東華大學圖書館 |

Building a Better Brain: Atomic Scale Control for Use in Neuromorphic Memristors.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Building a Better Brain: Atomic Scale Control for Use in Neuromorphic Memristors./
作者:	Goul, Ryan.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2022,
面頁冊數:	111 p.
附註:	Source: Dissertations Abstracts International, Volume: 84-08, Section: B.
Contained By:	Dissertations Abstracts International84-08B.
標題:	Condensed matter physics. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30242376
ISBN:	9798368477442

Building a Better Brain: Atomic Scale Control for Use in Neuromorphic Memristors.
Goul, Ryan.

Building a Better Brain: Atomic Scale Control for Use in Neuromorphic Memristors. - Ann Arbor : ProQuest Dissertations & Theses, 2022 - 111 p.

Source: Dissertations Abstracts International, Volume: 84-08, Section: B.

Thesis (Ph.D.)--University of Kansas, 2022.

Computational demands of both the scientific and corporate world have expanded tremendously over the past several decades. Companies such as Google, IBM, and Intel to name a few are actively searching for alternatives to conventional transistor-based computing architectures. As the dimensions of transistors approach atomic scales, they will no longer be able to be further scaled down, more importantly though, the added heat and energy consumption of these high-density arrays of devices will become prohibitive. Several alternative computing schemes such as quantum computing and neuromorphic computing have been suggested as the next evolution in computing.Neuromorphic computing seeks to improve computing efficiency in computer vision, voice pattern recognition, and other pattern recognition applications through emulation of one of the most efficient computers observed, the human brain. Using the neurons and synapses of the brain, humans can easily make judgements and decisions that would take enormous amounts of data and power for a trained artificial intelligence to make using conventional computer hardware. For this reason, much research has been done toward the fabrication of these artificial neurons and synapses. The main device made to mimic these biological components is the memristor, a two-terminal device whose resistance can be manipulated using electric current and field. This resistance change mimicking the synaptic weight change experienced in biological systems as the brain learns.Memristors for storage or for neuromorphic computational use are required to have a good on/off ratio (high resistance/low resistance) of at least 102, fast switching speeds (10's of nanoseconds), high write/read endurance (reported as high as 1011 cycles), and high memory retention times (on the order of months if not years) defining memristors as nonvolatile memory with no requirement for power to maintain stored data. However, balancing the high resistance needed in the on/off ratio, fast switching speeds, and desire for lower switching energies is a materials engineering issue since many materials cannot achieve all of these attributes simultaneously to the desired levels. Current memristor design mainly uses metal oxides deposited using physical vapor deposition and atomic layer deposition (ALD) at thicknesses of 5 - 10 nm, due to depositing the oxide on a poor-quality interface resulting in an inhomogeneous and leaky barrier in thinner oxides. Control of this metal/insulator (MI) interface would allow for atomic tunability in the qualities of the resistive switching layer. By beginning with a defect free interface, it becomes possible to control the density and type of defects introduced to the oxide layer, controlling vacancy formation and diffusion, while avoiding unintended defects from a poor MI interface.This MI interface also plays a critical role in other thin film devices such as capacitors, Josephson Junctions (JJs), Magnetic Tunnel Junctions (MTJs), and complementary metal oxide semiconductors (CMOS), among others. In order to solve this MI issues, our group developed an all in vacuo ALD process to better control this delicate MI interface. Using this technique, our group successfully fabricated a functioning ALD-Al2O3 JJ as thin as 0.2 nm, an MTJ using a 0.6 nm thick ALD-Al2O3 barrier, and ~4 nm thick capacitors of MgO and Al2O3 showing dielectric constants near the bulk value and effective oxide thicknesses comparable to high-k HfO2. These results in attaining high quality barriers through MI interface control laid the groundwork for this thesis focused on attaining tunable memristors by starting with a pristine oxide and then introducing defects.This thesis work exploits in vacuo ALD by inserting layers of ALD-MgO into pristine ALD-Al2O3 at various locations in the barrier in various sequences. Density functional theory (DFT) simulations of MgO/Al2O3 done by collaborators, showed that not only does inserting MgO lower the Fermi level in the sample (increasing resistance), but it also makes the formation of oxygen vacancies more favorable. Additional experiments utilizing scanning tunneling spectroscopy further confirmed that the initial ALD material (Al2O3 or MgO) grown on the MI interface was a determining factor in defect formation. These results confirmed that atomic tuning was dependent not only on the material composition of the resistive switching layer, but also dependent on the specific sequence of ALD layer growth. Through this atomic tuning, the on/off ratio of our memristors has been shown as tunable between 10 - 104, and the pulsed switching speed of these memristors has been demonstrated between single microseconds up to tens of milliseconds. However, this is only the beginning as this technique can readily be applied to a multitude of other ALD-grown materials. Ga2O3 is one such material that could replace ALD-Al2O3 and have a more easily tunable bandgap resulting in greater control of oxygen vacancy diffusion and therefore control over the switching dynamics. In a similar way, transparent materials could be grown using ALD to act as neuromorphic photodetectors triggered by incoming light as opposed to a conventional voltage pulse.With the number of materials growable using ALD, the potential amount of material combinations is vast. In the future we plan on exploring a multitude of combinations using different materials beyond the ones mentioned above, exploiting their various unique properties. Plans for applying our current atomically tunable memristors to neuromorphic computing by fabricating large crossbar arrays have already been made. To this purpose, photomasks were created and work has started towards fabricating a 64-bit crossbar array and applying it to basic image recognition tasks to begin with. Naturally this will carry more interesting challenges such as sneak path leakage current through neighboring memristors in the array, but existing solutions for this should be implementable using our existing fabrication methods, and it may be possible to improve on these commonly implemented solutions using atomic tuning.

ISBN: 9798368477442Subjects--Topical Terms:

3173567
Condensed matter physics.
Subjects--Index Terms:

Atomic layer deposition

Building a Better Brain: Atomic Scale Control for Use in Neuromorphic Memristors.
LDR:07305nmm a2200385 4500 001 2365908
005 20240918072801.5
006 m o d
007 cr#unu||||||||
008 241017s2022 ||||||||||||||||| ||eng d
020 $a 9798368477442
035 $a (MiAaPQ)AAI30242376
035 $a AAI30242376
040 $a MiAaPQ $c MiAaPQ
100 1 $a Goul, Ryan. $0 (orcid)0000-0003-0386-9986 $3 3707186
245 1 0 $a Building a Better Brain: Atomic Scale Control for Use in Neuromorphic Memristors.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2022
300 $a 111 p.
500 $a Source: Dissertations Abstracts International, Volume: 84-08, Section: B.
500 $a Advisor: Wu, Judy.
502 $a Thesis (Ph.D.)--University of Kansas, 2022.
520 $a Computational demands of both the scientific and corporate world have expanded tremendously over the past several decades. Companies such as Google, IBM, and Intel to name a few are actively searching for alternatives to conventional transistor-based computing architectures. As the dimensions of transistors approach atomic scales, they will no longer be able to be further scaled down, more importantly though, the added heat and energy consumption of these high-density arrays of devices will become prohibitive. Several alternative computing schemes such as quantum computing and neuromorphic computing have been suggested as the next evolution in computing.Neuromorphic computing seeks to improve computing efficiency in computer vision, voice pattern recognition, and other pattern recognition applications through emulation of one of the most efficient computers observed, the human brain. Using the neurons and synapses of the brain, humans can easily make judgements and decisions that would take enormous amounts of data and power for a trained artificial intelligence to make using conventional computer hardware. For this reason, much research has been done toward the fabrication of these artificial neurons and synapses. The main device made to mimic these biological components is the memristor, a two-terminal device whose resistance can be manipulated using electric current and field. This resistance change mimicking the synaptic weight change experienced in biological systems as the brain learns.Memristors for storage or for neuromorphic computational use are required to have a good on/off ratio (high resistance/low resistance) of at least 102, fast switching speeds (10's of nanoseconds), high write/read endurance (reported as high as 1011 cycles), and high memory retention times (on the order of months if not years) defining memristors as nonvolatile memory with no requirement for power to maintain stored data. However, balancing the high resistance needed in the on/off ratio, fast switching speeds, and desire for lower switching energies is a materials engineering issue since many materials cannot achieve all of these attributes simultaneously to the desired levels. Current memristor design mainly uses metal oxides deposited using physical vapor deposition and atomic layer deposition (ALD) at thicknesses of 5 - 10 nm, due to depositing the oxide on a poor-quality interface resulting in an inhomogeneous and leaky barrier in thinner oxides. Control of this metal/insulator (MI) interface would allow for atomic tunability in the qualities of the resistive switching layer. By beginning with a defect free interface, it becomes possible to control the density and type of defects introduced to the oxide layer, controlling vacancy formation and diffusion, while avoiding unintended defects from a poor MI interface.This MI interface also plays a critical role in other thin film devices such as capacitors, Josephson Junctions (JJs), Magnetic Tunnel Junctions (MTJs), and complementary metal oxide semiconductors (CMOS), among others. In order to solve this MI issues, our group developed an all in vacuo ALD process to better control this delicate MI interface. Using this technique, our group successfully fabricated a functioning ALD-Al2O3 JJ as thin as 0.2 nm, an MTJ using a 0.6 nm thick ALD-Al2O3 barrier, and ~4 nm thick capacitors of MgO and Al2O3 showing dielectric constants near the bulk value and effective oxide thicknesses comparable to high-k HfO2. These results in attaining high quality barriers through MI interface control laid the groundwork for this thesis focused on attaining tunable memristors by starting with a pristine oxide and then introducing defects.This thesis work exploits in vacuo ALD by inserting layers of ALD-MgO into pristine ALD-Al2O3 at various locations in the barrier in various sequences. Density functional theory (DFT) simulations of MgO/Al2O3 done by collaborators, showed that not only does inserting MgO lower the Fermi level in the sample (increasing resistance), but it also makes the formation of oxygen vacancies more favorable. Additional experiments utilizing scanning tunneling spectroscopy further confirmed that the initial ALD material (Al2O3 or MgO) grown on the MI interface was a determining factor in defect formation. These results confirmed that atomic tuning was dependent not only on the material composition of the resistive switching layer, but also dependent on the specific sequence of ALD layer growth. Through this atomic tuning, the on/off ratio of our memristors has been shown as tunable between 10 - 104, and the pulsed switching speed of these memristors has been demonstrated between single microseconds up to tens of milliseconds. However, this is only the beginning as this technique can readily be applied to a multitude of other ALD-grown materials. Ga2O3 is one such material that could replace ALD-Al2O3 and have a more easily tunable bandgap resulting in greater control of oxygen vacancy diffusion and therefore control over the switching dynamics. In a similar way, transparent materials could be grown using ALD to act as neuromorphic photodetectors triggered by incoming light as opposed to a conventional voltage pulse.With the number of materials growable using ALD, the potential amount of material combinations is vast. In the future we plan on exploring a multitude of combinations using different materials beyond the ones mentioned above, exploiting their various unique properties. Plans for applying our current atomically tunable memristors to neuromorphic computing by fabricating large crossbar arrays have already been made. To this purpose, photomasks were created and work has started towards fabricating a 64-bit crossbar array and applying it to basic image recognition tasks to begin with. Naturally this will carry more interesting challenges such as sneak path leakage current through neighboring memristors in the array, but existing solutions for this should be implementable using our existing fabrication methods, and it may be possible to improve on these commonly implemented solutions using atomic tuning.
590 $a School code: 0099.
650 4 $a Condensed matter physics. $3 3173567
650 4 $a Nanotechnology. $3 526235
650 4 $a Materials science. $3 543314
653 $a Atomic layer deposition
653 $a Memristor
653 $a Neuromorphic computing
653 $a MI interface
653 $a Atomic scale
690 $a 0611
690 $a 0652
690 $a 0794
710 2 $a University of Kansas. $b Physics & Astronomy. $3 3344311
773 0 $t Dissertations Abstracts International $g 84-08B.
790 $a 0099
791 $a Ph.D.
792 $a 2022
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30242376