights and inspired guides to the practical implementation of brain-like artificial intelligent computing.

more

목차

Chapter 1. Introduction 1
1-1. Introduction of neuromorphic engineering 1
1-1-1. Research background and development history 1
1-1-2. Advantages and novelty of neuromorphic architecture 2
1-2. Neural signal transmission process of the neural network 3
1-2-1. Structure of biological neural network and synapse 3
1-2-2. Synaptic plasticity 5
1-2-3. Implementation of artificial neural network and synapse 7
1-3. Artificial synapse based on ferroelectric transistor device 8
1-3-1. Device structure 8
1-3-2. Mechanisms of switching behavior 10
1-3-3. Implementation of diverse synaptic functions 13
Chapter 2. Fiber-shaped artificial synapse based on organic ferroelectric transistor for wearable neuromorphic application 18
2-1. Introduction 18
2-1-1. Wearable electronics 18
2-1-2. 1D artificial synapse for e-textile neuromorphic platform 19
2-2. Experiment 21
2-2-1. Device fabrication process and device structure 21
2-2-2. Method for pattern learning algorithms 24
2-3. Results and discussion 28
2-3-1. Implementation of synaptic behavior 28
2-3-2. Implementation of synaptic characteristics with mechanical flexibility 32
2-3-3. Recognition simulation for MNIST handwritten digit classification 35
2-3-4. ECG patterns diagnosis model for u-healthcare applications 43
2-4. Conclusion 50
Chapter 3. Single unit-based ferroelectric organic tactile synapse for artificial intelligent e-skin applications 51
3-1. Introduction 51
3-1-1. Artificial neural network for tactile perception 51
3-1-2. Challenges in implementation of neuromorphic tactile system 53
3-2. Experiment 54
3-2-1. Device fabrication process 54
3-2-2. Methods for pattern learning algorithms 57
3-3. Results and discussion 59
3-3-1. Device structure and switching mechanism 59
3-3-2. Implementation of essential synaptic functions 66
3-3-3. Implementation of tactile synaptic behaviors 70
3-3-4. Recognition simulation for MNIST digit classification 72
3-4. Conclusion 86
Chapter 4. Ferroelectric HZO-based synaptic barristor enabling CMOS-compatible neuromorphic system 87
4-1. Introduction 87
4-1-1. Ferroelectricity of hafnium oxide thin film 87
4-1-2. Vertical barristor-based artificial synapse 90
4-2. Experiment 92
4-2-1. Device fabrication process 92
4-2-2. Methods for pattern learning algorithms 95
4-3. Results and discussion 97
4-3-1. Device structure and material characteristics 97
4-3-2. Nonvolatile memory characteristics and switching mechanism 100
4-3-3. Implementation of synaptic behaviors 104
4-3-4. MNIST pattern learning simulation 110
4-3-5. Potential applications: diagonal-gated synapse array for efficient convolutional neural
network 113
4-4. Conclusion 119
Chapter 5. Conclusion 119
5-1. Summary 120
5-2. Outlook 122
Chapter 6. Bibliography 126

more
반출 Meta View 목록