WorksPhysics and Astronomy
PdNeuRAM: forming-free, multi-bit Pd/HfO2 ReRAM for energy-efficient neuromorphic computing
Abstract
Memristor technology shows great promise for energy-efficient computing, yet it grapples with challenges like resistance drift and inherent variability. For filamentary Resistive RAM (ReRAM), one of the most investigated types of memristive devices, the expensive electroforming step required to create conductive pathways results in increased power and area overheads and reduced endurance. In this study, we present novel HfO2-based forming-free ReRAM devices, PdNeuRAM, that operate at low voltages, support multi-bit functionality, and display reduced variability. Through a deep understanding and comprehensive material characterization, we discover the key process that allows this unique behavior: a Pd-O-Hf configuration that capitalizes on Pd innate affinity for integrating into HfO2. This structure actively facilitates charge redistribution at room temperature, effectively eliminating the need for electroforming. Moreover, the fabricated ReRAM device provides tunable resistance states for dense memory and reduces programming and reading energy by 43% and 73%, respectively, using spiking neural networks (SNN). This study reveals novel mechanistic insights and delineates a strategic roadmap for the realization of power-efficient and cost-effective ReRAM devices.
In this emerging era defined by Artificial General Intelligence (AGI) and Internet of Things (IoT) technologies [1], Computing-in-Memory (CIM), where computation and storage seamlessly converge in a single physical locale, has ascended as a visionary next-generation computing paradigm [2]. Among the many candidate technologies for computing cells, ReRAM has captivated both academia and industry with its compelling attributes, including non-volatility, zero leakage power, a simplified structure, high integration density, rapid switching speeds, and inherent CMOS compatibility [1, 3, 4]. However, beneath these luminous advantages lie formidable challenges, most notably, the prerequisite for electroforming pre-processing, where the electroforming phase necessitates the application of substantially elevated voltages to establish Conductive Filaments (CF) within the resistive medium [4]. It not only intensifies energy consumption and complicates fabrication but also compromises device reliability and endurance [5, 6, 7]. Moreover, the pronounced current density during ReRAM write operations further exacerbates concerns about power consumption and long-term durability, often compelling designers to enlarge ReRAM cell dimensions, a trade-off that diminishes circuit packing density and inflates manufacturing costs [5, 8].
To address these challenges, HfO has been widely explored as a premier dielectric candidate, celebrated for its superior breakdown voltage, robust thermal stability, and exceptional CMOS compatibility. These remarkable attributes promise not only a reduction in operational power consumption but also enhanced thermal performance over alternative dielectrics [9, 10, 12] [11]. Furthermore, the incorporation of Multi-Level Resistance (MLR) technology in HfO boosts memory density and slashes manufacturing costs [13, 14, 15], offering significant advantages for neuromorphic computing applications. Yet, current specialized fabrication methods, such as thermal annealing [16], X-ray irradiation [4], exotic element doping [16, 17], and plasma treatments [17, 18], are designed to introduce defect states and reduce forming voltages. Aforementioned techniques will increase manufacturing costs and complexity, typically demand substantial energy, and may inadvertently impair transistor performance [19].
Therefore, there is a compelling need for simpler, energy-efficient fabrication processes for forming-free ReRAM devices.
In this study, inspired by the remarkable efficiency of biological neural cells, as illustrated in Figure 1a, b, which underpin the acceleration of neuromorphic devices for fully connected artificial neural networks (Figure 1c) [1], we proposed a streamlined, forming-free ReRAM fabrication approach that leverages conventional Integrated Circuit (IC) manufacturing processes. Notably, the presence of Pd, serving as both an electron injection source and catalytic center in HfO, fosters a particular atomic configuration that diminishes charge diffusion barriers [20]. Synergistically, the cooperative action of these integrated stacks culminates in forming-free behavior. Electrical characterization reveals significantly lower initial resistivity in pristine PdHT devices compared to conventional Pt/HfO/Ti/Pt (PtHT) devices [21]. Moreover, the PdHT devices exhibit stable multilevel resistance states and low-power switching capabilities, achieved through modulation of RESET stopping voltages (), where the RESET refers to switching the memory device to its high-resistance state; the bias at the switching point is defined as ; and the sweeping stop voltage is defined as . These attributes render the proposed PdHT device a compelling candidate for energy-efficient neuromorphic computing applications. Experimental validations further confirm substantial energy savings during both training and inference phases of SNN, demonstrating effective performance in tasks such as image classification and gesture recognition.

Figure 1. Overview of neuromorphic systems and ReRAM device stacks. a An overview of potential applications for energy-efficient neuromorphic systems. b Schematics of a biological synapse. c Schematics of fully connected artificial neural networks. d Schematic illustration of crossbars. e Schematics of the designed ReRAM cell. f SEM image of fabricated crossbars. g The optical image of one cell (top) and STEM image of the cross section of one OxReRAM cell (bottom). h Cross-sectional Elemental Energy Loss Spectroscopy (EELS) mapping of the PdNeuRAM cell in Low Resistance States (LRS) with the structure of Pd/HfO/Ti/Pd (denoted as PdHT, also PdNeuRAM).
Results
Device characterization
We designed crossbar arrays of devices with node size (see Method), as displayed in Figure 1d, Figure 1e. The successful realization of the designed layout and stack is validated by optical microscopy and High-Resolution Scanning Transmission Electron Microscopy (HRSTEM), as depicted in Figure 1f, Figure 1g. The distinct layers of the device stacks are clearly visible in the cross-sectional HRSTEM image, as shown in Figure 1g, where the HfO layer exhibits a thickness of approximately 5 nm, sandwiched by 5 nm layers of Ti and Pd electrodes. Furthermore, HRSTEM-EELS imaging of PdHT (Figure 1h) confirms the precise elemental distribution.
To examine the electrical properties of the fabricated devices, I–V characterizations were performed, as depicted in Figure 2. Measurements on the electrical properties of the PdHT devices reveal three pivotal attributes for energy-efficient computing, including electroforming-free operation, low operating voltages, and tunable conductance. As illustrated in Figure 2a, the PdHT devices manifest forming-free bipolar switching behavior. A statistical analysis of 42 randomly selected devices from the same die (Figure 2b) reveals that the predominantly centers around 0.56 V, while the is near −0.58 V. For comparison, similar characterizations were conducted on PtHT devices fabricated using an identical process except for choosing Pt as the top and bottom electrodes (TE/BE). As shown in Figure 2d, these PtHT devices show bipolar switching, which is similar to PdHT devices, but they require an additional electroforming step (as illustrated in Figure 2e), a result consistent with previous reports [21]. The Device-to-Device variability (D2D) test (Figure 2e) confirms that PtHT devices require an electroforming voltage of approximately 2.3 V. Moreover, the statistical analysis presented in Table 1 confirms a significant alleviation in variability and a reduction in both and for the PdHT structure compared to both the PtHT devices and the state-of-the-art reports [4, 16, 17], as detailed in Table 2. Table 2 presents a comprehensive comparison of state-of-the-art forming-free ReRAM devices, encompassing both non-HfO and HfO-based systems. This analysis reveals that these devices typically exhibit one or more inherent limitations: they require elevated SET/RESET voltages, offer a constrained range of resistance states, or depend on specialized treatments to eliminate the conventional electroforming process. In contrast, our work successfully overcomes these issues, providing a more efficient and robust solution.

Figure 2. Electrical characterization of PdHT and PtHT memristors. a I-V curves of forming-free PdHT memristors. The inset schematic diagrams show the PdHT structures. b Statistical analysis of and for 42 PdHT devices. c I-V curves for tuning various . d I-V curves of conventional PtHT memristors. The inset schematic diagrams show the PtHT structures. e Statistical analysis of , and for PtHT memristors. f Reading stability test for 8 representative resistance states by readings under 0.1 V. g Retention test of resistance states for seconds with reading intervals of 60 seconds. h Endurance test of PdHT by sweeping over cycles and reading at 0.1 V. i Conductance change over including its mean, maximum, and minimum conductance values as statistically analyzed in f.
| Structures | () | () | ||
| PtHT | 11.0% | 19.2% | 0.68 V | -0.72 V |
| PdHT | 7.7% | 16.8% | 0.56 V | -0.58 V |
Table 1. Comparison of operating voltages and their variability for PtHT and PdHT devices, including cycle-to-cycle () and .
| Types | TE | BE | On/Off | Treatment | ||||
| 20 nm WO | Pd | 1.0, -3.0 | No | No | oxygen plasma | |||
| 7 nm TaO | Pt | 1.2, -0.9 | No | 600°C | annealing | |||
| no HfO | SrTiO [24] | 400 nm MoO | Pt | 3.5, -3.8 | No | 100°C | proton injection | |
| Al [25] | 150 nm SiN | p Si | 3.2, -1.5 | No | No | hydrogen plasma | ||
| TiN [26] | 5 nm HfTiO | TiN | 1.0, -1.0 | 2 | 400°C | annealing | ||
| HfO | Ge [16] | 5 nm HfO | Pd | 3.2, -0.8 | No | 600°C | Ge doping | |
| Al [27] | 60 nm HfO | p Si | 2.9, -2.5 | No | 550°C | elements doping | ||
| Pt [4] | 20 nm HfO | TiN | 0.70, -0.70 | No | No | X-Ray | ||
| TiN [7] | 3 nm HfO | TiN | 0.75, -0.6 | 2 | 450°C | annealing | ||
| Pd | 5 nm HfO | Pd | 0.56, -0.58 | No | No [This work] |
Table 2. The-state-of-the-art forming-free devices
Furthermore, the reliability of the PdHT devices was also studied and certified by further electrical measurements. The multibit capability of the PdHT devices was evaluated using a linear voltage sweeping mode. As shown in Figure 2c, the resistance is continuously modulated by varying the , thereby precisely controlling the rupture degree of the . To ascertain the stability of the distinct resistance states, a -number dependent stability test was conducted (Figure 2f), which demonstrates analog state functionality that permits the storage of eight tunable multi-bit weights within a single memory cell. Furthermore, retention (Figure 2g) and endurance (Figure 2h) assessments reveal the stable resistance states between and High Resistance States (), with outstanding non-volatility, where the resistance states remain stable for over seconds, and the device can endure cycles without noticeable performance degradation, though the performance is not presenting its optimum pattern (higher temperature of the retention test and higher number of cycles for the endurance test are needed for further investigation). Additionally, to gain deeper insight into the tunable resistance states, we conducted a comprehensive analysis of the conductance variation versus , as depicted in Figure 2i. The results reveal an exponential gradual decline in conductance with increasing , indicating that the extent of rupture can be modulated by adjusting the amplitude of . This tunability may be further enhanced by controlling additional pulse parameters, such as pulse numbers, pulse amplitudes, and pulse width of .
Understanding the forming-free behavior
We proposed a possible explanation for the forming-free behavior (see Figure 3a) based on a suite of comprehensive characterizations, including the Pd/HfO interface imaging by HRSTEM-iDPC (iDPC stands for integrated Differential Phase Contrast), which allows us to image light atoms in the presence of heavy atoms (as shown in Figure 3b and Figure S5), the surface profiles of Pd and Pt thin films obtained via Atomic Force Microscopy (AFM) (Figure S6, Figure S6), the Arrhenius activation energy measurements for both PdHT and PtHT memristors using a semiconductor analyzer (Figure 3c), and the elemental distribution profiles acquired from Rutherford Backscattering Spectroscopy (RBS) (Figure 3d).

Figure 3. Investigation of forming-free behavior. a schematic illustration of a forming-free mechanism. b STEM-HAADF and iDPC images of the PdHT device. The scale bar is 5 nm. c Activation energy study for PtHT (orange) and PdHT (dark cyan). d RBS elemental distribution line profile along PdHT cross section.

Figure S5. a Oxygen ions observation using STEM-iDPC. b STEM-iDPC cross section image of pristine PdHT to indicate the Pd atoms through red cycles across HfO layer

Figure S6. a AFM image of Pd surface with a roughness of 715 pm. b AFM image of Pt surface with roughness of 1087 pm. c AFM surface analysis of PdHT device and d AFM surface analysis of PtHT device
The activation energies for the pristine PtHT and PdHT devices are compared in Figure 3c via the Arrhenius equation, , where denotes the resistivity, is the absolute temperature in Kelvin, is the pre-exponential factor, represents the activation energy, and is the Boltzmann constant. As illustrated in Figure 3c, the activation energy exhibits two distinct regimes: a Low-Temperature (LT) regime (300 K to 400 K) marked by lower activation energies (denoted as ) and a High-Temperature (HT) regime (400 K to 470 K) characterized by larger activation energies (denoted as ).
Notably, the pristine PdHT devices exhibit activation energies of eV and eV, which are significantly lower than those of the PtHT devices ( eV and eV). The activation energy of PtHT is possibly attributed to the diffusion barrier associated with the migration of doubly positively charged oxygen vacancies inside HfO, in accordance with previous reports [20, 28, 29, 30]. It may arise from the diffusion barrier of doubly negatively charged interstitial oxygen ions, based on existing ab initio calculations [20, 30, 31]. Meanwhile, it is noteworthy that the Frenkel-Pair (FP) dissociation process in HfO typically requires activation energies on the order of 5.2 eV [20, 32]. Furthermore, if the FP process comprises a doubly positively charged pair (), its formation energy is 5.8 eV [33], a value closely related to the electroforming process, as reported previously [20, 34], implying that FP dissociation is highly unlikely to occur under both LT and HT conditions without an applied voltage bias.
In pristine PtHT devices, the presence of is expected in amorphous hafnia devices and is introduced during the sputtering process [35], though it remains at a low level (see more details in the ESI part 1 about doubly positively charged oxygen vacancies formation). In contrast, in PdHT devices, activation energy appears to be attributed to singly positively charged () defects or interstitial oxygen defects. This difference is likely due to the incorporation of Pd in HfO at the Pd/HfO interface (as observed in Figure 3b), adsorbing oxygen ions around the Pd atoms (Pd–), generating more oxygen vacancies along the conductive path. Interstitial Pd atoms alter the defect landscape in hafnia by creating more oxygen vacancies and donating electrons, thereby stabilizing oxygen vacancies as singly positively charged () defects and forming – pairs, which serve as traps for electrons hopping [33]. Additional activation energy measurements on the HRS state of both PdHT and PtHT memristors (Figure S3) demonstrate that after electroforming (i.e., completion of FP dissociation) in PtHT devices, the dominant charges exhibit an activation energy of approximately 0.22(2) eV, similar to that observed in PdHT devices, which is also consistent with the fact that after electroforming of the PtHT devices, the IV characteristics of both types of devices are basically indistinguishable.

Figure S3. Activation energy comparison for PdHT and PtHT
Considering above all, the forming-free phenomenon likely originates from Pd atoms initially forming a conductive bridge through bonding oxygen ions from hafnia, generating more oxygen vacancies (see Figure 3a). The incorporation of Pd supplies electrons that effectively reconstruct oxygen vacancies/ions in the Pd–O–Hf configuration relative to the conventional Hf–O configuration in PtHT memristors. Furthermore, under minimum bias, it provides the maximum electrons among the Pd bridges, and its inherent special configuration provides electron injection by effectively providing electrons to defective sites and eases the forming process [20] and conductive path for the initial states.
To substantiate the hypothesis of the Pd-induced conductive bridges in of PdHT, RBS measurements were conducted. As depicted in Figure 3d, the layer in the PdHT memristors exhibits a higher Pd concentration compared to the Pt concentration in PtHT devices. Remarkably, even after annealing the PdHT and PtHT devices at 573 K for 5 minutes, the Pd atomic ratio within the layer remains higher than that of Pt in the PtHT devices, as further illustrated in Figure S7. This observation agrees with the lower migration barrier of Pd into semiconductors compared with other metals such as Li, Cu, Ag, Pt, and Au [36].

Figure S7. a Atomic ratio comparison of PtHT and PdHt under 573K (denoted as 573 K). b STEM-EELS line profile of LRS state for PdHT, where it shows the Hf ions and oxygen ions dissociation, and Hf mixed with Pd layer while oxygen ions migrate into the Ti layer.
We employed HRSTEM-HAADF to examine the Pd/ interface. As shown in Figure 3b, the images reveal that the Pd-Hf atoms are clearly observed above Pd electrodes taken in the [011] zone, showing a region where crystallized (evidenced by small, moderately bright atoms, marked in purple spheres as shown in the most right HRSTEM-iDPC image) intermixed with Pd (characterized by larger, highly bright atoms, marked in orange spheres), which is further emphasized in Figure S5 with red cycles. For further validation of the Pd–O–Hf configuration, please refer to ESI part 2 for further discussion on it.
The influence of Pd on forming voltages was examined via the I–V characteristics of various devices with various electrodes (see Figure S4, Figure S4). The critical role of Pd electrodes is underscored by their ability to markedly reduce or even eliminate the need for electroforming voltages, in stark contrast to the PtHT structure (Figure S4). Moreover, as the thickness increases, a distinct electroforming behavior emerges (Figure S4), likely due to the finite extent of the Pd–O–Hf configuration within , which results in an exponential increase in resistance with a linear increase in thickness.

Figure S4. Different combination of Pt and Pd as BE/TE and their I-V curve with 5nm of HfO (a, b, c) and 5nm of HfO with Pd electrodes for BE/TE (d).
Investigation of conduction mechanism
In ReRAM devices, multiple conduction mechanisms [37, 38], such as Ohmic conduction, trap-mediated conduction, Trap-Assisted Tunneling (TAT), Space-Charge-Limited Conduction (SCLC), Poole–Frenkel emission, Schottky emission, and valence change-based mechanisms, collectively dictate the overall device behavior. In particular, SCLC is a prominent mechanism observed in ReRAM devices during high-voltage regime [38]. Here, the current is constrained by a space-charge-limited region in which charge carriers, injected from the electrodes, traverse localized trap states within the resistive switching material. The SCLC behavior is critically influenced by the trap density and trap energy levels, which govern the mobility and transport of the charge carriers.
A trap-controlled SCLC regime can be segmented into two distinct portions: the trap-unfilled (trap-limited) regime and the trap-filled (trap-free) regime. In the Mott-Gurney region, characterized by the power law , a steep increase in current is observed at high electric fields. This behavior is typically identified by an initial Ohmic conduction at low fields, followed by a transition to a power-law dependence as electrode-injected electrons surpass the equilibrium concentration. Consequently, SCLC conduction is more likely when the electrode contact exhibits high carrier injection efficiency [38]. The SCLC conduction mechanism is quantitatively described by the Mott-Gurney law (Eq. 1), which relates the current density to the applied voltage and material properties:
Here, denotes the dielectric permittivity, is the carrier mobility, represents the film thickness, and is the applied voltage.
The slope () of the double logarithmic I–V curve evolves from approximately 1.0 (indicative of trap-limited Ohmic conduction) through 1.5 (reflecting the trap-unfilled transitional region) to nearly 2.0 (characteristic of trap-free SCLC) for both HRS and LRS, as observed in Figure 4a and Figure 4d for PdHT devices under a 0.1 mA compliance current. On the contrary, when the compliance current is increased to 0.5 mA, the slope for HRS (see Figure 4b and Figure 4e) is dominated by Ohmic conduction in both HRS and LRS, a consequence of reduced interface state density at the Pd/HfO interfacial region owing to filament formation. Conversely, PtHT devices exhibit ohmic conduction under a 0.1 mA compliance current for both HRS and LRS (Figure 4c and Figure 4f). These results indicate that at the Pd/HfO interface in PdHT devices, shallow defect states are preferentially formed, enabling electrons to enter the conduction band (see Figure 4g and Figure 4h). In contrast, PtHT devices tend to form deep defect states, wherein electrons tunnel into vacancy defect sites [38] (see Figure 4i).

Figure 4. Investigation of conduction mechanism from I-V curves fitting for PtHT and PdHT. a I-V curves of PdHT device with 0.1 mA compliance current, b 0.5 mA compliance current. c Present PtHT device I-V curve with 0.1 mA compliance current. d, e, f The fitted I-V plots of positive bias parts a, b, c, respectively. g, h Illustration of HRS and LRS band diagrams of PdHT and electrons drifting between shallow vacancy states, respectively. i The band diagram of PtHT LRS states and electron hopping between deep defective sites.
Taking all the aforementioned analysis in Figure 3 and Figure 4, including the activation energy study, HRSTEM observations of Pd atoms, EELS and RBS elemental distribution profiles of PdHT, and I–V curve fitting, into consideration, in PdHT devices, the Pd–O–Hf configuration, including the Pd-induced conductive bridges and electron injection from Pd atoms in , facilitates oxygen vacancies reconfiguration and electron drift toward the Pd electrode via tunneling. It hereby contributes to the elimination of the electroforming process. After the RESET operation for PdHT devices (from initial state to HRS), the Pd-induced conductive bridges will be partially damaged irreversibly by scavenging the oxygen vacancies through oxygen ions from , leaving the shallow defective states thanks to the oxygen affinity nature of Pd. In contrast, in PtHT devices, electrons predominantly hop between deep defect sites generated by high oxygen vacancy concentrations after the electroforming process under higher electric fields.
Device calibration and electrical simulation
To evaluate the energy consumption of the PdHT and PtHT devices, we calibrated the I– curves of the real devices using the JART model [39]. In the case of PdHT as illustrated in Figure 5, the simulated I–V performance exhibited excellent concordance with the experimental results (see Figure 5a). Regarding the multilevel resistance state performance, the rupture of the displayed a well-controlled, gradual transition upon tuning the , as demonstrated in Figure 5b, where multiple distinguishable I–V curves were observed from Cadence simulation, which implies that the results of PdHT device calibration agree well with experimental results. Similarly, the incremental change in conductance by varying bias amplitude for showed remarkable agreement between experimental observations and simulation results (Figure 5c). Furthermore, distinct resistance states can be programmed using variable RESET stop voltages with a pulse duration, leading to the conclusion that 8 discernible resistance states can be established with corresponding programming schemes, as depicted in Figure 5d. The detailed programming protocol was further illustrated in Figure 5e, where each programmed resistance state is successfully achieved using a writing/reading pulse of at V/ mV, respectively. To assess the power consumption during programming, we integrated the instantaneous power over time, as shown in Figure 5f. Here, both RESET/SET energy consumptions contribute to the total programming energy, while the sum of HRS and LRS reading energies represents the total energy consumption during reading. The programming energy consumption for 8 distinct conductance states for both PdHT and PtHT devices is summarized in Table S1 and Table S2, respectively.

Figure 5. Power consumption evaluation a Experimental I– curve calibration with JART memristor model. b MLR verification using sweeping by changing . c Comparison of conductance gradual change for experimental and simulation results. d 8 resistance states programming schemes, including the results of currents and resistance over time. e Programming scheme of the resistance state with . f Power consumption evaluation for programming the conductance states in e.
| (V) | (V) | (W) | (s) | (J) | (W) | (J) | (J) | ||
| G1 | HRS1 | -1.5 | 0.1 | 7.48e-07 | 1.00e-05 | 7.48e-12 | 4.20e-04 | 4.20e-09 | 4.21e-09 |
| G2 | HRS2 | -1.4 | 0.1 | 9.15e-07 | 1.00e-05 | 9.15e-12 | 3.96e-04 | 3.96e-09 | 3.97e-09 |
| G3 | HRS3 | -1.3 | 0.1 | 1.14e-07 | 1.00e-05 | 1.14e-12 | 4.80e-04 | 4.80e-09 | 4.80e-09 |
| G4 | HRS4 | -1.2 | 0.1 | 1.44e-06 | 1.00e-05 | 1.44e-11 | 5.72e-04 | 5.72e-09 | 5.73e-09 |
| G5 | HRS5 | -1.1 | 0.1 | 1.88e-06 | 1.00e-05 | 1.88e-11 | 6.30e-04 | 6.30e-09 | 6.32e-09 |
| G6 | HRS6 | -1.0 | 0.1 | 2.57e-06 | 1.00e-05 | 2.57e-11 | 7.06e-04 | 7.06e-09 | 7.09e-09 |
| G7 | HRS7 | -0.9 | 0.1 | 3.80e-06 | 1.00e-05 | 3.80e-11 | 7.76e-04 | 7.76e-09 | 7.80e-09 |
| G8 | LRS | -0.5 | 0.1 | 5.09e-06 | 1.00e-05 | 5.09e-11 | 3.97e-04 | 3.97e-09 | 4.02e-09 |
Table S1. Energy consumption (PdHT). Measured parameters for various conductance states (G1–G8).
| G | (V) | (V) | (V) | (W) | (s) | (J) | (W) | (J) | (J) | |
| G1 | HRS1 | -1.5 | 1.1 | 0.1 | 5.96e-06 | 1.00e-05 | 5.96e-11 | 8.92e-04 | 8.92e-09 | 8.98e-09 |
| G2 | HRS2 | -1.4 | 1.1 | 0.1 | 6.27e-06 | 1.00e-05 | 6.27e-11 | 9.44e-04 | 9.44e-09 | 9.51e-09 |
| G3 | HRS3 | -1.3 | 1.1 | 0.1 | 6.69e-06 | 1.00e-05 | 6.69e-11 | 1.00e-03 | 1.00e-08 | 1.01e-08 |
| G4 | HRS4 | -1.2 | 1.1 | 0.1 | 7.35e-06 | 1.00e-05 | 7.35e-11 | 1.07e-03 | 1.07e-08 | 1.08e-08 |
| G5 | HRS5 | -1.1 | 1.1 | 0.1 | 8.39e-06 | 1.00e-05 | 8.39e-11 | 1.12e-03 | 1.12e-08 | 1.13e-08 |
| G6 | HRS6 | -1.0 | 1.1 | 0.1 | 8.95e-06 | 1.00e-05 | 8.95e-11 | 1.07e-03 | 1.07e-08 | 1.07e-08 |
| G7 | HRS7 | -0.9 | 1.1 | 0.1 | 9.02e-06 | 1.00e-05 | 9.02e-11 | 9.78e-04 | 9.78e-09 | 9.87e-09 |
| G8 | LRS | -0.5 | 1.1 | 0.1 | 4.51e-06 | 1.00e-05 | 4.51e-11 | 5.88e-04 | 5.88e-09 | 5.93e-09 |
Table S2. Energy consumption (PtHT). Measured parameters for various conductance states (G1–G8). Averaged Power and Energy
System-level energy simulation for neuromorphic applications
For the demonstration of the proposed PdHT ReRAM device in realistic neuromorphic applications, we designed two deep convolutional SNNs targeted at the classification of the N-MNIST [40] (Figure 6a) and IBM DVS128 Gesture [41] (Figure 6b) datasets, respectively. These networks are implemented in Python using primitives from the open-source Spike LAYer Error Reassignment (SLAYER) [42] and PyTorch [43] frameworks, and they are trained with a variant of the backpropagation algorithm. Spiking neurons are modeled on the Spike Response Model (SRM), an advanced generalization of the ubiquitous Integrate-and-Fire (I&F) neuron model [44]. At the output, the class corresponding to the neuron that produced the highest number of spikes is selected as the winning class. For further methodological details, please refer to the Methods section.

Figure 6. System-level energy simulation for neuromorphic applications a, b Model architectures of the N-MNIST and Gesture SNNs, respectively. c, d Overall per-layer write energy consumption after one-time programming of pre-trained SNN instances and read energy consumption during the inference of the test set for the N-MNIST and Gesture SNNs, respectively. e, f Overall write energy consumption per training epoch in the case of online learning for the N-MNIST and Gesture SNNs, respectively.
To estimate the energy consumption at the system level, we perform a series of experiments and report the overall read and write energy measured collectively for all ReRAM devices needed to store the synaptic weights across each network. In the first set of experiments, we train the two SNNs in software, and then we write the final synaptic weights directly onto the ReRAM devices. For simplicity and generalization purposes, there is no limit to the maximum number of employed devices. The first two bars in Figure 6c and Figure 6d report the write energy, , per layer on a logarithmic scale for the N-MNIST and Gesture SNNs, respectively. As anticipated, the write energy scales proportionally with the number of synapses, since a greater number of weights necessitates more devices. Notably, the PdHT devices achieve an overall reduction in write energy of approximately 43% with respect to that of PtHT for each SNN.
In the case of online learning, the synaptic weights are updated and written back on the ReRAM devices at every epoch of the training of each SNN and the overall write energy per epoch is illustrated in Figure 6e and Figure 6f for the two case studies, respectively. Initially, the synaptic weights are set to small random values, resulting in a peak write energy equal to the total write energy occurring by summing the corresponding bars for all layers in Figure 6c and Figure 6d. This is because in the first iteration almost all ReRAM devices need to be programmed to a new value. As training progresses, only those devices that require updates contribute to the write energy, leading to a gradual reduction that mirrors the convergence of the synaptic weights, thereby lowering the overall energy consumption.
Regarding the read energy , during inference, represented by the rightmost two bars in Figure 6c and Figure 6d, the energy consumption per layer exhibits a pattern distinct from that of the write energy. In SNNs, each layer functions as a filter that reduces the number of spikes transmitted to subsequent layers, resulting in progressively sparser spiking activity. Therefore, the read energy depends not only on the number of synapses but also on the density of incoming spikes. For example, layer SF1 of the Gesture SNN consumes only more energy than its predecessor, despite having more synapses. Overall, both networks exhibit an approximate 73% reduction in read energy during the complete inference of their respective test sets. The average energy per spike during inference is summarized in Table S7 and demonstrates a comparable energy reduction when utilizing the proposed PdHT devices over PtHT ones.
| N-MNIST SNN | Gesture SNN | |
| PtHT | 33.6 nJ | 65.7 nJ |
| PdHT | 9.4 nJ | 17.7 nJ |
Table S7. Average read energy per spike.
Discussion
In this work, we reported the successful CMOS-compatible fabrication of PdHT ReRAM devices, PdNeuRAM. High-resolution imaging analyses, including SEM, HRSTEM, and AFM, confirm well-defined device stacks with minimal interfacial roughness. Electrical characterization reveals forming-free switching at relatively low operating voltages, with around 0.56 V and averaged at V. Notably, the PdHT devices exhibit reduced C2C and D2D variability comparable to their PtHT counterparts, indicating robust switching behavior and enhanced reliability.
From previous studies on oxygen chemisorption on Pd and Pt films and Schottky barriers [45] and MIGS [46, 47], it seems unlikely that the Schottky barrier differences cause the electroforming differences. The absence of an electroforming step in PdHT devices is primarily attributed to the presence of Pd atoms within the layer for a few nanometers, which probably contributes to the charge redistributions, acting as Pd-induced conductive bridges, thus eliminating the need for a high-voltage electroforming process. Furthermore, regardless of the role of Pd in the conduction mechanism, the PdHT presents similar resistive switching properties to that of PtHT as indicated by the identical activation energy, which is in turn in agreement with the behavior reported in the previous devices [4, 7].
Current-voltage (-) analysis indicates that both PdHT and PtHT devices operate under a SCLC mechanism. At low bias, conduction follows an ohmic behavior. As the voltage increases, trap-mediated processes dominate; injected carriers fill shallow defect states, transitioning the conduction to a quadratic dependence on voltage (), consistent with the Mott-Gurney law. In PdHT devices, intrinsic insulating layer defects and interfacial imperfections, enhanced by Pd-O-Hf interactions, create shallower defect states that promote efficient electron transport and lower variability. In contrast, PtHT devices exhibit deeper defect states, which vary with electroforming voltage and material modifications, leading to increased variability.
The integration of these forming-free multibit ReRAM devices into neuromorphic systems has been successfully demonstrated by employing them as synaptic elements in SNNs. Using three-device (-bit) configurations to represent 9-bit quantized synaptic weights, the networks are implemented for real-world pattern recognition tasks on the N-MNIST and IBM’s DVS128 Gesture datasets. The SNN architectures, developed using frameworks such as SLAYER and PyTorch, achieve competitive classification accuracies (94.6% for N-MNIST and 85.6% for IBM’s DVS128 Gesture). However, energy consumption analyses reveal that the PdHT devices yield substantial energy savings, with write and read operations reduced by approximately 43% and 73%, respectively. This energy efficiency, combined with stable multilevel conductance programming and robust endurance, makes the proposed PdNeuRAM technology highly promising for next-generation, low-power neuromorphic computing applications.
Methods
ReRAM cell fabrication
First, the SiO substrate was sequentially cleaned by fuming nitric acid, acetone, isopropyl alcohol, and deionized (DI) water in an ultrasonic oscillator. Then, the metal films were deposited on the chemically cleaned SiO substrate through electron beam evaporation in a vacuum chamber with the pressure of torr. 5 nm Ti (the adhesion layer) and 50 nm Pd (BE) were deposited by ATC 2400 Sputtering System under a vacuum level of torr; 5 nm or 10 nm HfO (the oxide layer) was deposited by Sputter (Alliance Concepts system) under a vacuum level of torr; 5 nm or 10 nm Ti (interface layer) and 50 nm Pd (TE), which were deposited by ATC 2400 Sputtering System under a vacuum level of torr. By lifting off technique, the nodes with dimensions are 5 µm × 5 µm are fabricated by negative photoresist (AZnLof@2020) and positive photoresist (S1805) for photolithography (Heidelberg microMLA) (see the profiles in Figure S2). The pads dimensions are 15 µm × 15 µm. The BE via was etched by CHF/Ar gases (Sentech Etchlab 200) for 270 s. The optical microscopy image of the device was obtained by an optical microscope (Olympus BX51) 50×. Scanning Electron Microscope (Hitachi S-4800) was applied to observe the details of the nodes and cross-section.

Figure S2. Pd deposition on positive resist (a) and negative resist (c). (b), (d) are the cross section profile of as-fabricated different resists
Electrical performance characterization
To study the electrical performance of both structures cells, a probe station (CAS-CADE) equipped with a semiconductor analyzer (Keysight B1500A) was utilized to conduct the electrical measurement for the ReRAM cell at room temperature (300 K), during which a bias voltage was applied to the TE while the BE was grounded. The double linear sweeping bias is from 0 V to 1.5 V with a step 100 mV for the SET sweeping process and 0 V to −1.5 V step −100 mV for RESET sweeping process.
HRSTEM characterization
In our study, we implemented a multifaceted STEM methodology combining HAADF, EELS, and iDPC techniques to achieve a comprehensive characterization of advanced materials. We began with STEM-HAADF imaging, which utilizes a high-angle annular dark-field detector to collect thermally diffuse scattered electrons. This method provides Z-contrast images where the intensity scales roughly as , allowing us to resolve atomic columns and differentiate heavy and light elements based on their scattering power. Precise alignment and calibration are performed with dwell times in the order of µs to achieve high resolution and reliable compositional mapping.
STEM-EELS was employed to complement the HAADF images by providing elemental and chemical state information through electron energy loss spectroscopy. EELS spectra, collected in parallel with HAADF imaging, enabled the identification of subtle variations in bonding and composition at the nanoscale.
To enhance the detection of light elements and obtain electrostatic potential maps, we applied the iDPC method. In this mode, the segmented DF4 detector is selected via the Velox DPC/iDPC control panel, and the resulting differential signals are integrated. This process improves image contrast and reduces sensitivity to defocus and thickness variations, yielding an image where contrast is roughly proportional to atomic number.
RBS characterization
RBS was employed to determine the elemental depth profiles of the layers in the forming-free PdHT devices. RBS measurements were conducted using a Kobe Steel HRBS-V500 system with a ion beam accelerated at 400 keV. The beam was directed onto the sample at an incidence angle of 45°, and backscattered ions were detected at a scattering angle of 107.5°. A 512 channel detector with an energy resolution of 2 keV was used. The RBS data was analyzed using the Kobe Steel AnalysisIB software. The measurements revealed the Pd ratio in the layer, confirming the intermixing of Pd-O-Hf. This configuration is critical for reducing the oxygen diffusion barrier and facilitating the forming-free behavior.
Activation energy measurement
The activation energy of the forming-free -based ReRAM devices was measured using a Keithley B1500A Semiconductor Parameter Analyzer in conjunction with a thermionic heater. In this setup, the devices were mounted on a temperature-controlled stage that provided stable and uniform heating over a defined temperature range. To prevent oxidation and moisture interference during high-temperature measurements, a continuous flow of high-purity gas was maintained throughout the experiment. The B1500A recorded the current-voltage (I-V) characteristics at incremental temperature steps. The temperature-dependent resistivity was then derived from the I-V data, and an Arrhenius plot of versus was constructed. The slope of the linear fit, when divided by Boltzmann’s constant (), yielded the activation energy. This method ensured precise thermal control and reproducible measurements, critical for understanding the conduction mechanisms in the devices.
Device calibration and power consumption evaluation
The device model is developed based on the JART model, which accurately replicates the electrical behavior of the forming-free PdHT ReRAM devices. Calibration of the model is achieved by fitting the experimental DC I–V curves, obtained using a Keisight B1500A, with the simulation results. In our setup, both DC sweep and pulse-based measurement configurations are implemented using Cadence simulation tools. The DC sweep methodology enables the extraction of key parameters such as SET/RESET voltages and memory window, while pulse-based simulations reveal the dynamic switching characteristics and multi-resistance states under various programming schemes. This integrated simulation approach provides critical insights into the energy consumption during programming and reading operations, ensuring consistency between experimental and simulated results.
Device-based system-level SNN simulation
The synaptic weights of the two SNNs are quantized to integer values of 9 bits, which are represented by a combination of three multistate ReRAM devices with 8 states each, i.e., -bit ReRAM devices. In the presented experiments, we performed the training and inference for the two case studies and reported the write and read energy footprint of the PdHT and PtHT ReRAM devices in different scenarios. The focus is kept on the energy impact of the crossbar arrays hosting the devices since it is not dependent on the actual realization of the periphery circuitry, allowing space for generalization no matter the actual implementation of the neuromorphic hardware accelerator. In all cases, we present a comparison between state-of-the-art PtHT ReRAM devices and the proposed PdHT ones.
N-MNIST SNN: The N-MNIST dataset is a neuromorphic, that is, spiking, version of the MNIST dataset, which comprises images of handwritten arithmetic digits in grayscale format [40]. It consists of sample images that are generated from the saccadic motion of a Dynamic Vision Sensor (DVS) in front of the original images in the MNIST dataset. The samples in the N-MNIST dataset are not static, i.e. they have a duration in time of 300 ms for each. The dataset is divided into a training set of samples and a test set of samples. The SNN architecture shown in Figure 6a comprises 3 convolutional layers (SC1, SC2, and SC3) followed by 2 fully connected ones (SF4 and SF5). The classification accuracy on the test set is 94.56%, which is comparable to the performance of state-of-the-art level-based DNNs.
Gesture SNN: The IBM’s DVS128 Gesture dataset consists of 29 individuals performing 11 hand and arm gestures in front of a DVS, such as hand waving and air guitar, under 3 different lighting conditions [41]. Samples from the first 23 subjects are used for training, and samples from the last 6 subjects are used for testing. In total, the dataset comprises 1342 samples, each of which lasts about 6s, making the samples 20 times longer than those in N-MNIST. To speed up the neuromorphic simulations, we trimmed the length of the samples to about 1.5 s. The proposed SNN architecture consists of a pooling layer SP0 to reduce the input samples, 2 convolutional layers (SC1 and SC2) followed by a pooling layer each. The data coming from the last pooling layer is flattened and fed to 2 fully connected layers (SF3 and SF4). The network architecture is presented in Figure 6b. The network performs with an accuracy of 85.61% in the test set, which is acceptable considering the shortened samples in the dataset and the shallower architecture compared to the originally proposed architecture in [41].
In both use cases, the SNNs are implemented with SRM spiking neurons, effectively capturing the temporal dynamics of spiking activity. Training uses a variant of the backpropagation algorithm, in which error is calculated on the probability of each neuron to change spike state, i.e., fire a spike if it was in a resting state or stop firing, in the next timing instance [42].
Data availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
The code that support the findings of this study are available from the corresponding author upon reasonable request.
Acknowledgements
This work was supported by Delft University of Technology. We acknowledge the kind support of the Delft Kavli Nanofabrication cleanroom, especially Marinus Fischer, Roald van der Kolk, Charles de Boer, and Eugene Straver, for their unconditional assistance. We also specially thank Lennart P. L. Landsmeer and Emmanouil Arapidis for their kind support and discussion.
Author contributions
EH, HA, RI: Conceptualization. EH: device fabrication, electrical measurement. EH, HA, RI, SH, GG: Results Analysis and Visualization. MA, BN, EH, HA: STEM, EELS, and iDPC results acquisition and analysis. EH, BN, HA: AFM, activation energy analysis. EH, HX: Device calibration based on model. TS, EH, HA, AG: SNN simulation, visualization, and analysis. AMS, EH, HA, LB, NE: RBS analysis. EH, BN, TS, and HA: manuscript draft & finalization. All coauthors contribute to the manuscript review and revision.
Competing interests
The authors declare no competing interests.
Additional information
Supplementary information
The online version contains supplementary material available at https://doi.org/xxx/xxxx.
Electronic Supplementary Information
Doubly positively charged oxygen vacancies formation
Within pristine PtHT devices, the formation of doubly charged oxygen vacancies, , is inherently expected in amorphous hafnia-based PtHT devices. These vacancies are primarily generated during the sputtering process [35] and through reactive interactions between titanium (Ti) and nonstoichiometric hafnium oxide, . In these reactions, doubly negatively charged oxygen ions combine with Ti to yield the stable electride compound at the Ti/ interface, thereby functioning as an effective electron reservoir while concomitantly leaving behind vacancies in . The preferential stabilization of Ti-based electrides thanks to the higher electron affinity nature of Ti than that of Hf (according to the work function of the elements), as opposed to the inherently unstable electron-localized variants without the presence of Ti or , further substantiates this mechanism [20, 48]. Conversely, in PdHT devices, the incorporation of palladium (Pd), which possesses a lower electron affinity relative to platinum (Pt), may promote electron donation. This process facilitates the reduction of to while simultaneously enhancing the movement of interstitial oxygen ions with the presence of Ti beyond room temperature so that it will facilitate the formation of .
Further validation of Pd-O-Hf configuration presence
Meanwhile, a finding that is corroborated by AFM measurements (Figure S6, Figure S6). Statistical analysis shows that the Pd film exhibits a roughness of 715 pm, markedly lower than the 1087 pm measured for the Pt film, implying that the Pd surface is less prone to form spikes [49] and thereby reduces potential measurement errors in RBS analysis. From HRSTEM observation, as shown in Figure S2 and Figure S2, we can also conclude that there are no Pd spikes generated. Furthermore, the STEM-EELS line profiles confirm the anticipated elemental composition in LRS (Figure S7), with the distinct elemental layers aligning precisely with the design illustrated in Figure 1d. Notably, the measurements verify the Pd-O-Hf configuration at the Pd/ (LRS) interface and in the layer, a possible consequence of the lower diffusion barrier and higher propensity for formation [36, 50] of Pd.
MIGS model explanation
The MIGS model attributes SBH pinning to a finite density of metal Fermi energy, we use the relation:
where is the Schottky barrier height, represents the effective work function of the metal, is the pinning factor at the metal/ interface, is the charge neutrality level of the dielectric layer, is the electron affinity of , and is the conduction band minimum. Based on this model (with , see ref. [46]), we calculate the SBH of PtHT and PdHT,
Study on the SBH for Pd/HfO and Pt/HfO
Apart from the aforementioned observations and analyses, we also investigated the band structures and Schottky barriers of the PdHT and PtHT memristors to ascertain whether the forming-free behavior arises from differences in Schottky barrier heights. From previous studies [45, 51, 52, 53], it is well established that oxygen can chemisorb on noble metals such as Pt and Pd at room temperature, forming a stable layer of atomic oxygen or surface oxides. This phenomenon occurs due to the strong interaction between oxygen and the d-electrons of these metals under ambient conditions. In Figure S1, we summarize the oxygen chemisorption energies on Pd(111) and Pt(111), along with the oxygen dissociation barriers, from which it is evident that Pd(111) exhibits a higher propensity to form Pd(111)-O than Pt(111). Such behavior significantly influences the work functions of Pd and Pt, as reported in previous studies [52, 53], where the effective work functions of Pt and Pd are in the ranges of 6.00–6.10 eV and 6.02–6.37 eV, respectively. Moreover, oxygen adsorption affects the work function differently: for Pt, it increases by 0.35–0.45 eV, whereas for Pd, the increase is 0.90–1.25 eV, a disparity attributed to the distinct dipole moments in Pt–O and Pd–O bonding. This distinct impact is corroborated by the observation of oxygen ions via the STEM-iDPC technique at the Pd lattice in the Pd/HfO interface (Figure S5). In HRSTEM-iDPC images, as shown in Figure S5, especially when combined with HAADF imaging, it serves as a complementary imaging mode. While HAADF primarily provides Z-contrast (sensitive to atomic number, thus highlighting heavier elements), iDPC collects the differential phase contrast signals, often using a segmented or pixelated detector, and then integrates them to reconstruct the projected electrostatic potential of the sample. This integration allows for enhanced contrast of light elements and detailed structural information that might be missed by HAADF alone. From this image, we can observe that at the interface of Pd/HfO oxygen is identified among Hf atoms and on Pd atoms.

Figure S1. Atomic schematics of oxygen adsorption on Pd or Pt including adsorption energy and dissociation barriers, according to the previous study [45]. it
Furthermore, the Charge Neutrality Level (CNL) for HfO in both the Pd/HfO and Pt/HfO systems has been determined experimentally and theoretically, with the extracted CNL () reported as 4.36 eV [54]. From Metal-Induced Gap States (MIGS) [46, 47], the calculated Schottky barrier heights for Pd/HfO and Pt/HfO are 4.17–4.22 eV and 4.18–4.37 eV, respectively. This result implies that the difference in activation energy E observed in Figure 3c between the PdHT and PtHT memristors does not originate from disparities in Schottky barrier height. Consequently, this finding indirectly underscores the critical impact of the electrode materials on the forming-free behavior.
| Write Energy (J) | Average Write Energy per Synapse (J) | ||||
| Layer | Synapses # | PtHT | PdHT | PtHT | PdHT |
| SC1 | 588 | 1.57e-05 | 9.15e-06 | 2.67e-08 | 1.56e-08 |
| SC2 | 2400 | 6.31e-05 | 3.68e-05 | 2.63e-08 | 1.54e-08 |
| SC3 | 48000 | 1.34e-03 | 7.69e-04 | 2.78e-08 | 1.60e-08 |
| SF1 | 10080 | 2.79e-04 | 1.62e-04 | 2.77e-08 | 1.61e-08 |
| SF2 | 840 | 2.22e-05 | 1.40e-05 | 2.64e-08 | 1.67e-08 |
| Total | 61908 | 1.72e-03 | 9.91e-04 | 2.77e-08 | 1.60e-08 |
Table S3. Write energy of one-time programming of synaptic weights for the pre-trained N-MNIST SNN.
| Write Energy (J) | Average Write Energy per Synapse (J) | ||||
| Layer | Synapses # | PtHT | PdHT | PtHT | PdHT |
| SC1 | 800 | 2.15e-05 | 1.25e-05 | 2.69e-08 | 1.57e-08 |
| SC2 | 4608 | 1.21e-04 | 7.06e-05 | 2.63e-08 | 1.53e-08 |
| SF1 | 1048576 | 2.91e-02 | 1.66e-02 | 2.77e-08 | 1.58e-08 |
| SF2 | 5632 | 1.55e-04 | 8.86e-05 | 2.76e-08 | 1.57e-08 |
| Total | 1059616 | 2.94e-02 | 1.68e-02 | 2.77e-08 | 1.58e-08 |
Table S4. Write energy of one-time programming of synaptic weights for the pre-trained Gesture SNN.
| Layer | Synapses # | PtHT | PdHT | PtHT | PdHT |
| SC1 | 588 | 2.61e-04 | 7.36e-05 | 4.43e-07 | 1.25e-07 |
| SC2 | 2400 | 8.04e-03 | 2.29e-03 | 3.35e-06 | 9.55e-07 |
| SC3 | 48000 | 2.17e-03 | 5.55e-04 | 4.52e-08 | 1.16e-08 |
| SF1 | 10080 | 3.65e-04 | 9.93e-05 | 3.62e-08 | 9.83e-09 |
| SF2 | 840 | 1.82e-05 | 5.85e-06 | 2.17e-08 | 6.96e-09 |
| Total | 61908 | 1.08e-02 | 3.03e-03 | 1.75e-07 | 4.89e-08 |
Table S5. Read energy per layer during inference of the -MNIST SNN. Read Energy (J); Average Read Energy per Synapse (J).
| Layer | Synapses # | PtHT | PdHT | PtHT | PdHT |
| SC1 | 800 | 9.93e-03 | 2.63e-03 | 1.24e-05 | 3.29e-06 |
| SC2 | 4608 | 3.23e-02 | 9.23e-03 | 7.01e-06 | 2.00e-06 |
| SF1 | 1048576 | 4.69e-02 | 1.23e-02 | 4.48e-08 | 1.17e-08 |
| SF2 | 5632 | 4.85e-05 | 1.58e-05 | 8.61e-09 | 2.81e-09 |
| Total | 1059616 | 8.92e-02 | 2.42e-02 | 8.42e-08 | 2.28e-08 |
Table S6. Read energy per layer during inference of the Gesture SNN. Read Energy (J); Average Read Energy per Synapse (J).
References
- [1]Kudithipudi, D., Schuman, C., Vineyard, C.M., Pandit, T., Merkel, C., Kubendran, R., Aimone, J.B., Orchard, G., Mayr, C., Benosman, R., et al.: Neuromorphic computing at scale. Nature 637(8047), 801–812 (2025)DOI
- [2]Hamdioui, S., Xie, L., Du Nguyen, H.A., Taouil, M., Bertels, K., Corporaal, H., Jiao, H., Catthoor, F., Wouters, D., Eike, L., et al.: Memristor based computation-in-memory architecture for data-intensive applications. In: 2015 Design, Automation & Test in Europe Conference & Exhibition (DATE), pp. 1718–1725 (2015). IEEEDOI
- [3]Chen, I.-W., Lee, J.: Ultrathin film resistive memory devices. Google Patents. US Patent App. 14/507,957 (2016)
- [4]Wang, Y.-B., Chang, T.-C., Lin, S.-K., Wu, P.-Y., Zhang, Y.-C., Tan, Y.-F., Chen, W.-C., Wu, C.-W., Chou, S.-Y., Zhou, K.-J., et al.: Forming-free hfo 2-based resistive random access memory by x-ray irradiation. IEEE Transactions on Electron Devices 69(12), 6705–6709 (2022)DOI
- [5]Tsai, C.-Y., Ting, Y.-W., Huang, K.-C.: Structure and method for a forming free resistive random access memory with multi-level cell. Google Patents. US Patent 9,053,781 (2015)
- [6]Park, J., Kumar, A., Zhou, Y., Oh, S., Kim, J.-H., Shi, Y., Jain, S., Hota, G., Qiu, E., Nagle, A.L., et al.: Multi-level, forming and filament free, bulk switching trilayer rram for neuromorphic computing at the edge. Nature Communications 15(1), 3492 (2024)DOI
- [7]Chen, Y.-S., Lee, H.-Y., Chen, P.-S., Wu, T.-Y., Wang, C.-C., Tzeng, P.-J., Chen, F., Tsai, M.-J., Lien, C.: An ultrathin forming-free hfox resistance memory with excellent electrical performance. IEEE electron device letters 31(12), 1473–1475 (2010)DOI
- [8]Huang, Y.-J., Pan, S.C., Lee, S.-C.: RRAM devices. Google Patents. US Patent 9,525,008 (2016)
- [9]Stecconi, T., Guido, R., Berchialla, L., La Porta, A., Weiss, J., Popoff, Y., Halter, M., Sousa, M., Horst, F., Dávila, D., et al.: Filamentary taox/hfo2 reram devices for neural networks training with analog in-memory computing. Advanced electronic materials 8(10), 2200448 (2022)DOI
- [10]Wedig, A., Luebben, M., Cho, D.-Y., Moors, M., Skaja, K., Rana, V., Hasegawa, T., Adepadlli, K.K., Yildiz, B., Waser, R., et al.: Nanoscale cation motion in tao x, Nukala, P., Ahmadi, M., Wei, Y., De Graaf, S., Stylianidis, E., Chakrabortty, T., Matzen, S., Zandbergen, H.W., Björling, A., Mannix, D., et al.: Reversible oxygen migration and phase transitions in hafnia-based ferroelectric devices. Science 372(6542), 630–635 (2021)
- [12]Falcone, D.F., Clerico, V., Choi, W., Stecconi, T., Horst, F., Begon-Lours, L., Galetta, M., La Porta, A., Garg, N., Alibart, F., et al.: All-in-one analog ai accelerator: On-chip training and inference with conductive-metal-oxide/hfox reram devices. arXiv preprint arXiv:2502.04524 (2025)DOI
- [13]Beckmann, K., Holt, J., Manem, H., Van Nostrand, J., Cady, N.C.: Nanoscale hafnium oxide rram devices exhibit pulse dependent behavior and multi-level resistance capability. MRS Advances 1(49), 3355–3360 (2016)DOI
- [14]Nakayama, K., Takata, M., Kasai, T., Kitagawa, A., Akita, J.: Pulse number control of electrical resistance for multi-level storage based on phase change. Journal of Physics D: Applied Physics 40(17), 5061 (2007)
- [15]Xu, C., Niu, D., Muralimanohar, N., Jouppi, N.P., Xie, Y.: Understanding the trade-offs in multi-level cell reram memory design. In: Proceedings of the 50th Annual Design Automation Conference, pp. 1–6 (2013)DOI
- [16]Ding, X., Yu, X., Lan, Z., Li, J., Zhou, S., Lee, C., Zhao, Y.: Forming-free hfo x-based resistive memory with improved uniformity achieved by the thermal annealing-induced self-doping of ge. IEEE Transactions on Electron Devices 70(4), 1671–1675 (2023)
- [17]Kim, W., Hardtdegen, A., Rodenbücher, C., Menzel, S., Wouters, D.J., Hoffmann-Eifert, S., Buca, D., Waser, R., Rana, V.: Forming-free metal-oxide reram by oxygen ion implantation process. In: 2016 IEEE International Electron Devices Meeting (IEDM), pp. 4–4 (2016). IEEEDOI
- [18]Wu, E., Ando, T., Jamison, P.: Reversing a decades-long scaling law of dielectric breakdown for reram forming voltage reduction-modeling competition among defect generation and annihilation processes. In: 2023 International Electron Devices Meeting (IEDM), pp. 1–4 (2023). IEEEDOI
- [19]Ikraiam, F.A.: Investigating the effects of ionizing radiation on cv characteristics of soi-mos capacitors. In: Proceedings of the 14th Arabic Conference of the Arab Union for Astronomy and Space Sciences: AUASS-CONF23, 13-16 November 2023, Sharjah, United Arab Emirates, vol. 420, p. 266 (2025). Springer NatureDOI
- [20]Traoré, B., Blaise, P., Vianello, E., Perniola, L., De Salvo, B., Nishi, Y.: Hfo 2-based rram: Electrode effects, ti/hfo 2 interface, charge injection, and oxygen (o) defects diffusion through experiment and ab initio calculations. IEEE Transactions on Electron Devices 63(1), 360–368 (2015)
- [21]Hardtdegen, A., Zhang, H., Hoffmann-Eifert, S.: Tuning the performance of pt/hfo2/ti/pt reram devices obtained from plasma-enhanced atomic layer deposition for hfo2 thin films. ECS Transactions 75(6), 177 (2016)
- [24]Wang, Z., Yang, R., Huang, H.-M., He, H.-K., Shaibo, J., Guo, X.: Electroforming-free artificial synapses based on proton conduction in α-moo3 films. Advanced Electronic Materials 6(4), 1901290 (2020)DOI
- [25]Sun, Y., Ma, Z., Tan, D., Shen, Z., You, J., Li, W., Xu, L., Chen, K., Feng, D.: Hybrid channel induced forming-free performance in nanocrystalline-si: H/a-sinx: H resistive switching memory. Nanotechnology 30(36), 365701 (2019)DOI
- [26]Chakrabarti, B., Galatage, R.V., Vogel, E.M.: Multilevel switching in forming-free resistive memory devices with atomic layer deposited HfTiOₓ nanolaminate. IEEE electron device letters 34(7), 867–869 (2013)DOI
- [27]Kumar, S., Rath, C.: Oxygen vacancy mediated stabilization of cubic phase at room temperature and resistive switching effect in sm-and dy-doped hfo2 thin film. physica status solidi (a) 217(1), 1900756 (2020)DOI
- [28]Capron, N., Broqvist, P., Pasquarello, A.: Migration of oxygen vacancy in hfo2 and across the hfo2/ sio2 interface: A first-principles investigation. Applied Physics Letters 91(19) (2007)DOI
- [29]Mueller, M.P., De Souza, R.A.: Sims study of oxygen diffusion in monoclinic hfo2. Applied Physics Letters 112(5) (2018)DOI
- [30]Clima, S., Chen, Y., Degraeve, R., Mees, M., Sankaran, K., Govoreanu, B., Jurczak, M., De Gendt, S., Pourtois, G.: First-principles simulation of oxygen diffusion in hfox: Role in the resistive switching mechanism. Applied Physics Letters 100(13) (2012)DOI
- [31]Hou, Z., Gong, X., Li, Q.: Al-induced reduction of the oxygen diffusion in hfo2: an ab initio study. Journal of Physics: Condensed Matter 20(13), 135206 (2008)DOI
- [32]Öttking, R., Kupke, S., Nadimi, E., Leitsmann, R., Lazarevic, F., Plänitz, P., Roll, G., Slesazeck, S., Trentzsch, M., Mikolajick, T.: Defect generation and activation processes in hfo2 thin films: Contributions to stress-induced leakage currents. physica status solidi (a) 212(3), 547–553 (2015)
- [33]Foster, A.S., Gejo, F.L., Shluger, A., Nieminen, R.M.: Vacancy and interstitial defects in hafnia. Physical Review B 65(17), 174117 (2002)DOI
- [34]Waser, R., Aono, M.: Nanoionics-based resistive switching memories. Nature materials 6(11), 833–840 (2007)
- [35]Calka, P., Sowinska, M., Bertaud, T., Walczyk, D., Dabrowski, J., Zaumseil, P., Walczyk, C., Gloskowskii, A., Cartoixà, X., Suñé, J., et al.: Engineering of the chemical reactivity of the ti/hfo2 interface for rram: Experiment and theory. ACS applied materials & interfaces 6(7), 5056–5060 (2014)DOI
- [36]Tahini, H.A., Chroneos, A., Middleburgh, S.C., Schwingenschlögl, U., Grimes, R.W.: Ultrafast palladium diffusion in germanium. Journal of Materials Chemistry A 3(7), 3832–3838 (2015)DOI
- [37]Wang, C., Wu, H., Gao, B., Zhang, T., Yang, Y., Qian, H.: Conduction mechanisms, dynamics and stability in rerams. Microelectronic Engineering 187, 121–133 (2018)DOI
- [38]Funck, C., Menzel, S.: Comprehensive model of electron conduction in oxide-based memristive devices. ACS Applied electronic materials 3(9), 3674–3692 (2021)
- [39]Bengel, C., Siemon, A., Cüppers, F., Hoffmann-Eifert, S., Hardtdegen, A., Von Witzleben, M., Hellmich, L., Waser, R., Menzel, S.: Variability-aware modeling of filamentary oxide-based bipolar resistive switching cells using spice level compact models. IEEE Transactions on Circuits and Systems I: Regular Papers 67(12), 4618–4630 (2020)DOI
- [40]Orchard, G., Jawawant, A., Cohen, G.K., Thakor, N.: Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in Neuroscience 9 (2015). Article 437arxiv.org/abs/1507.07629
- [41]Amir, A., Taba, B., Berg, D., Melano, T., McKinstry, J., Nolfo, C.D., Nayak, T., Andreopoulos, A., G´arreau, G., Mendoza, M., Kusnitz, J., Debole, M., Esser, S., Delbruck, T., Flickner, M., Modha, D.: A low power, fully event-based gesture recognition system. In: IEEE Conference on Computer Vision and Pattern Recognition (2017)DOI
- [42]Shrestha, S.B., Orchard, G.: SLAYER: Spike layer error reassignment in time. In: Advances in Neural Information Processing Systems, pp. 1412–1421 (2018)arxiv.org/abs/1810.08646
- [43]Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gिमelshéin, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems 32, pp. 8024–8035 (2019)arxiv.org/abs/1912.01703
- [44]Gerstner, W.: Time structure of the activity in neural network models. Phys. Rev. E 51, 738–758 (1995)pubmed.ncbi.nlm.nih.gov/9962697
- [45]Shen, X., Liu, W., Gao, X., Lu, Z., Wu, X., Gao, X.: Mechanisms of oxidase and superoxide dismutation-like activities of gold, silver, platinum, and palladium, and their alloys: a general way to the activation of molecular oxygen. Journal of the American Chemical Society 137(50), 15882–15891 (2015)DOI
- [46]Tse, K.-Y., Robertson, J.: Control of schottky barrier heights on high-k gate dielectrics for future complementary metal-oxide semiconductor devices. Physical review letters 99(8), 086805 (2007)pubmed.ncbi.nlm.nih.gov/17930973
- [47]Tung, R.T.: Chemical bonding and fermi level pinning at metal-semiconductor interfaces. Physical review letters 84(26), 6078 (2000)pubmed.ncbi.nlm.nih.gov/10991128
- [48]Hua, E., Choi, S., Ren, S., Kim, S., Ali, G., Kim, S.J., Jang, W.-S., Joo, S., Zhang, J., Ji, S., et al.: Negatively charged platinum nanoparticles on dititanium dioxide electr...
- [49]Hu, R., Li, X., Tang, J., Li, Y., Zheng, X., Gao, B., Qian, H., Wu, H.: Investigation of resistive switching mechanisms in ti/tiox/pd-based rram devices. Advanced Electronic Materials 8(8), 2100827 (2022)
- [50]Do, H.B., Luc, Q.H., Ha, M.T.H., Huynh, S.H., Nguyen, T.A., Lin, Y.C., Chang, E.Y.: Study of the interface stability of the metal (mo, ni, pd)/hfo2/aln/ingaas mos devices. AIP Advances 7(8) (2017)
- [51]Yamashita, D., Ishizaki, A.: In situ measurements of change in work function of pt, pd and au surfaces during desorption of oxygen by using photoemission yield spectrometer in air. Applied Surface Science 363, 240–244 (2016)DOI
- [52]Giner, J., Lange, E.: Elektronenau...spannungen von reinem und sauerstoffbedecktem au, pt und pd auf grund ihrer voltaspannungen gegen ag. Naturwissenschaften 40(19), 506–506 (1953)DOI
- [53]Jaeckel, R., Wagner, B.: Photo-electric measurement of the work function of metals and its alteration after gas adsorption. Vacuum 13(12), 509–511 (1963)DOI
- [54]Gu, D., Dey, S.K., Majhi, P.: Effective work function of pt, pd, and re on atomic layer deposited hfo2. Applied Physics Letters 89(8) (2006)DOI