Shining light on devices: New perspectives in non-volatile memory device and material investigation Open Access

Over the past 65 years, CMOS devices owing to increased demand for computing and miniaturization have pushed Moore’s law to its limit.¹ A significant amount of scientific effort has since been aimed toward alternative technologies with competitive or lower power consumption, higher speed, and enhanced endurance, with the International Roadmap for Devices and Systems (IRDS) suggesting that from 2025, reduced power consumption and heterogeneous integration will be the main drivers in the industry.² Novel developments include neuromorphic architectures inspired by brain structures, combining logic and memory, which have particular relevance given the recent surge of interest in new approaches to computing for artificial intelligence (AI). To address this momentous challenge, alternatives to CMOS, like oxide-based metal–insulator–metal (MIM) and heterojunction devices, are at the forefront.³ With the advent of oxide electronics, (Ba, Sr)TiO₃-based capacitors,⁴ lead zirconate titanate-based transducers,⁵ and hafnium (Hf)-based oxinitrides for high-k gate insulators⁶ mark significant landmarks in the development of functional oxide materials, particularly aimed at low-voltage applications and high scalability.

Materials that can be reversibly and reliably switched between electrical resistance states can be employed as non-volatile memory (NVMe) devices to encode information.⁷ In 1971, Leon Chua proposed a resistor with memory, or a “memristor,” as a theoretical fourth fundamental circuit element.^8,9 Although this original theoretical definition involves magnetic induction,¹⁰ the term “memristor” is commonly applied to any resistive switching memory.¹¹ Memristors are generally formed by a two-terminal device in an MIM structure. Under bias, they may be switched between a low resistance state (LRS) and a high resistance state (HRS).

While such resistive switching behavior was first reported in oxides in 1962,¹² the demonstration of a memristor based on TiO₂ in 2008 by Hewlett-Packard Labs sparked interest in resistive random access memory (RRAM) and the possible use of memristors to simulate synaptic functions,¹³ resulting in a significant surge in number of patents published for technologies referring to memristive function in the early 2000s [Fig. 1(a)]. Non-volatile memories (NVMe) are emerging as a feasible alternative to bridge the performance gap between traditional memory technologies like DRAM and flash. NVMes include memory classes like resistive, ferroelectric (FE), and 2D materials, each offering unique advantages.

The second class of NVMes popular today is based on ferroelectrics. Ferroelectricity, the inherent electric polarization of certain minerals, was first identified in Rochelle salt by Valasek in 1920.¹⁴ Memory devices based on ferroelectric (FE) materials were first proposed in 1952.¹⁵ Ferroelectric materials exhibit a switchable electric polarization, which can be utilized as a parameter to save information.¹⁶ In 1963,¹⁷ a novel solid-state memory resistor was introduced, designed for variable gain components, leveraging ferroelectric material memory capabilities. It utilized remanent polarization to control semiconductor field-effect conductance. The ability of ferroelectric materials to retain a polarization state even in the absence of an externally applied field¹⁸ means they can be used for non-volatile static memory functionality, which was studied throughout the 1970s–1980s using materials including ABO₃ perovskites such as BaTiO₃ (BTO) and lead zirconate-titanate Pb(Zr, Ti)O₃ (PZT).^19–21 A series of milestone developments in ferroelectric technologies in the early 2010s [Fig. 1(b)], including the discovery of ferroelectricity in doped HfO₂ in 2011,²² resulted in increased interest by the NVMe community [Fig. 1(a)]. While Pb(Zr_1–xTi_x)O₃ and BaTiO₃, as mentioned, are limited in scaling down film thickness,^23,24 the HfO₂-based devices show excellent ferroelectric properties at sub-10 nm thicknesses.^25–31 Moreover, HfO₂ offers a distinct advantage over PZT due to its lead-free composition, mitigating environmental concerns associated with the toxicity of lead-containing materials and compatibility with CMOS technology processing.⁵ 

Third, the discovery of two-dimensional (2D) or Van der Waals (VdW) materials in the early 2000s is of great significance for achieving high-density integration of devices. VdW materials have a layered structure with covalent in-plane bonding but weaker Van der Waals-type interplanar bonds. This strong in-plane bonding gives mechanical strength while maintaining flexibility, and the weak interplanar bonds mean that layers can be cleaved and separated. Monolayer graphite, now commonly referred to as graphene, was studied both theoretically and experimentally. Early theoretical studies date back to the 1940s³² and experimental observations, including via transmission electron microscopy (TEM), were reported decades prior to the mechanical exfoliation technique introduced by Novoselov et al.³³ The implementation of 2D materials into electronics was rapidly investigated following the discovery of graphene in 2004,³³ as is evident in Fig. 1(a), with multiple memristors utilizing graphene oxide as the active material demonstrated by 2010.^34–37 Closely following this, various electronic devices were constructed incorporating single or few layers of semiconductive MoS₂, a VdW transition metal dichalcogenide (TMD).^38–42 VdW interactions between 2D materials like graphene, hexagonal boron nitride (hBN), black phosphorus (BP), and 2D TMDs do not require lattice matching, making heterostructures easy to form and regulate switching characteristics via interface control.⁴³ VdW heterostructures have been used to fabricate devices with optical transparency and ultralow switching powers,^44–47 while materials including HfSe₂ and MoS₂ have been successfully integrated into crossbar arrays.^48–50 In addition, with demonstrations of FE in SnTe⁵¹ and CuInP₂S₆.⁵² Among others,^52–59 2D FE VdW-based NVMe offers novel functionalities such as multi-level storage, synaptic-like behavior, and neuromorphic computing applications.

This review explores emerging non-volatile memories as functional devices and the concomitant challenges that are now being investigated in recent literature. The obstacles, including unpredictability in operating parameters, insufficient comprehension of switching processes, the influence of vacancies and other defects, inadequate interface management between the active layer and functional electrode, and many intrinsic physical limits, impede the progress of non-volatile memories. This study examines both established and novel characterization approaches that may tackle these problems and promote the advancement of NVMe technology. Section II provides an overview of the basic NVM devices elaborating on their baselines working mechanisms and the major challenges impeding large scale commercialization back. Section III illustrates how these challenges are investigated and addressed using established techniques like electron, probe and optical microscopy techniques. Section IV introduces the novel approaches for atomistic, operando measurements in ambient environment, offering insights into the challenges and potential solutions toward commercialization of non-volatile memory systems.

Oxide-based NVMes are very promising materials in the emerging NVMe applications, yet, many obstacles hinder large-scale commercialization.^6,25,30 Addressing these roadblocks requires insight into fundamental device dynamics. For commercial relevance, a high degree of consistency from device to device is required, as well as reliable switching over time.⁶⁰ Integration and miniaturization are also key challenges that must be addressed for widespread implementation. Technical challenges include scaling down active layers while maintaining correct functionality, non-uniformity, poorly understood switching mechanisms, 3D monolithic integration, endurance, and reliability.⁶¹ 

Resistive switching memories, a desirable non-volatile memory technology, depend on the regulated formation and dissolution of conductive filaments (CFs) inside a dielectric layer. The creation and elimination of an atomic conductive filament (CF) inside a device underlies the switching process of resistive memory. In a Ni/HfO₂/SiO_X/Si memristor, Fig. 2(a) schematically illustrates the evolution of the metal nanofilament in RRAM devices across different operational stages. It depicts the transition from the SET state to the RESET state. The current flows from the Ni electrode to the n⁺-doped Si substrate. The resistive switching mechanism is governed by ion migration within the dielectric layers, leading to the formation and rupture of conductive filaments. During the SET process, voltage application induces the formation and migration of oxygen vacancies, establishing a conductive path. Simultaneously, Ni atoms from the top electrode diffuse downward, forming an MNF with a truncated cone morphology spanning the SiO_x and HfO₂ layers. In the RESET stage, a high current through the MNF generates Joule heating, promoting thermally assisted Ni diffusion and localized melting at narrow regions of the filament. This leads to an abrupt filament rupture, switching the device to its OFF state. After RESET, residual Ni remains in the silicon as a compound, while disconnected Ni fragments persist in the dielectric layers, potentially serving as nucleation sites for subsequent metal nanofilament formation in the next SET cycle.⁶² 

Parameters used to describe the quality of resistive switching devices include switching speed, switching energy cost, on/off ratios, retention, and endurance. Conventional memristors generally use bulk oxides as their active layer. Table I summarizes the performance metrics for some exemplar memristive devices using oxides. A detailed comparison of Table I clearly reveals significant variations in performance metrics when using the same material with different electrodes, as well as when using different materials with the same electrodes.

Resistive memories, although a promising candidate for in-memory computing and energy-efficient neuromorphic devices to meet the computational requirements of future applications, are held back by high variability: poor cycle-to-cycle and device-to-device uniformities, eventually limiting their mass production.⁶¹ The difficulties of reliability and variability in RRAM (evident in Table I) arise from the stochastic formation of conductive filaments, which might differ in shape and position, thereby affecting the device's resistance state. Variability in device performance especially in endurance and memory windows among the most researched materials—TiO_x,^80,81 NiO_x,⁸² ZnO_x,^83,84 HfO_x,^25,81,85 WO_x,^86,87 and ZrO_x^88,89—produces a bigger challenge in understanding these multi-functional materials. Parameter variability in set and reset values spread over a voltage range [Fig. 2(b)] causes stochastic switching dynamics, which leads to erratic device performance in resistive memory devices. Desired values for ideal memory of >10¹² cycles are needed (although limited, some studies have reported Pt/Ta2O5−x/Pt memristors exhibiting endurance on the order of 10¹² cycles⁹⁰) to produce a technologically relevant product.⁹¹ Here, endurance failure often results from the formation of high-resistance interfacial layers or the enlargement of ion migration zones, which may further impair the switching mechanism and intensify device inconsistency.⁹² In addition, the endurance-retention dilemma presents a prevailing challenge. The current-retention conundrum is a significant factor for low-power applications.⁹³ A broad filament can attain substantial retention at the expense of power, while a thin filament can accomplish minimal power usage at the expense of retention. The current retention difficulty is a significant issue that requires deeper examination. In situ examination of filamentary and interfacial switching processes is still being investigated. However, owing to the smaller dimensions of the region of interest in filamentary processes, the investigation is hampered by physical and phenomenological challenges.

The optimization of the device metrics discussed above requires knowledge of the fundamental switching mechanism, a detailed investigation of device evolution during cycling, and a sufficient understanding of the material structure to identify failure modes. The conductive mechanism in oxide-materials-based resistive switching devices is complex. The switching behavior often has been reported to involve interactions among defects,⁹⁴ electrons,⁹⁵ metal ions,⁹⁶ interfaces,⁹⁷ and even grain boundaries.⁹⁸ In the published literature, it is challenging to examine an atomic device and fully elucidate the physical mechanisms of resistance switching. The unregulated movement of ions inside the memristor is a critical component contributing to its performance deficiencies [Fig. 2(b)]. In contrast to conventional MOS transistors, which depend on the movement of a large number of electrons, memristors depend on the migration of a restricted number of ions via a designated layer or interface. The restricted number of ions, along with the stochastic characteristics of their movement and the morphology of the resultant ion migration region, complicates the precise regulation and quantification of the memristor's performance, resulting in randomness and inconsistency.⁹² Sophisticated characterization techniques capable of elucidating switching pathways are needed to understand the device kinetics at the atomic scale. This is imperative toward the development of ultrathin switching memristor devices.

The discovery of graphene in 2004 paved the way for explosive growth in the field of 2D materials.³³ Advantageous characteristics of 2D materials include strong in-plane covalent bonding, which provides mechanical strength while maintaining flexibility, fast switching speeds, low power consumption, and excellent device scalability.^99–107 Using monolayers enables the formation of devices that are at the physical limit of minimum thickness, maximizing the potential for high-density integration. For example, where conductive filaments one atom wide are formed in a monolayer-based memristor [Fig. 2(c)], the term “atomristor” is sometimes used. A significant benefit is the potential for an atomically flat electrode contact, enabled by the absence of dangling bonds on the surface of a van der Waals material. The use of 2D materials facilitates the circumvention of scalability challenges associated with reducing bulk oxides to similar thicknesses. Nonetheless, achieving satisfactory levels of memory and endurance now constitutes an obstacle to scaling up.

Among various 2D materials proposed over the years, graphene is a mechanically strong yet flexible VdW electrical conductor with the highest electrical conductivity of any known material at room temperature. It has an electron mobility of up to 2 × 10⁵ cm²/Vs,¹⁰⁸ but this is strongly affected by the fabrication process.¹⁰⁹ The monolayer shows optical transparency to white light of 97.7%.¹¹⁰ These properties make graphene highly appropriate for transparent conductive electrodes, and competitive with the standard oxide electrode indium tin oxide (ITO). The ideal graphene lattice is a zero-band gap semimetal,¹¹¹ with modification of the electronic band structure achieved by the introduction of point defects, wrinkles and crumples, dislocations, or grain boundaries.¹¹² These many degrees of freedom for controlling the electronic structure of graphene offer flexibility to optimize the material properties according to the device application.^113,114 Graphene is therefore commonly used as an electrode material in 2D materials-based memristive devices and is also combined with a secondary material as a composite to form the active switching layer.^115–123 In the years immediately following the discovery of graphene, graphene oxide was explored as a 2D insulating dielectric for use in NVMe devices.^34–37 Subsequently, hBN has emerged as the preeminent VdW material used as an insulator and is used to perform a range of functions in 2D electronics,^124,125 including as a switching layer for ReRAM with ultralow energy consumption.^{118–120,126–139} However, the dielectric constant of hBN is relatively low, and fabrication challenges lead to the presence of defect states in the bandgap, which lead to undesirable tunnel currents in few-layer hBN,^140,141 so other 2D insulators are sought after. Candidate materials include bismuth oxyhalides BiOX (X = Br, I),^142–144 Bi₂SeO₅,¹⁴⁵ and quasi-2D ultrathin crystalline CaF₂.¹⁴⁶ 

Transition metal dichalcogenides (TMDs) are a class of semiconducting VdW materials with a trilayer structure utilized as transistor channels or switching layers in 2D NVMe devices.¹⁴⁷ Various semiconductive VdW materials, including Group IV Monochalcogenides (SnS¹⁴⁸ and GeSe¹⁴⁹), 2D III-V semiconductors (GaSe and InSe^150–152), ultrathin metal oxides (MoO₃ and TiO_x,^153–155), black phosphorus,^156–158 and organic–inorganic quasi-2D halide perovskites.^159–161 2D TMDs (NbTe₄^162,163 ReS₂,^164–166 SnS₂,¹⁶⁷ MoTe₂,^168–171 and MoS₂^{116,121,132,172–183}) currently dominate discussion of industry-relevant 2D semiconductors.^184,185 MoS₂ is highly suitable for use as the active layer in memristors, due to high carrier mobility, well-established fabrication methods, robust Young’s modulus, direct bandgap, and chemical stability under ambient conditions.^186–188 The construction of NVMe based on ferroelectricity in 2D materials has been an object of recent interest.^189–191 Ferroelectric polarization does not normally persist for ultrathin oxides, although some successful preparations of ultrathin films of bulk oxides with preservation of ferroelectric behavior have been reported.^192–194 Ferroelectric behavior in VdW materials has been reported for In₂Se₃,^53–54,195 CuInP₂S₆,^52,58,59 CuCrSe₂,⁵⁷ and γ-GeSe.⁵⁶ In-plane ferroelectric polarization has also been achieved in a single layer of SnTe⁵¹ and in few layers of SnS at room temperature.¹⁹⁶  Table II summarizes the performance metrics for some exemplar 2D memristive devices.

Low-power memristors are highly sought after for their potential in fast-speed, energy-efficient artificial neuromorphic networks. These devices are expected to consume less power than conventional CMOS devices due to their non-volatile nature and the absence of standby power dissipation.^107,202 Unlike CMOS devices, which continuously dissipate power due to leakage currents, memristors only consume power during the switching process. Accurate power and energy calculations for memristors, however, require detailed knowledge of their I–V characteristics during switching, including parameters like threshold voltage, on-state current, and switching speed. Unfortunately, such data are not consistently available across all the memristors reported in the literature, thus not included in Table II.

Heterostructures in 2D memristor research improve switching behaviors, power consumption, and on/off ratios by customized interfaces and bandgap engineering.^{47,164,199,203} For example, it has been demonstrated that the addition of a hBN monolayer as a tunneling barrier reduces the electrode contact resistance for MoS₂-based devices.^204,205 The design of these heterostructures is highly flexible since the construction is based on VdW bonding, which eliminates the requirement for lattice matching.

There are a range of roadblocks to the widespread implementation of 2D materials in nano-electronic devices. A group of significant challenges is encountered in device fabrication. The growth and transfer of large-area, high-quality films are indeed challenges for all 2D material platforms, where fabrication compatible with current silicon chip technologies is particularly desirable.^105,206 Damage to the 2D material can occur at various stages of device manufacture, including common crystallographic damage to the active layer during the deposition of the top electrode, as illustrated in Fig. 2(d). Since any damage to a 2D material is likely to impact a large fraction of the material volume, issues such as damage upon the deposition of the top electrode²⁰⁷ must be carefully accounted for. Stable and consistent doping of 2D materials for controlling device polarity is another prominent challenge.^100,208,209 Another issue is incompatibility of 2D materials with currently widely available high-k dielectrics such as Al₂O₃ and HfO₂.^100,210,211 The properties of 2D materials can be significantly impacted by the substrate, roughness, charged impurities, or the presence of any contaminants. This frequently requires dielectric encapsulation of 2D materials for stability, adding a further layer of complexity to device construction. The VdW insulator hBN may be employed for this purpose, but its suitability is limited by relatively low dielectric constant and by the difficulty of scaling up fabrication methods.²¹² 

Given that 2D materials-based devices have thicknesses on the atomic scale, device performances are highly sensitive (evident from performance metrics listed in Table II) to sources of intrinsic and extrinsic disorder, such as crystallographic defects, interfacial Coulomb impurities, or fluctuations of the dielectric environment.^213–215 Temporal and spatial variability is generally high in 2D materials-based NVMe devices, particularly in vertical devices based on filamentary switching. The probabilistic nature of the formation and fracture of the filament means that the threshold voltages follow a stochastic distribution, which can have a particularly large width for 2D materials. Where the variability is very high, this property has been harnessed to mimic spiking action in neurons and for encryption purposes.^116,158 Figure 2(d) illustrates one such approach used in a study that employed electron irradiation to increase the number of sulfur vacancies in a monolayer of MoS₂, thereby improving both device yield and endurance.⁶⁴ While non-filamentary mechanisms have also been demonstrated, such as voltage-bias-induced motion of sulfur vacancies in ReS₂-based lateral memristor,¹⁶⁵ this issue of variability must be addressed for proper utilization of vertical filamentary 2D materials based memristive technologies. This can only be achieved with complete characterization of local structure, dynamics, and sources of disorder during operation, to identify device mechanisms and design for optimized performance accordingly.

Ferroelectric NVMes operate on the principle of the use of electric polarization orientation for encoding information. The thermodynamic properties of ferroelectrics can be explained by Landau theory,^216,217 which assumes that the ferroelectric phase (lower symmetric phase) transition occurs through a small distortion of the structure of the paraelectric phase (higher symmetric phase). Figure 2(e) depicts the influence of strain via the substrate and top electrode (capping effect) on the development of ferroelectricity in HfO₂. As the thickness of the deposited film diminishes, smaller grains develop post-annealing. This influences the energy landscape of the film, perhaps stabilizing the ferroelectric orthorhombic phase. Ultrathin films may benefit from elevated surface energy, which promotes higher-symmetry phases. The application of stress during manufacture may result in the development of ferroelectric phases. In contrast, thicker films (about 20 nm) are inclined to develop the non-polar monoclinic phase. The best thickness for achieving superior ferroelectric characteristics in HfO₂-based thin films is around 5–10 nm, striking a balance between surface energy advantages and the reduction of detrimental interfacial layer effects.²⁵ Ferroelectric memories offer a promising future for in-memory computing and energy-efficient miniature devices to meet computation-intensive processes. These devices have especially shown big promise in low power consumption and fast read/write speed.²¹⁸ Several fluorite- and perovskite-based oxide materials have been investigated, showing variability in device performance and reliability in terms of endurance (Table III).

There are, however, several challenges such as high coercive field, wake-up, and fatigue that are blocking the large-scale commercialization of these devices; some of those are discussed below.

Thin-film ferroelectricity observed in HfO₂ and ZrO₂ led to a paradigm shift from conventional perovskites (PZT, BTO) in realizing extremely scaled down NVMs.²⁸ A major impediment in realizing this has, however, been high coercive fields observed in the former, which are essential to achieve low voltage/low power consumption. The higher E_C value indicates the higher kinetic energy barrier between the two polarization states,²²³ which is reported to be the kinetic energy that can switch o-HfO₂ by displacing oxygen ions along the polar axis.²⁵ The study of the possible intermediate phases was also conducted by Maeda et al.,²²⁴ who established that there are two intermediate states in the spontaneous switching pathway, o-phase (Pbcm) and t-phase (P42/nmc), with the lowest energy pathways seeming the most plausible. The energy barrier for the switching pathway involving the o-phase (Pbcm) at zero polarization was about twice as high (200 meV/f.u.) compared to that of the t-phase (P42/nmc) at zero polarization (75 meV/f.u.). Consequently, the pathway involving the t-phase (P42/nmc) is considered the most viable route for polarization switching in polar o-HfO₂.^225–227 A direct experimental in-operando verification of these events is limited by phenomenological and instrumental limitations. A verification of phase transformation at several points in the polarization switching pathway will go a long way in having a deeper understanding of the coercive field problems.

While the reduction in the coercive field remains a challenge, an increase or decrease in polarization charge over several thousand read/write cycles introduces unreliability. The gradual increase in P_r upon the application of a cycling electric field is referred to as the wake-up effect.²²⁸ This is a primary difficulty hindering the large-scale commercializing of HfO₂-based ferroelectric devices. In published literature, two processes are identified as probable reasons behind the wake-up effect; the electric field-induced non-polar to polar phase transition and domain wall pinning and depinning influenced by point defects, particularly oxygen vacancies.²²⁹ The STEM analysis [Fig. 2(f)] of the HfO₂ film indicated a phase transition from a non-polar t- or m-phase to a polar o-phase with the application of an electric field.²³⁰ The transition, corroborated by high-angle annular dark-field (HAADF)-STEM analysis in Fig. 2(f), accounts for the observed rise in Pr. This phase change may be a direct consequence of applied bias or may result from the redistribution of oxygen vacancies under the influence of a strong electric field.²³⁰ The second prominent reason is anticipated to be point defects, particularly oxygen vacancies.^30,231,232 A direct investigation of these phenomena requires direct mapping of atomistic processes inside the ultrathin functional layer under applied bias and essentially tracking under ambient conditions to isolate stray effects.

Like wake-up, another significant issue limiting the commercialization of fluorite-based oxide NVMes is the fatigue phenomenon. The remnant polarization in ferroelectric thin films diminishes after ∼10⁵ cycles [Fig. 2(f)], hence progressively narrowing the memory window and leading to device failure. Various mechanisms are hypothesized to produce fatigue in different types of ferroelectrics during electric field cycling.^233,234 The growth of a non-polar and low permittivity (ɛ_r) layer between the functional electrode and the FE thin film has been reported in several studies.²³⁴ A different process proposed relies on the buildup of oxygen vacancies.^235,236 Oxygen vacancies were reported to move and accumulate at the film/electrode interfaces and film grain boundaries during field cycling, resulting in domain-wall pinning. The buildup of oxygen vacancies at the grain boundaries, which restrict domain walls and reduce polarization switching in the presence of an applied electric field, can lead to a gradual decline in the P_r value.^235,236 Consequently, the development of large grains throughout the fabrication process could potentially elevate the number of cycles until fatigue occurs.^237,238

We have seen how challenges, including high variability, damage during fabrication, and sensitivity to disorder in 2D materials; high coercive field, wake-up, and fatigue in ferroelectrics; the challenge of balancing retention and power requirements in filamentary devices; and poor spatial and temporal uniformity, as well as low endurance, are all the result of a lack of understanding of materials and atomic dynamics. Therefore, characterization methods become essential tools for the development of new NVM technologies as they unravel intrinsic material characteristics, device functionalities, and failure causes. Advanced microscopy, spectroscopy, and in situ electrical measurements may offer comprehensive insights into atomic-scale phenomena, with the aim to optimize material interfaces, develop dependable switching mechanisms, and enhance device reliability, thus expediting the incorporation of novel memory solutions into practical applications.

Electron microscopy (EM) is often utilized to analyze the surface and internal structure of devices at the level of atomic structure. EM techniques can reach sub-nm resolution and yield multidimensional datasets, including tomography, diffraction patterns, and elemental analysis, but also come with a set of stringent limitations.²³⁹ Scanning Electron Microscopy (SEM) collects information about the topography and composition, with a spatial resolution in the range of 0.5–5 nm.²⁴⁰ SEM on Fe-doped STO shows structural changes in an Au/Fe:STO/Nb:STO memristive device²⁴¹ revealing the appearance of a feature marked as the “forming crater,” formed following a current spike during [Fig. 3(a)]. Higher magnification of the contrast spot and subsequent XPS points to Sr enrichment and Ti³⁺ and Fe²⁺ ions in the developing crater; crucial to generating conductive filaments. In a similar study, filament formation dynamics have been observed for electrochemical metallization (ECM) memory Pt/H₂O/Ag, with the cell in a lateral geometry to facilitate the use of SEM [Fig. 3(a)].²⁴² The conclusions drawn about the interaction of the morphology of the Ag dendrites with the switching characteristics of the system are limited by the discrepancy between the vacuum environment of the SEM and the ambient conditions under which the devices are generally operated, as well as the requirement for lateral geometry rather than the usual vertical three-dimensional structure of memristors.

Preparing samples for transmission electron microscopy (TEM) requires resource-intensive methods to achieve the required sample thickness and cleanliness, meaning TEM has high costs in terms of both time and money (Fig. 3). In an ex situ study on FE HfO₂,²³⁴ TEM was utilized to analyze Gd:HfO₂ film structural changes at electrode interfaces during pristine, wake-up, and fatigue cycling conditions. The bulk monoclinic and tetragonal (both paraelectric) symmetry relaxed at the electrode interfaces toward the orthorhombic (ferroelectric) phase as per TEM results [Fig. 3(b)]. This relaxation trend was most pronounced in the pristine state and least pronounced in the fatigued state. This shows that the wake-up effect is caused by interface phase transformations from non-ferroelectric to ferroelectric, increasing switchable material volume. With sustained cycling, TEM observations of interface relaxation decreasing coincide with fatigue, suggesting a relationship between switchable areas at interfaces and leftover polarization. In memristive devices, TEM revealed filament dynamics in ZnO-based resistive memory. Researchers employed in situ TEM to study conducting filament development, confirming their conical shape and elemental Zn composition.²⁴³ As the applied voltage increases, the CF, seen by TEM, extends across the ZnO layer, finally establishing a connection between the top and bottom electrodes [Fig. 3(c)]. The study clearly demonstrated that the SET process initiates from a residual filament left behind from the previous reset process. TEM also showed filament breakdown at the inert electrode contact during erasing. The capacity to directly observe filaments and their behavior at various phases of operation revealed filament development dynamics. It has been highlighted that observations made under the vacuum conditions of EM may differ from the characteristics under the devices’ normal operating conditions, with the presence of atmosphere having strong influence on the electrochemical behavior of ferroelectric and other devices.^244,245 For example, RESET failure in HfO2-based CBRAM was found to occur under vacuum and was attributed to a dual CF mechanism associated with moisture desorption, an effect not present in atmosphere.²⁴⁶ 

Scanning Tunneling Microscopy (STM) provides exceptional spatial resolution and is utilized for both visualization and modification of surfaces at the atomic level. It may be considered a less complex method than EM, with less time required per device measurement but still significant financial investment required (Fig. 3). STM characterization of MoS₂ nanoclusters [Fig. 3(d)] helps comprehend their atomic-scale electrical characteristics.²⁴⁷ Specifically, STM imaging and spectroscopy show one-dimensional metallic edge states in MoS₂ nanoclusters. The conductive states at the margins of MoS₂ nanoclusters are caused by electronic structural changes, not dangling bonds. Localized edge states provide bright features at the edges of MoS₂ clusters in STM pictures. The Fermi level is pinned 1.2 eV above the valence band maximum in these states. These metallic edge states may increase material conductivity, particularly in nanoscale devices with a high edge-to-bulk ratio. When a 2D material approaches the nanoribbon limit, edge states as observed via STM are projected to dominate electron conductance. However, STM is limited to the study of the top surface of electrically conducting materials and is generally performed under high-vacuum and low-temperature conditions, limiting its applicability to the study of material dynamics within NVMe devices.

Conductive Atomic Force Microscopy (C-AFM) blends high-resolution topographical imaging with local electrical characterization, with intermediate cost and time requirements. It detects current flow through a material under an applied voltage using a conductive AFM tip to explore nanoscale electrical characteristics.²⁴⁸ This method helps study local conductivity, resistance, and charge transport processes in semiconductor devices, thin films, and organic electronics. C-AFM on MoS₂ measures the uniformity of the Schottky barrier height (SBH) at metal–MoS₂ device interfaces. SBH determines the sort of charge carriers (electrons for n-type, holes for p-type) that may enter TMD-based transistor channels. Figure 3(e) illustrates how oxygen plasma treatment adjusts the SBH of MoS₂, enabling low SBH areas for electrons and holes.²⁴⁹ This discovery enables the creation of ambipolar MoS₂ transistors and 2D memristors, a vital step in creating complementary MOS technology. While silicon-based CMOS technology requires both n- and p-type silicon, complementary MOS technology requires both types of MoS₂.

X-ray techniques are also employed to provide information about surface composition and chemical states. The application of bias during XPS experiments allows for in situ study of how electron binding energies vary during cycling.²⁵⁰ On the other hand, Fig. 3(g) illustrates the XPS spectra collected ex situ to investigate the resistive switching mechanism in a p-Si/SiO_x/Sn device.²⁵¹ Spectra were collected from the active layer after forming and after reset processes, by etching away the top electrode. A comparison of the Sn⁰ peaks shows a significant increase after reset as compared to after forming, indicating a reduction of Sn ions, which the authors note supports an oxygen vacancy-supported mechanism of resistive switching. This study was limited to post-mortem analysis due to the requirement to strip the top electrode for access to the active layer. X-ray absorption spectroscopy (XAS) is a resource-heavy technique that generally uses synchrotron radiation to study local atomic and electronic structure. In situ XAS allows for a detailed study of compositional changes over the lifetime of a memristive device and has been used to identify ring-like patterns of oxygen migration as a contributing factor to device failure in tantalum oxide memristors.²⁵² 

While other methods may be intrusive or destructive, Raman spectroscopy shines for being generally a non-invasive analytical method. As a tabletop method performed under ambient conditions with no special sample preparation requirements, Raman is low-complexity and low-cost while still yielding rich real-time data. This unique position relative to other operando characterization methods is illustrated in Fig. 3. Measuring the inelastic scattering of monochromatic light, usually from a laser, yields comprehensive information on molecular structure, chemical composition, and interactions. Figure 3(f) tackles the problem of establishing Grüneisen characteristics by studying strain-phonon frequency relationships for 2D materials like MoS₂.²⁵³ MoS₂ bending data are shown in Fig. 3(f) employing a flexible cruciform substrate. Researchers can correctly assess strain-induced vibrational property changes by comparing Raman shifts with applied strain.

Although the techniques discussed above are all beneficial, none of them can be considered a universal solution. In fact, there is a significant demand to be able to characterize and observe two-dimensional materials and thin films in their natural/ambient environment. In addition, there is a need to concurrently evaluate the topography, electrical characteristics, and spectroscopic behavior of these materials.

NVMe devices are generally based on interactions between defects,⁹⁴ electrons,⁹⁵ metal ions,⁹⁶ interfaces,⁹⁷ and grain boundaries.⁹⁸ Efficient characterization of NVMe devices should examine the correlation of these dynamic interactions with electrical characteristics in real time and under normal operating conditions. The rapid device development and iteration demanded by the current market pace means that characterization must also be fast, scalable, repeatable, and high-throughput. Measurement techniques based on electron microscopy or those requiring the use of a synchrotron x-ray source cannot be considered to meet these requirements. Scanning probe techniques face challenges including probe-sample interactions, tip reproducibility, and electrode durability. Electronic characterization approaches are often the principal research focus; however, although they provide extensive information on device performance characteristics, they often fail to provide crucial insights into the nano-kinetics inside the functional material, which are typically obscured by electrical measurements alone. Therefore, novel operando microscopy methods are emerging as highly effective approaches for rapidly developing an understanding of functional materials. Non-perturbative methods based on optical and plasmonic phenomena offer a unique route to understanding device dynamics.^254,255

Photoluminescent (PL) emissions are widely used to characterize semiconductors, yielding information about the electronic structure, including insight into any defect energy levels. Measurement of real-time variation in these optical signals offers a powerful in situ approach to real-time tracking of material modifications. Operando Raman spectroscopy can be used to detect real-time symmetry changes in the active medium of a device during electrical switching. For example, an operando Raman study of a thin-film niobium oxide has demonstrated that sodium (de) intercalation drives interfacial resistive switching, a factor relevant to device lifetime.³ Operando Raman and PL have been used to evidence activation of sodium cations in Na-doped WO₃ resistive switching devices.²⁵⁶ 

Plasmon (the resonant oscillation of electrons within a material when illuminated by light at the resonance frequency)²⁵⁷ phenomena can boost the capability to read out optical signals and have been used to enhance signal detection due to high field confinement within active devices, as well as to gather information about the materials where they are confined, as the plasmonic resonant frequency strongly depends on the materials composition and device morphology.^254,255,258 A key advantage of the plasmonic enhancement of optical emissions is the ability to collect information from a nano-volume of material. The construction of resistive switches in a “nanoparticle-on-mirror” (NPoM) geometry has been shown to generate a tightly confined plasmonic hotspot whose volume overlaps with the switching material. This generates an optical system highly sensitive to local changes within the switching material, resulting in a non-destructive tool for rapid device characterization under ambient conditions. Morphological changes due to electrical cycling of a nanoscale electrochemical metallization (ECM)-based switch in NPoM geometry have been detected via the modulation of light scattered by the nano junction. The formation of CFs caused a change in the overall geometry of the optical nano cavity (coinciding with the switching gap), hence modifying the plasmonic resonance condition.²⁵⁹ The same approach has also been used in valence change memories (VCM), where the change f refractive index due to accumulation of oxygen vacancies in strontium titanate film has also been detected via frequency shift of plasmonic resonances.²⁶⁰ 

Plasmonic junctions have also been leveraged to combine plasmon-enhanced dark field (DF) nano-spectroscopy, nano-Raman, and nano-PL to examine the volatile switching of a device based on MoS₂.¹⁷⁴ With applied voltage, the Au PL increase [Fig. 4(a)] suggests metal intercalation into MoS₂, while the sulfur vacancy-associated PL peak at 740 nm shows no change. Contrary to earlier theories, this clearly shows that sulfur vacancies do not participate in switching and that Au atoms are intercalating between Mo and S atoms, without modifying the MoS₂ structure. The change in plasmonic gap mode intensity [Fig. 4(a)] has been able to prove [supported by finite-difference time-domain (FDTD) simulations] the creation of many Au nanofilaments rather than a single one. The possibility to optically “count” the number of single Au nanofilaments per cycle event unveiled the key role of citrate ligands present at the interface of the top electrode of the MoS₂ device switch. In fact, ligands had a “semi-memristive” action that slows Au atom movement and prevents Au from retracting back to the Au electrode, this way contributing to device nonvolatility.²⁶¹ This is a clear example on how operando optical tracking of device switching contributed to give pivotal insights on the switching mechanism, opening surface engineering as a new route toward MoS₂-based switches.

The use of a combination of operando DF, operando PL, and operando Raman has been illustrated by a range of studies.³¹ In fact, excitation of PL-enhanced emissions allows interrogation of the defect states in the active material within the hotspot volume. As a result, the local defect dynamics during electrical switching can be examined. For example, a study of AuNP/HZO/TiN devices connected “wake-up” and “fatigue” effects to oxygen vacancy (Vo) migration.³⁰ Highly localized crystallographic data collected by surface-enhanced Raman spectroscopy (SERS) indicated a reduction in the monoclinic phase and a rise in the orthorhombic/tetragonal phases [Fig. 4(b)]. Interrogation of defect states by photoluminescence (PL) data showed a rise in oxygen vacancies in the film during a “pre-wake-up” stage, but a return to a defect-free film level with continued cycling, implying that the Vo buildup became more uniformly diffused. Hence, the operando optical data indicated that a crystallographic phase transition caused by migration of Vo was the source of the “wake-up” in these devices.

Clearly, plasmonic nanocavities have been evidenced as a powerful investigative technique for the in situ study of device dynamics. To make this approach technologically relevant, statistics should be gathered from multiple such devices for thorough characterization of the material. This can be time consuming. This bottleneck has been circumvented by attaching a plasmonically active top electrode to a portable conductive cantilever, free to be moved across the material of interest to apply bias at any desired location. This opto-electrical technique is termed Scanning Plasmon-Enhanced Microscopy (SPEM) and has been demonstrated on 2D-TMDs MoS₂ and WSe₂.²⁶² SPEM analyzes nanomaterials simultaneously using topographical, electrical, and spectroscopic methods, unlike previous scanning probe methods. Gold nanoparticles' van der Waals interaction permits precise electrical characterization without harming the tested materials. Electrical and optical nanoscale mapping of MoS₂ (conductivity, dark-field, Raman, and photoluminescence) had been performed simultaneously in a study where MoS₂'s optoelectronic capabilities were observed to change with thickness²⁶² [Fig. 4(c)]. SPEM's plasmonic hotspot boosted Raman and PL signals, enabling higher-resolution mapping than laser mapping. SPEM enhanced spectral resolution from 2.2 µm (without plasmonic enhancement) to 600 nm. This improved resolution reveals small structural differences that affect MoS₂'s optical and electrical characteristics. Dark-field measurements at several points on the flakes were utilized to analyze MoS₂'s local optical characteristics using SPEM. In the field of NVMe, the aim is to have a facile method to understand how topographical or crystallographic features such as defects, domains, or sample edges may influence the local switching characteristics of the material. This is facilitated by SPEM’s ability to simultaneously collect electrical and optical information that is both spatially and temporally resolved.

Modulation of plasmonic resonances can be used as a mechanism for optical readout of memristor switching. For example, state switching of an RRAM integrated with a plasmonic waveguide has been detected via the disruption of SPP propagation by the formation of the metallic conductive filament [Fig. 4(d)].²⁶³ In another interesting work,²⁶⁴ an electrically controlled plasmonic switch functioning at the atomic scale was developed. The switch enables rapid and repeatable transitions by the repositioning of a single atom or, at most, a few atoms inside a plasmonic cavity. This study discusses the development of an atomic-scale plasmonic switch exhibiting step-like switching capability, as seen in Fig. 4(e). A fundamental issue in comprehending memristors resides in monitoring conductance quantization and conductance fingerprints, which reveal functioning at the atomic level. This work employs a method that fabricates an integrated atomic-scale optical switch on a silicon-on-insulator wafer. The researchers developed almost planarized waveguides using a modified local oxidation of silicon (LOCOS) technique. They then constructed metallic pads of silver and platinum, filling the interstice between the two metals with amorphous silicon (a-Si). An atomic-scale filament structure was ultimately created by applying a voltage, which facilitated the movement of silver ions to form a filamentary channel. This filament, serving as a plasmonic cavity, is modulated by moving a single or a few atoms, resulting in either a conductive or non-conductive condition that varies the plasmonic cavity characteristics, leading to unique plasmonic resonances. This procedure allows the observation of conductance quantization, signifying the effective functioning of a memristor at the atomic scale. Furthermore, where the atomic filament is formed due to light-induced thermal changes, the memristive switching can be used as a mechanism of photodetection.²⁶⁵ 

Novel techniques that have been employed for the in situ study of other devices such as batteries may also have potential applications in the study of NVMes. To tackle the difficulty of comprehending dynamic lithium-ion intercalation processes in battery materials, which constrains advancements in battery technology, particularly rapid charging, researchers used interferometric scattering microscopy (iSCAT). Figure 4(f) illustrates the application of this technique for tracing lithium-ion heterogeneity within a nickel-rich manganese cobalt oxide (NMC) cathode. This laboratory-based, cost-effective method has been used to trace lithium-ion heterogeneity within a nickel-rich manganese cobalt oxide (NMC) cathode,²⁶⁶ with iSCAT intensity increasing (or decreasing) during de-lithiation or (lithiation). Lithium distribution during charging was closely examined [Fig. 3(f)] and found to follow a “core–shell” pattern, with de-lithiation occurring at the particle periphery first. When the particle was allowed to rest after charging, a more uniform distribution was obtained, as shown in the intensity maps of the NMC particle in Fig. 3(f). These observations were found to be consistent with kinetically-driven intra-particle heterogeneity during charging and allowed the authors to estimate the lithium-ion diffusion coefficient for the cathode material. iSCAT has also been utilized to track lithium-ion dynamics in LixCoO₂, the prototypical cathode material.²⁶⁷ The same intensity change upon lithiation was used to map the presence of propagating phase boundaries during the biphasic and lithium ordering transitions in LixCoO₂. These results underscore the capability of iSCAT microscopy to provide insights at the nm-length scale into lithium-ion intercalation processes in battery materials under ambient/natural settings, with potential for extension to the study of ionic motion in memristive systems.^268,269 The real-time tracing of the dynamics of ions has also been achieved by detecting perturbations of surface plasmon waves hosted by an inert plasmon optical fiber probe.²⁷⁰ 

Variability, reliability, and scalability continue to pose substantial difficulties in the commercialization of non-volatile memories (NVMes) like RRAM and FeRAM, which are considered among the most promising next-generation memory technologies. Two-dimensional materials offer a flexible platform for designing high-density, low-power RRAM, and FeRAM devices, but challenges arising from the lack of understanding of switching mechanisms are currently hindering progress. This study emphasizes the transformative capacity of advanced characterization techniques, such as plasmonic spectroscopy, nanoparticle-on-mirror (NPoM) structured nanoelectrodes for plasmon-enhanced mapping, interferometric scattering detection for ion migration, and state-of-the-art atomic-scale optoelectronic switches in tackling these challenges. Through an exhaustive examination of how these techniques elucidate the nanoscale intricacies of NVMs, we provide an innovative viewpoint on expediting the advancement and use of these essential technologies. The integration of traditional and novel characterization techniques can provide practical insights into material behavior with competitive precision, illuminating pathways to device optimization. These contributions provide a basis for forthcoming advancements in NVM technology and their extensive commercialization.

G.D.M. acknowledges support from the EPSRC (Grant Nos. EP/R511675/1 and EP/X034593/1; selected by the ERC, funded by UKRI) and the Royal Society (Grant No. RGS/R1/221262). D.M.K. acknowledges support from the EPSRC (Grant No. EP/S022953/1) and the Cambridge Trust. All authors acknowledge support from the Winton Program for the Physics of Sustainability.

The authors have no conflicts to disclose.

A.J. and D.M.K. contributed equally to this work.

Atif Jan: Investigation (equal); Visualization (lead); Writing – original draft (lead); Writing – review & editing (equal). Dawn M. Kelly: Investigation (equal); Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). Giuliana Di Martino: Conceptualization (lead); Funding acquisition (lead); Project administration (lead); Resources (lead); Supervision (lead); Writing – review & editing (equal).

Data sharing is not applicable to this article as no new data were created or analyzed in this study.