Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing
The development of artificial intelligence (AI) has changed the overall landscape of information technology.[1] However, conventional computing hardware is energetically unfavorable to handle multifarious AI tasks because the frequent data transfer between the physically separated microprocessors an...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2023
|
| Online Access: | https://hdl.handle.net/1721.1/152130 |
| _version_ | 1826198112632307712 |
|---|---|
| author | Ryu, Seungchan |
| author2 | Li, Ju |
| author_facet | Li, Ju Ryu, Seungchan |
| author_sort | Ryu, Seungchan |
| collection | MIT |
| description | The development of artificial intelligence (AI) has changed the overall landscape of information technology.[1] However, conventional computing hardware is energetically unfavorable to handle multifarious AI tasks because the frequent data transfer between the physically separated microprocessors and data storage units results in ‘memory wall bottleneck’, leading to high energy consumption.[2] In contrast, a human brain is energy efficient because it fuses computing and storage functionalities at the same place. Inspired by the human brain, neuromorphic computing using non-volatile memristors has been proposed as a novel computing hardware to enable an energy-efficient computing.[3,4] This thesis contributes to develop a highly energy efficient and CMOS-compatible protonic electrochemical artificial synapse employing the nanoporous gadolinium (Gd)-doped ceria (Gd:CeO₂) as an inorganic solid proton electrolyte. The Gd:CeO₂ was carefully chosen based on the consideration of its high proton conductance, doping effect, and fabrication compatibility with current semiconductor technology.[5-8]
We synthesize the Gd:CeO₂ thin-film electrolytes using pulsed laser deposition. By the tuning of growth temperature and partial oxygen pressure, the obtained thin-film electrolytes have various structures ranging from the preferred oriented dense structures to nanoporous structures. We investigate proton conduction properties through these thin-films using electrochemical impedance spectroscopy, and show that nanoporous structures can improve the proton conductance by more than 5 orders of magnitude. We explain the plausible origin of the enhanced proton conduction via the water layer formation on the surface or porous columnar structures. We process to fabricate electrochemical memristors using the Gd:CeO₂ electrolyte, PdHX as the gate/proton reservoir and tungsten oxide (WO₃) as the switching channel, and characterize the performance of these devices. We examined the thickness effect of the electrolyte on the device operation. We show that devices built on porous Gd:CeO₂ electrolyte can perform highly energy-efficient computing as the protonic artificial synapse. The device required ~1 fJ/(µm²×nS) per synaptic event, which is energetically efficient compared to the conventional non-volatile memristors,[9,10] and also comparable to that of the human brain. Our analysis of the thermodynamic and kinetic contributions of computing energy show that the high proton conductance in the nanoporous structure results in the decrease in the ohmic drop in the electrolytes and mainly contributes to small computation energy. This study shows that Gd:CeO₂ electrolyte based memristors provide a promising solution of CMOS-compatible energy-efficient computing for AI applications. |
| first_indexed | 2024-09-23T10:59:31Z |
| format | Thesis |
| id | mit-1721.1/152130 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T10:59:31Z |
| publishDate | 2023 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1521302023-09-14T03:40:22Z Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing Ryu, Seungchan Li, Ju Massachusetts Institute of Technology. Department of Mechanical Engineering The development of artificial intelligence (AI) has changed the overall landscape of information technology.[1] However, conventional computing hardware is energetically unfavorable to handle multifarious AI tasks because the frequent data transfer between the physically separated microprocessors and data storage units results in ‘memory wall bottleneck’, leading to high energy consumption.[2] In contrast, a human brain is energy efficient because it fuses computing and storage functionalities at the same place. Inspired by the human brain, neuromorphic computing using non-volatile memristors has been proposed as a novel computing hardware to enable an energy-efficient computing.[3,4] This thesis contributes to develop a highly energy efficient and CMOS-compatible protonic electrochemical artificial synapse employing the nanoporous gadolinium (Gd)-doped ceria (Gd:CeO₂) as an inorganic solid proton electrolyte. The Gd:CeO₂ was carefully chosen based on the consideration of its high proton conductance, doping effect, and fabrication compatibility with current semiconductor technology.[5-8] We synthesize the Gd:CeO₂ thin-film electrolytes using pulsed laser deposition. By the tuning of growth temperature and partial oxygen pressure, the obtained thin-film electrolytes have various structures ranging from the preferred oriented dense structures to nanoporous structures. We investigate proton conduction properties through these thin-films using electrochemical impedance spectroscopy, and show that nanoporous structures can improve the proton conductance by more than 5 orders of magnitude. We explain the plausible origin of the enhanced proton conduction via the water layer formation on the surface or porous columnar structures. We process to fabricate electrochemical memristors using the Gd:CeO₂ electrolyte, PdHX as the gate/proton reservoir and tungsten oxide (WO₃) as the switching channel, and characterize the performance of these devices. We examined the thickness effect of the electrolyte on the device operation. We show that devices built on porous Gd:CeO₂ electrolyte can perform highly energy-efficient computing as the protonic artificial synapse. The device required ~1 fJ/(µm²×nS) per synaptic event, which is energetically efficient compared to the conventional non-volatile memristors,[9,10] and also comparable to that of the human brain. Our analysis of the thermodynamic and kinetic contributions of computing energy show that the high proton conductance in the nanoporous structure results in the decrease in the ohmic drop in the electrolytes and mainly contributes to small computation energy. This study shows that Gd:CeO₂ electrolyte based memristors provide a promising solution of CMOS-compatible energy-efficient computing for AI applications. Ph.D. 2023-09-13T18:07:17Z 2023-09-13T18:07:17Z 2023-02 2023-03-01T20:03:25.934Z Thesis https://hdl.handle.net/1721.1/152130 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Ryu, Seungchan Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing |
| title | Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing |
| title_full | Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing |
| title_fullStr | Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing |
| title_full_unstemmed | Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing |
| title_short | Protonic All-Solid-State Electrochemical Device as an Artificial Synapse for CMOS-Compatible Neuromorphic Computing |
| title_sort | protonic all solid state electrochemical device as an artificial synapse for cmos compatible neuromorphic computing |
| url | https://hdl.handle.net/1721.1/152130 |
| work_keys_str_mv | AT ryuseungchan protonicallsolidstateelectrochemicaldeviceasanartificialsynapseforcmoscompatibleneuromorphiccomputing |