Nonvolatile Memristive Materials and Physical Modeling for In-Memory and In-Sensor Computing

Abstract

Separate memory and processing units are utilized in conventional von Neumann computational architectures. However, regarding the energy and the time, it is costly to shuffle data between the memory and the processing entity, and for data-intensive applications associated with artificial intelligence, the demand is ever increasing. A paradigm shift in traditional architectures is required, and in-memory computing is one of the non-von-Neumann computing strategies. By harnessing physical signatures of the memory, computing workloads are administered in the same memory element. For in-memory computing, a wide range of memristive material (MM) systems have been examined. Moreover, developing computing schemes that perform in the same sensory network and that minimize the data shuffle between the processing unit and the sensing element is a requirement, to process large volumes of data efficiently and decrease the energy consumption. In this review, an overview of the switching character and system signature harnessed in three archetypal MM systems is rendered, along with an integrated application survey for developing in-sensor and in-memory computing, viz., brain-inspired or analogue computing, physical unclonable functions, and random number generators. The recent progress in theoretical studies that reveal the structural origin of the fast-switching ability of the MM system is further summarized.

Keywords: brain-inspired neuromorphic computing; in-memory computing; in-sensor computing; molecular dynamics simulations; nonvolatile memristive materials; physical unclonable functions.

Figure 1

In‐memory computing and in‐sensor computation using MM systems. a) In a traditional computational platform, the data D are transferred to a processing element, when a function f is implemented on the D, resulting in substantial costs in power and time. b) With reference to the in‐memory computation, by harnessing physical characters of the memory hardware, the f(D) is administered in the same computational storage entity, therefore avoiding the requirement to transfer the D to the processing element. The MM technologies including phase‐change memory (PCM), resistive switching memory (RSM), and magnetic tunnelling memory (MTM) can operate as units of the computational‐storage element. c) Traditional sensory computational design. The analog outputs from different sensors are altered to digital signals, which are retained in the storage unit. The processing elements access the data from the storage and subsequently transfer the output signal back to the storage element for long‐term retention. d) The in‐sensor computational design. The processing operations are included in distinct sensors for front‐end processing. To avoid data transfer between sensors and processors, the sensor can cooperate to implement information aggregation and compression, and data processing.

Figure 2

Memristive material development. Various types of memrisitive material systems, i.e., phase‐change memory layers, resistive switching memory layers, and magnetic tunneling memory layers, are disclosed. This survey presents four key modes of research domains that utilizes the memristive material system.

Figure 3

PCM layers and AIMD simulations. a) The germanium–antimony–telluride ternary phase diagram revealing the variation of the crystallization temperature for different PCM compositions. Adapted with permission.^{[ 509 ]} Copyright 2011, Institute of Electrical and Electronics Engineers. b) Temperature variations result in reversible phase transitions between the crystallized phase and the disordered amorphous phase. The T _cryst represents the crystallization point, while the T _melt denotes the melting temperature. c,d) Schematic representations of the crystallized state (c) and the glassy amorphous state (d) of the PCM element. The heating above the T _melt to melt and subsequently quench the PCM volume rapidly results in switching the crystallized phase into the disordered‐amorphous phase. On the other hand, the heating above the T _cryst for an intermediate time to crystallize the PCM element leads to switching the amorphous phase into the ordered crystallized phase. e,f) Schematic diagrams of the (e) growth‐dominated and (f) nucleation‐dominated crystallization procedures. g) Snapshots of a GeSbTe model utilized in an ab initio molecular dynamics (AIMD) simulation at ≈600 K. g) Adapted with permission.^{[ 132 ]} Copyright 2011, American Physical Society. h) The progression of the effective radius (R _eff) of crystallized clusters from a crystallization trajectory of the GeSbTe. Adapted with permission.^{[ 132 ]} Copyright 2011, American Physical Society. i) The snapshots of the crystal growth from crystallized–amorphous boundaries in the AgInSbTe at ≈590 K. In the interface growth procedure, negligible substantial crystallized clusters were generated. Adapted with permission.^{[ 114 ]} Copyright 2014, The Authors, published by Springer Nature.

Figure 4

The RSM layer and physical‐modeling methodologies. a) Schematic representation of an oxygen‐vacancy filament in the Off state (top) and the On state (bottom) of an anion‐based RSM system. b) A high‐resolution transmission electron microscopy (HRTEM) image of a nanocrystalline Ti₄O₇ filament with a conical configuration in a Pt/TiO₂/Pt anion‐based RSM system. Adapted with permission.^{[ 510 ]} Copyright 2010, Springer Nature. c) Schematic diagram of a metal atom filament in the Off state (top) and the On state (bottom) of a cation‐based RSM system. d) The experiment reveals an Ag filament in a planar Au/SiO _x :Ag/Au cation‐based RSM system using in situ high‐resolution TEM, exhibiting the creation of a conducting Ag bridge and spontaneous withdrawal following bias removal. Adapted with permission.^{[ 406 ]} Copyright 2016, Springer Nature. e–g) Illustrations of the first‐principles approach to compute oxygen‐vacancy characters (e), the MD strategy to simulate ionic‐migration dynamics (f), and the kinetic Monte Carlo (KMC) methodology to model oxygen vacancy and interstitial‐ion distributions (g).

Figure 5

MTM layers and modeling strategies. a) The spin‐transfer torque (STT) process. Electrons are spin polarized by the magnetic moment in the fixed layer when the electron pass through the magnetically fixed layer. Moreover, electrons oriented in the opposite direction to that of the fixed layer are reflected, while the electron oriented in the same direction as that of the magnetically fixed layer flows through the tunneling barrier. The spin‐angular momentum of electrons that flow through the tunneling barrier are transferred to the magnetic‐free layer moments, thus altering the free‐layer magnetization or material state. b) The voltage‐controlled magnetic anisotropy (VCMA) mechanism. The charge depletion or accumulation at the magnesium oxide‐magnetic free layer interface alters the anisotropy of the magnetic‐free layer when an input voltage is administered across the magnetic‐tunneling junction, resulting in a precession, viz., oscillation, of the free‐layer magnetization. The free‐layer magnetization or material state is switched when the length of the input signal is half that of the precession process. c) The SOT mechanism. The magnetic‐tunneling junction is deposited on top of a high spin–orbit coupling (SOC) layer, to enable this process. Electrons with a specified spin direction aggregate near the top of the SOC layer, whereas the electron with an opposite spin direction accrues close to the bottom of SOC layers, when a current flows through the same SOC layer. The electrons assembled adjacent to the top of the SOC layer pass into the magnetic‐free layer and transfer the spin momenta to the same magnetic‐free layer, thus altering the free‐layer magnetization or material state. d,e) Simulation tools, viz., (d) the quantum ab initio model and (e) ab initio spin models, are utilized for MTM development, e.g., phenomena, materials, and other areas, to understand the material signature, and consequently steer the material optimization.

Figure 6

MM system signatures and application specifications. The plot compares the characters of three different types of MM systems and their influence on applications such as analogue in‐memory computing, spiking neural network (SNN) neurons, physical unclonable functions (PUFs), and artificial neural network (ANN) synapses. Short switching times, small dimensions, low programming energy, and excellent endurance are beneficial for the sample applications shown in this survey. PUFs are facilitated through a high degree of randomness, whereas SNN neurons and ANN synapses require minimized stochasticity. Furthermore, SNN neurons utilize spontaneous, rapid conductance relaxation, while ANN neurons and PUFs are assisted by prolonged retention.

Figure 7

MM‐powered brain‐inspired and analog computation. a) Variation in the output conductance of an MM system for different stimulus numbers. An increased number of stimuli results in an increasing structurally ordered volume, leading to an increase in the output conductance value. b) The arithmetic addition of addends “three” and “four” through the stimulus aggregation in an MM element. c) An integrate‐and‐fire neuron, wherein the integration was performed based on accruing input spikes in an MM unit. d) The synaptic potentiation utilizing the gradual structural ordering of an MM entity. a,c,d) Adapted with permission.^{[ 398 ]} Copyright 2016, Springer Nature. e) The schematic diagram compares illustrations of biological neurons with phase‐change systems, demonstrating that a synapse between two neurons can be mimicked utilizing a single phase‐change system, and potentiation and depression can be generated via crystallization and amorphization of GeSbTe. f) The experiment demonstrates the use of phase‐change systems to mimic the spike‐timing dependent plasticity rule, where conduction depicts synaptic weight. e,f,) Adapted with permission.^{[ 392 ]} Copyright 2012, American Chemical Society.

Figure 8

The physical unclonable functions and random number generators based on MM systems. a) Conductance switching in traditional RSM layers may be stochastic and is controlled by varying the length and voltage of programming stimuli. b) Illustration of a digital circuitry that harnesses the MM stochasticity for the creation of true random numbers. The MM unit is linked to a resistive element in a voltage divider configuration type, and a specified‐duration set stimulus is administered to the MM unit. The comparator element generates an output bit “1” owing to the set transition in the MM unit after a stochastic delay period. The counter in units of a designated clock time determines the variation between the stimulus length and the delay duration. If the variation between the stimulus length and the delay duration is an even multiple of the clock time, the output bit “0” results, whereas the output bit “1” occurs for the case when the variation between the stimulus length and the delay duration is an odd multiple of the clock time. A stochastic‐bit assemblage appears upon the application of a series of set stimuli. b) Adapted under the terms of the CC‐BY Creative Commons Attribution 4.0 International license (https://creativecommons.org/licenses/by/4.0).^{[ 534 ]} Copyright 2017, The Authors, published by Springer Nature. c) Schematic representation of a Bayesian network, wherein each node describes random variables, and every link denotes a specified dependence among nodes, determined based on transitional conditional probabilities. From a stipulated observation, these networks can be harnessed to determine the probability of concealed origins. Utilizing stochastically switching MM systems, the probability distribution required for implementing probabilistic inferences can be created. For instance, model probabilities can be encrypted within Poisson‐distributed binary bit ensembles created in an MM system. By multiplying two separate bit assemblages using an AND gate, corresponding calculations including the intersection operation can be administered. Adapted under the terms of the CC‐BY Creative Commons Attribution 4.0 International license (https://creativecommons.org/licenses/by/4.0).^{[ 535 ]} Copyright 2017, The Authors, published by Springer Nature. d) Schematic depiction of MM crossbar arrays utilized for creating physical unclonable functions (PUFs). To achieve an increased group of challenge‐response pairs (CRPs), a broad distribution of conductance values is harnessed. For instance, a challenge comprises an N‐bit vector administered to N rows in an N × N crossbar PUF. The output current from N columns was recorded and modified to an N‐bit response. 2 ^N is the theoretical number of CRPs. Adpated with permission.^{[ 536 ]} Copyright 2016, Institute of Electrical and Electronics Engineers.

Figure 9

MM systems for in‐sensor computing. a) Illustration of the low‐level visualization processing methodology using edge‐removal and contrast improvement processes. The visualization processing and detection schemes are integrated in the same sensor for in‐sensor computation. b) Schematic diagram of the low‐level acoustic processing approach. To achieve clear signals for subsequent processing, the pure acoustic signal is refined utilizing band‐pass filters with a noise suppresion in every channel. c) Schematic representation of low‐level aromatic processing strategies. The distinction of gas modes in subsequent high‐level processing methodologies is influenced by the baseline difference in pure sensory information. The baseline was extracted from input signals. d,e) Various sensory response characters for in‐sensor computing. Linear response signatures enhance precision in in‐sensor computing on artificial neural networks, as shown in (d), whereas nonlinear sensor signatures, e.g., sublinear, threshold, and superlinear, generate intensity‐dependent information, enabling spatiotemporal information encoding and improving computation functions at sensory terminals, as depicted in (e). d,e) Adapted with permission.^{[ 465 ]} Copyright 2022, Springer Nature.