Morphology analysis of the reconstruction of nano-filaments during aborted reset operation in Cu-based ReRAM devices Authors: Shehla Yasmeen, Aaron DiFilippo, Marius Orlowski Published: 2025 ABSTRACT This paper presents a detailed physical analysis of the counterintuitive strengthening effect observed in Cu/TaOₓ/Pt-based resistive random-access memory (ReRAM) devices during an aborted reset operation. While our prior work [26] announced this phenomenon in a short letter, here, we develop a comprehensive electro-thermal model that links the observed resistance reduction directly to an atomic reconstruction of the conductive filament (CF). We demonstrate that the transformation from a truncated conical filament to a lower-resistance, hourglass-shaped structure is driven by the synergistic action of intense electric fields and localized Joule heating, which are both highly confined to the filament’s tip. This reconstruction of CF is governed by the fundamental physics of Cu atom diffusivity and ion electromigration within the TaOₓ matrix. Our model, constrained by volume conservation and supported by electrostatic and thermal simulations, successfully optimizes the filament geometry to predict a maximum resistance reduction factor of ~ 11, which aligns precisely with experimental data for high-resistance filaments. Furthermore, we identify a critical two-phase reset process: a reinforcement phase dominated by electromigration and a final rupture phase. The experimental dependence on voltage ramp rate is quantitatively explained by calculating the timescales required for Cu ion mobility, conclusively showing that slower ramps allow sufficient time for atomic rearrangement. This detailed analysis resolves the controversy surrounding filament morphology and provides profound insights into CF dynamics that are critical for advancing the reliability and design of ReRAM devices, particularly for neuromorphic computing applications. Introduction Resistive Random Access Memory (ReRAM) has emerged as a pivotal technology for next-generation non-volatile memory and neuromorphic computing, offering excellent scalability, a simple metal–insulator-metal (MIM) structure, and low power consumption1–3. Resistive RAM technologies are primarily categorized by their physical switching mechanism. The two main classes of ReRAM are filamentary and non-filamentary (or interfacial). Filamentary types are further divided in: (i) Electrochemical Metallization (ECM) Memory; this type relies on an electrochemically active electrode (e.g., Cu, Ag) that provides metal cations to form a conductive filament within an insulating solid electrolyte (e.g., TaOₓ, SiO₂). The switching is bipolar and the filament’s formation (SET) and dissolution (RESET) are governed by electrochemical processes. In this work our analysis is based on Cu/TaOx/Pt devices belonging to the ECM category.(ii) The Valence Change Mechanism (VCM) type Memory uses an electrochemically active electrode (e.g., TiN, Pt) and a metal oxide where oxygen vacancies are the mobile species (e.g., HfO₂, TaOₓ). The applied field drives the migration of oxygen anions/vacancies, changing the local stoichiometry and creating a highly conductive, oxygen-deficient filament. This is also a bipolar switching mechanism. An example of non-filamentary memory is the Phase-Change Memory (PCM) which operates on a fundamentally different principle. It uses a chalcogenide material (e.g., Ge₂Sb₂Te₅) that is switched between a crystalline (low-resistance) and an amorphous (high-resistance) state through the application of Joule heat (a melt-quench process for RESET, and an anneal for SET), not through the formation of a nanoscale metallic filament. Other notable types include Thermochemical Memory (TCM), driven purely by Joule heating, and non-filamentary ReRAM, where the resistance change occurs uniformly across an interface rather than being localized to a single filament. In filamentary-type ReRAM cells, such as the Cu/TaOₓ/Pt system, the fundamental switching mechanism involves the formation and rupture of a conductive nanofilament (CF) via electrochemical metallization and thermochemical processes4–6. The SET operation (transition to a Low-Resistance State, LRS) is characterized by the field-driven growth of a metallic filament, while the RESET operation (transition to a High-Resistance State, HRS) is dominated by Joule-heat-induced dissolution of the filament at a critical current, Ireset7–11. While numerous excellent studies have elucidated the mechanisms of filament rupture12–25, the intricate morphological evolution of the filament during the aborted reset operation—particularly the counterintuitive phenomenon of resistive strengthening—has remained unexplored. This effect is most manifest during an aborted reset operation 26. The preliminary reinforcement phase actively preconditions the final morphology of the ruptured filament, dictating the geometry of the gap. This, in turn, establishes the initial conditions for the subsequent SET operation, inevitably influencing the stochasticity and energy requirements of the next filament formation cycle. This is particularly true for incomplete or aborted reset operations26, where the applied voltage bias is halted near, but before reaching, the full rupture threshold. Such operations can lead to counterintuitive phenomena, including a robust reduction in device resistance, that challenge the conventional binary view of filament dynamics. The ability to precisely control and modulate the conductive filament’s resistance is not merely an academic curiosity; it is the very foundation of ReRAM’s potential in neuromorphic computing27–29. In neuromorphic systems, devices must act as artificial synapses, capable of exhibiting a continuous range of stable conductance states—or “weights”—that can be incrementally potentiated or depressed. Standard ReRAM switching, however, often tends towards binary, stochastic behavior, which is a significant hurdle for reliable analog computing. This work moves beyond the initial announcement of the resistive strengthening effect26 to provide a detailed physical and mathematical framework explaining the fundamental mechanisms behind this phenomenon during aborted reset operations in Cu/TaOₓ/Pt devices. We establish a comprehensive model that successfully bridges the gap between experimental electrical characteristics and the nanoscale electro-thermal atomic processes governing filament dynamics. We demonstrate that the observed reduction in resistance is a deterministic outcome of a fundamental morphological reconstruction of the filament, driven by the confluence of extreme electric fields and localized Joule heating at the filament tip. A critical insight from our study is the introduction of a two-phase mechanism for the reset operation: an initial reinforcement phase governed by electromigration and atomic diffusivity, followed by the final rupture phase. We show experimentally and theoretically that the voltage ramp rate is the decisive parameter controlling which phase dominates, thus determining the final filament morphology and device resistance. This new paradigm resolves long-standing debates about filament evolution and reveals that the reset outcome is a complex competition between reinforcement and rupture. Our findings have significant implications for device engineering, providing a novel methodology to control filament morphology through operational parameters. Most importantly, the ability to intentionally strengthen a filament presents a intriguing novel functionality that could be directly exploited in multi-state memory and neuromorphic computing systems, where precise and stable control of synaptic weights is paramount. Experimental Cu/TaOₓ/Pt resistive random-access memory (ReRAM) cell arrays were fabricated in a crossbar configuration on thermally oxidized silicon substrates. The structure began with a 300 nm SiO₂ layer grown on a silicon wafer 2, followed by the sequential electron-beam physical vapor deposition (PVD) of Ti (20 nm), Pt (50 nm), and Cu (150 nm) thin films. The Pt electrode lines were patterned using a lift-off process with a 2 μm-thick photoresist layer to achieve gently sloped sidewall profiles. Between the definition of Pt and Cu electrodes, a 25 nm-thick, oxygen-deficient TaOₓ layer was deposited in a blanket fashion by evaporating TaOₓ pellets without introducing oxygen gas into the PVD chamber. Further fabrication and electrical characterization details are available in references30–32. The current–voltage (I-V) characteristics were measured at room temperature using a Keithley 4200-SCS parameter analyzer. The fundamental resistive switching mechanism arises from the formation and rupture of a copper conductive filament (CF). Figure 1a shows typical initial six set–reset I–V switching sequences of the Cu/TaOx/Pt devices. Figure 1b presents a schematic of the conventional reset operation (black) and an aborted reset operation (red), together with the key operational parameters. During the set process, a positive voltage sweep induces a sudden current increase at the set voltage (Vset), with the current limited by a compliance current, Icc(+), typically between 10 and 300 μA. The subsequent reverse sweep displays ohmic behavior, characterized by the low-resistance state (LRS), Ron(+). To reset the device, a negative voltage is applied. The current remains ohmic until it reaches a critical reset current (Ireset; typically 1–6 mA), at which point it collapses at the reset voltage (Vreset), returning the device to a high-resistance state (HRS). Under optimal conditions, devices can switch repeatedly without significant degradation.Fig. 1(a) Typical first 6 set–reset I–V switching sequences of a Cu/TaOx/Pt device. (b) Schematic of conventional reset (black) and aborted reset (red) operations. (a) Typical first 6 set–reset I–V switching sequences of a Cu/TaOx/Pt device. (b) Schematic of conventional reset (black) and aborted reset (red) operations. A so-called "aborted or constrained reset" operation is shown in red in Fig. 1b. This process mirrors a standard reset but uses a compliance current, Icc(-), set somewhat lower than the typical Ireset. The compliance current prevents sufficient Joule heating from accumulating to rupture the filament. We denote the resistance measured after this aborted reset operation as Ron(ar) , while Ron is the low-resistance state of the filament directly after a standard SET operation, measured before any aborted reset is attempted. A surprising finding is that for filaments with a high initial resistance (Ron shown in Fig. 1b), the value of Ron(ar) can be up to an order of magnitude smaller, demonstrating that the aborted reset operation significantly reinforces, rather than weakens, the filament. The resistance Ron(ar) is not visible in the I-V sweep of Fig. 1b because the compliance current, Icc, masks the true resistance value. Therefore, Ron(ar) is ascertained independently by measuring the current at a low read voltage of 0.05 V after the aborted reset operation is complete. In contrast, for low-resistance filaments, the ratio Ron/Ron(ar) remains unity, with no observable reinforcement or weakening. This distinction is evident in the experimental data presented in Table 1, which shows resistance values before (Ron) and after (Ron(ar)) the aborted reset operation, along with the calculated reduction factor. The key parameters in Tables 1 and 2 exhibit a tight distribution, reflected also in the error bars of Fig. 2, with a standard deviation (σ) of less than 30 Ω for the critical resistance states. This low device-to-device and cycle-to-cycle variability strongly attests to the reliability and reproducibility of the reported phenomenon.”Table 1presents the resistance reinforcement results for filaments with initial resistances (Ron) ranging from 107 Ω to 3850 Ω following an aborted reset operation. The table lists the resistance after the operation (Ron(ar)) and the reinforcement ratio, Ron/Ron(ar). The first row indicates the level of compliance current (Icc) applied during the SET operation to form the initial filament with resistance values given in the second row and denoted by Ron.ICC[µA]1500100060050030015050Ron[Ω]1072103294355481,2803,850Ron(ar)[ Ω]107209328430520352342Ron/Ron(ar)1.001.001.001.011.063.6411.25Table 2The effect of voltage ramp rate (rr) during the aborted reset operation on Ron. It shows the resulting Ron(ar) values for highly resistive filaments with similar Ron(-) values subjected to the same operation at three different ramp rates, rr. Time t1V (1V/rr) denotes the time to sweep an interval of 1V around the original Vreset value.rr [V/s]0.030.402.00\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{on}$$\end{document}[Ω]128012171327\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{on}^{(ar)}$$\end{document}[Ω]3528411321t1V [s]33.32.50.5Fig. 2Dependence of on-resistance Ron on compliance current Icc applied during the SET operation. Table 1presents the resistance reinforcement results for filaments with initial resistances (Ron) ranging from 107 Ω to 3850 Ω following an aborted reset operation. The table lists the resistance after the operation (Ron(ar)) and the reinforcement ratio, Ron/Ron(ar). The first row indicates the level of compliance current (Icc) applied during the SET operation to form the initial filament with resistance values given in the second row and denoted by Ron.ICC[µA]1500100060050030015050Ron[Ω]1072103294355481,2803,850Ron(ar)[ Ω]107209328430520352342Ron/Ron(ar)1.001.001.001.011.063.6411.25 presents the resistance reinforcement results for filaments with initial resistances (Ron) ranging from 107 Ω to 3850 Ω following an aborted reset operation. The table lists the resistance after the operation (Ron(ar)) and the reinforcement ratio, Ron/Ron(ar). The first row indicates the level of compliance current (Icc) applied during the SET operation to form the initial filament with resistance values given in the second row and denoted by Ron. Table 2The effect of voltage ramp rate (rr) during the aborted reset operation on Ron. It shows the resulting Ron(ar) values for highly resistive filaments with similar Ron(-) values subjected to the same operation at three different ramp rates, rr. Time t1V (1V/rr) denotes the time to sweep an interval of 1V around the original Vreset value.rr [V/s]0.030.402.00\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{on}$$\end{document}[Ω]128012171327\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{on}^{(ar)}$$\end{document}[Ω]3528411321t1V [s]33.32.50.5 The effect of voltage ramp rate (rr) during the aborted reset operation on Ron. It shows the resulting Ron(ar) values for highly resistive filaments with similar Ron(-) values subjected to the same operation at three different ramp rates, rr. Time t1V (1V/rr) denotes the time to sweep an interval of 1V around the original Vreset value. Dependence of on-resistance Ron on compliance current Icc applied during the SET operation. The data confirms that the reinforcement effect is pronounced for fragile, high-resistance filaments but diminishes for robust, low-resistance ones. Furthermore, even for very high-resistance filaments, the reduction factor seldom exceeds a value of 10. The filaments with various resistance values have been formed during a SET operation according to the relation2,33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{on}=K/{I}_{cc}^{n},$$\end{document}where for our devices K = 0.17 V and n = 0.996 have been obtained by fitting experimental data as shown in Fig. 2 Conventional models attribute the reset operation to a combination of Joule-heating-assisted Cu atom thermodiffusion and electromigration of Cu ions, leading to filament rupture at a constriction point at critical Vreset and Ireset values. This rupture creates a nanoscale gap, causing the current to collapse to a nA-range values. The observed reinforcement during an aborted reset has been hypothesized to result from an atomic reconstruction of the filament, where high Cu atom diffusivity and strong electromigration under high electric fields alter its structure13,15. The magnitude of this effect depends on the compliance current Icc(-), Vstop, and voltage ramp rate used during the aborted operation. Electrostatic simulations suggest that the high electric fields present in high-resistance filaments redistribute copper atoms, broadening the filament’s tip at the Cu electrode interface and transforming its shape from a cone to a more stable, low-resistance hourglass structure. This reconstruction reduces the local electric field, halting the process. It is worth noting that the observed strengthening behavior is a local filamentary effect and is therefore independent of the device’s lateral dimensions. This is because the nanoscale conductive filament, whose morphology dictates the resistance, is several orders of magnitude smaller than the overall cell area. We have experimentally confirmed this by obtaining consistent results across devices with vastly different electrode dimensions, including 5 × 5 µm2 cells, 10 × 10 µm2 cells, and 35 × 35 µm2 cells. These findings provide new insights into filament morphology and reveal a complex dynamic where the reset outcome is determined by a competition between filament reinforcement and rupture. This paper analyzes the mechanisms and effects of this atomic reconstruction in high-resistance filaments, comparing results with experimental data and linking them to the fundamental processes of Cu diffusion and electromigration in dielectric materials. In summary, the phenomenon reported here is specific to filamentary ReRAM cells and is not applicable to non-filamentary resistive switching devices, such as phase-change memory (PCM). We have experimentally demonstrated this effect in a Cu/TaOₓ/Pt cell as a representative example of an Electrochemical Metallization (ECM) device. Furthermore, we are confident that the underlying mechanism likely extends to vacancy-controlled filaments in Valence Change Mechanism (VCM)-type ReRAM. Therefore, we propose that this effect is a general characteristic of filamentary resistive switching in metal-oxide systems. Theoretical models The reconfiguration of the filament is governed by two key elements: the specific distribution of electric field and temperature within and around a highly resistive filament. This filament is geometrically represented as a truncated cone34 with a height, H, equal to the thickness of the TaOₓ layer, a base radius R, and a much smaller top radius r (r < < R). The resistance of the conductive filament is intrinsically linked to its truncated cone geometry. As defined by Eq. (7), the resistance is proportional to the filament’s height (H) and inversely proportional to its top (r) and bottom (R) radii. Using an established Cu filament resistivity of 4 × 10⁻4 Ω·cm35 and holding the bottom radius and height constant at R = 30 nm and H = 25 nm, we can calculate how the top radius alone dictates the resistance. A moderately narrow cone with top radius r = 5 nm yields a resistance of ~ 666 Ω, a narrow cone with r = 1 nm yields ~ 3332 Ω, and a wide cone or a cylindrical filament where r = R = 30 nm yields ~ 111 Ω. This calculated range from 111 to 3332 Ω, achieved by varying only the top radius, aligns precisely with the experimental values reported in Table 1, validating the model’s geometric basis. The tip of this cone is a critical region where the electric fields are most intense. Figure 3 shows explicit MATLAB simulations of the potential and electric field distributions for this geometry. For an applied voltage of 1 V—which is within the typical range of observed Vreset values—the maximum electric field near the cone’s tip reaches approximately 1.6 × 10⁶ V/cm. These field strengths scale linearly with the applied bias.Fig. 3MATLAB simulation of the electric field distribution between the cone tip and the Cu electrode, with the Cu electrode biased at –1 V and the Pt electrode grounded. The filament has a total height of 25 nm, with the tip height being 8 nm. MATLAB simulation of the electric field distribution between the cone tip and the Cu electrode, with the Cu electrode biased at –1 V and the Pt electrode grounded. The filament has a total height of 25 nm, with the tip height being 8 nm. The second critical factor driving filament reconfiguration is temperature. As reported in several publications10,11,14, the local temperature within the filament can reach 1000 °C, sometimes even causing partial melting of the Cu electrode. To model the temperature gradient along the central axis of the cone (inset, Fig. 3), we assume a temperature of 980 °C at the tip. The base of the Cu filament is anchored to the cooler Pt electrode, which reflectance measurements suggest is maintained at about 60 °C. We solve for the steady-state heat flow using the thermal conductivity, k, of the filament material. The heat flow Q in the cone10,11 is given then by1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q=-kA\left(DT/dz\right)$$\end{document}where A(z) is the cross-sectional area of the cone at height z. Integrating from the base to tip of the cone gives the temperature distribution :2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T(z)=T\left(0\right)+[(T\left(H\right)-T\left(0\right)]\bullet {\int }_{0}^{z}\frac{ds}{A\left(s\right)}/{\int }_{0}^{H}\frac{dz}{A(z)}$$\end{document} For a cone with dimensions H = 25 nm, r = 5 nm, and R = 30 nm, the resulting temperature distribution T(z) with the boundary conditions T(0) = 60 °C and T(25) = 980 °C3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T\left(z\right)=60+184z/(30-z)$$\end{document}is depicted in Fig. 4. The solution shows that temperatures exceeding 700 °C are confined to a very small region at the tip, extending only 4–5 nm. The temperature drops rapidly away from this hotspot, falling to near-ambient levels in the lower portion of the cone.Fig. 4Temperature distribution along the cone height, showing a peak temperature of 980 °C at the cone tip. Temperature distribution along the cone height, showing a peak temperature of 980 °C at the cone tip. In summary, two simultaneous conditions are necessary for filament reconfiguration:High temperature is required to enhance the diffusion of Cu atoms and to promote the redox reaction (Cu ⇌ Cu⁺ + e⁻).A strong electric field is needed to drive the electromigration of ions, facilitating the structural rearrangement. High temperature is required to enhance the diffusion of Cu atoms and to promote the redox reaction (Cu ⇌ Cu⁺ + e⁻). High temperature is required to enhance the diffusion of Cu atoms and to promote the redox reaction (Cu ⇌ Cu⁺ + e⁻). A strong electric field is needed to drive the electromigration of ions, facilitating the structural rearrangement. A strong electric field is needed to drive the electromigration of ions, facilitating the structural rearrangement. Both the highest temperatures and the strongest electric fields are localized to the tip of the conical filament. Figure 5 illustrates how these two effects work in tandem to reconfigure the filament, transforming an initial sharp, truncated cone into a more stable double-cone or hourglass shape through an aborted reset operation.Fig. 5Reconstruction of the cone. (a) Initial filament shape, approximated as a sharply truncated cone in our analysis. (b) Electromigration of Cu⁺ ions near the hot spot under high electric fields at the tip. (c) Final filament geometry resembling a double cone (hourglass shape). Reconstruction of the cone. (a) Initial filament shape, approximated as a sharply truncated cone in our analysis. (b) Electromigration of Cu⁺ ions near the hot spot under high electric fields at the tip. (c) Final filament geometry resembling a double cone (hourglass shape). In contrast, a low-resistance filament of a truncated cone with r \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\approx R$$\end{document} i.e. a shape that approximates a cylinder, does not generate such high fields or a significant temperature hotspot at its tip. Consequently, as explained in prior work26, there is no driving force for reconfiguration in these structures. This paper focuses specifically on elucidating the mechanisms behind the reconfiguration of highly resistive, conical Cu filaments. To facilitate further analysis, we propose a simplified model to approximate the potential drop and electric field between the conical filament tip and the Cu electrode. The potential distribution along the cone can be derived from the voltage division principle, where the potential at a given height z is equal to the resistance of the cone segment at z divided by the total cone resistance, multiplied by the applied voltage. This yields the function V(z), where the coordinate z = 0 is defined at the cone’s tip. The resulting equation for the potential on the cone’s surface at height z is given by:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V\left(z\right)={V}_{o}\frac{r}{H}\cdot (H-z)/\left(r+\frac{R-r}{H}z\right)$$\end{document}where Vo = 1 corresponds to the negative bias applied to the Cu electrode, while the Pt electrode remains grounded. Notably, this potential distribution along the cone’s surface is independent of the resistivity ρ of the cone and depends solely on the geometry of the filament. An example of this potential distribution is shown in Fig. 6 for a cone with dimensions H = 25 nm, R = 40 nm, and r = 2 nm. The figure also illustrates the electric field in the TaOₓ gap between the filament tip and the Cu electrode. This field was calculated by modeling the gap as the arc to the base of an isosceles triangle, with the tip-to-Cu electrode distance defining its height. The simulation shows that the electric field peaks at approximately 1.6 × 106 V/cm.Fig. 6Potential and electric field distribution along the cone surface using the approximation in Eq. (4), with the Cu electrode biased at − 1 V and the Pt electrode grounded. The approximation shows excellent agreement with full MATLAB simulations (Fig. 2). Potential and electric field distribution along the cone surface using the approximation in Eq. (4), with the Cu electrode biased at − 1 V and the Pt electrode grounded. The approximation shows excellent agreement with full MATLAB simulations (Fig. 2). These simplified approximations have been validated against explicit numerical MATLAB calculations (Fig. 6) and confirmed to be highly accurate. Their primary advantage is that they allow for the direct calculation of Cu ion electromigration forces without the need to extract field data from specialized MATLAB simulation software. Optimal reconfiguration of the filament This section analyzes the transformation of the filament from a truncated cone into a double cone with the goal of determining the lowest possible resistance of the reconfigured structure. Since this reconfiguration involves only the redistribution of intrinsic Cu atoms of the filament, it must occur under the constraint of volume conservation. To this end, we require formulae for the volume and resistance of both structures. Here, Vtc and Vdc denote the volume of the single truncated (tc) cone and the double cone (dc), respectively.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{tc}=\frac{\pi }{3}H({R}^{2}+Rr+{r}^{2})$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{dc}=\frac{\pi }{3}[\left(H-x\right)\left({R}^{2}+Rr+{r}^{2}\right)+x\left({y}^{2}+Rr+{r}^{2}\right)]$$\end{document} The corresponding equations for the resistance of both cones are given by Rtc and Rdc. The parameters for both geometries are visualized in Fig. 7.Fig. 7Geometry of a single truncated cone and a double truncated cone, with parameters used for resistance minimization under the constraint of volume preservation.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{tc}=\frac{\rho H}{\pi Rr}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{dc}=\frac{\rho }{\pi }\left[\frac{H-x}{Ry}+\frac{x}{zy}\right]$$\end{document} Geometry of a single truncated cone and a double truncated cone, with parameters used for resistance minimization under the constraint of volume preservation. The optimization problem is to find the parameters x, y, and z such that the resistance Rdc is minimized under the constraint that the volume of the original and reconfigured filament are identical. It should be noted that this process performs a mapping of the experimental values Ron and Ron(ar) on(+) and on(-) onto their geometrical representations, Rtc and Rdc, respectively. If no constraints are placed on the parameters x, y, and z, the optimization becomes trivial and yields the intuitively expected result: x = 0 and y = z = c, forming a cylinder, which would have the lowest possible resistance for the volume. The volume, Vcyl, and resistance, Rcyl, of such a cylinder are given as:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{cyl}=H\pi {c}^{2} - with - c=\frac{{R}^{2}+Rr+{r}^{2}}{3}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{cyl}=\frac{\rho H}{\pi {c}^{2}}$$\end{document} It is instructive to observe the resistance ratio of the two volume-equivalent filaments as a function of r, the smaller radius of the original truncated cone. This function is given as:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{tc}/{R}_{cyl}=3\left[\frac{R}{r}+\frac{r}{R}+1\right]=3\left[\frac{1}{\zeta }+\zeta +1\right] - with- \zeta =r/R$$\end{document} This resistance ratio is plotted as a function of ζ. As r varies between 0 and R, ζ ranges from 0 to 1. If ζ = 1, the original truncated cone is a cylinder and the resistance ratio is 1. The resistance ratio as a function of ζ is plotted in Fig. 8. It can be seen that the ratio is high only within a very narrow interval of ζ, approximately [0, 0.15]. For larger ζ values, the ratio is very close to unity, indicating no significant resistance reduction.Fig. 8Resistance ratio of a truncated cone to a cylinder of equal volume, plotted as a function of ζ = r/R (the ratio of the cone’s top radius to bottom radius). Resistance ratio of a truncated cone to a cylinder of equal volume, plotted as a function of ζ = r/R (the ratio of the cone’s top radius to bottom radius). A transformation from a truncated cone into a cylinder would require that all Cu atoms in the filament are free to move via electromigration throughout the TaOx matrix. However, as established in earlier discussions, the electric field is strong only in the top part of the frustum (see Figs. 2and 6). Simultaneously, the temperature due to Joule heating is also high only in this same top segment of the cone. Therefore, it is clear that the bulk lower part of the truncated cone will remain unaffected by the reconstruction process. We can thus assume that only the upper part undergoes reconstruction. From our calculations of the electric field and temperature distributions, we estimate that only a small upper part of the cone, with a height of approximately 0.15H, is affected by the reconstruction, while the bulk lower part of height 0.85H remains unchanged. If we freeze the lower part, the optimization is subject to the following volume conservation condition, where the parameter x from eq.(5) and eq.(6) has been eliminated:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.85\left({R}^{2}+Ry+{y}^{2}\right)+0.15\left({y}^{2}+zy+{z}^{2}\right)={R}^{2}+Rr+{r}^{2}$$\end{document} The resistance of this double cone can now be written as:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{dc}=\frac{\rho H}{\pi }\left[\frac{0.85}{Ry}+\frac{0.15}{zy}\right]$$\end{document} Since we are interested in highly resistive filaments with a small radius at the top, we can safely assume r < < R, so that R2 + Rr + r2 ≈ R2. The volume conservation equation can then be rewritten as:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.85\left({R}^{2}+Ry+{y}^{2}\right)+0.15\left({y}^{2}+zy+{z}^{2}\right)={R}^{2}$$\end{document} Let us define the normalized variables u = y/R and v = z/R. The previous equation then transforms into. 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u^2+0.15uv+0.15v^2+0.85u-0.15=0$$\end{document} and the resistance of the double cone structure becomes:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{dc}=\frac{\rho H}{\pi {R}^{2}}\left[\frac{0.85}{u}+\frac{0.15}{uv}\right]=\frac{\rho H}{\pi {R}^{2}}F\left(u,v\right)$$\end{document} We wish to minimize Rdc subject to volume conservation, which is equivalent to minimizing the function F(u,v):17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F\left(u,v\right)=\left[\frac{0.85}{u}+\frac{0.15}{uv}\right]$$\end{document} Using the method of Lagrange method with Lagrange multiplier λ, we define the Lagrange function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{L}\left(u,v,\lambda \right)$$\end{document}:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{L}\left(u,v,\lambda \right)=F\left(u,v\right)-\lambda ({u}^{2}+0.15uv+0.15{v}^{2}+0.85u-0.15)$$\end{document} Taking partial derivatives and setting them to zero yields the condition:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u\left(0.85+2u+0.15v\right)=v\left(0.85v+0.15\right)\left(u+2v\right)$$\end{document} Solving Eqs. (15) and (19) numerically yields u ≈ 0.126 and v ≈ 0.367, or y ≈ 0.126R and z ≈ 0.367R. This choice of parameters gives the lowest possible resistance for the reconfigured filament under the volume conservation constraint for the initial truncated cone. Discussion and results Interestingly, for this optimal parameter set, the reconfigured cone also possesses the lowest electric fields along its surface, as shown in Fig. 9a. This is contrasted with a less optimal double cone choice (Fig. 9b that displays much higher electric fields. This result is intuitively clear because the optimal double cone structure has not only the smallest total resistance but also the smallest local resistance, which minimizes the electric fields along its walls. Thus, the shape transformation will continue as long as Cu ions are mobile in the upper part, stopping only when the lowest resistance and lowest electric field density are achieved.Fig. 9Comparison of surface electric fields for different filament geometries. (a) The optimized shape minimizes electrical resistance, resulting in a peak electric field of 0.11 × 10⁶ V/cm. (b) In contrast, a non-optimized double-cone geometry exhibits higher resistance and a consequently higher electric field of 0.6 × 10⁶ V/cm. Comparison of surface electric fields for different filament geometries. (a) The optimized shape minimizes electrical resistance, resulting in a peak electric field of 0.11 × 10⁶ V/cm. (b) In contrast, a non-optimized double-cone geometry exhibits higher resistance and a consequently higher electric field of 0.6 × 10⁶ V/cm. Assuming R = 70 nm, the ratio of the initial resistance of the Cu CF to that of the optimally reconfigured CF is plotted as a function of the radius r at the top of the original CF in Fig. 10. It can be seen that for highly resistive CFs with a sharp tip (small r), the reduction factor can exceed 11. However, this factor decreases rapidly with only a slight increase in r; for r > 4 nm, the reduction factor is essentially unity, meaning no resistance reduction. Notably, the model’s predicted range for the resistance reduction factor is between 1 and 11, which aligns essentially perfectly with the experimental range of 1 to 11.3 as shown in Table 1.Fig. 10The factor of improvement in resistance achieved by the optimized filament design. The ratio is calculated as the resistance of the baseline truncated cone structure Rtc corresponding to Ron(ar) on(-) divided by the significantly lower resistance of the reconstructed, shape-optimized filament and Rdcopt corresponding to Ron(+). The factor of improvement in resistance achieved by the optimized filament design. The ratio is calculated as the resistance of the baseline truncated cone structure Rtc corresponding to Ron(ar) on(-) divided by the significantly lower resistance of the reconstructed, shape-optimized filament and Rdcopt corresponding to Ron(+). Thus, we see that the reset operation can, in principle, occur in two phases. The first phase of filament reinforcement happens when the current is near but below the Ireset value. During this phase, the temperature and electric fields is high enough to cause Cu atom displacement, leading to filament reconfiguration if sufficient time is allowed. The second phase is the rupture itself, which occurs when the temperature is so high that Cu atoms diffuse out of the filament at the hottest spot (i.e., near the constriction), creating a permanent gap. The first phase can be probed experimentally with an aborted reset operation at a high Icc(-) < Ireset , e.g. around 0.8 Ireset. If the voltage ramp rate is slow (rr = 0.03 V/s, as used for the experiments in Table 1), then the time near Icc(-) = 0.8Ireset is sufficient for Cu ions to electromigrate and effect filament reconfiguration. Here, Vstop can be slightly larger than Vreset. However, if the same operation is performed with a fast ramp rate (rr = 2.0 V/s), there is insufficient time for significant Cu ion displacement, and reconfiguration does not occur. This is borne out by the results of experiments shown in Table 2 where the strengthening effect is shown as a function of ramp rate. From the reset curves, the voltage interval from the point where the current is 0.7.Ireset < to a Vstop of -3 V is about 1 V. From this data we can assess the time t1V available for Cu ion displacement when the measurement sweeps an interval of 1 V; this time is therefore approximately 1 V/rr. These values for the resistance Ron(ar) on(-) and Ron on(+) are given in Table 2. Here, t1V is the time to traverse a voltage interval of 1 V when the current during the aborted reset operation is highest. This time must be compared with an estimate of the time required for a Cu ion to move under high electric fields. For that, we need an estimate of the Cu ion mobility μ, which is related to the diffusivity D by the Einstein relation: μ = D / (kT/q). The Cu diffusivity in dielectric materials can be found in the literature36,37 and is given by D = 2.1 × 10–10 exp(-0.98 eV/kT), with an activation energy Ea = 0.98 eV, which falls within the reported range of 0.9 eV to 1.1 eV. We now calculate the time required for a Cu ion to travel a distance s in a field E = 1 × 106 V/cm at a temperature near 980 °C. The equation for the time ts o traverse a distance s is given below:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{s}=\frac{s}{v}=\frac{s}{\mu \cdot E}=\frac{s\cdot (kT/q)}{D\cdot E}$$\end{document} For the reconstruction at the tip, a Cu ion needs to travel approximately 5 nm for effective reconstruction. Using the estimated diffusivity for Cu in TaOx36,37 of D = 2.1 × 10-10exp(-0.98 ev/kT) at T = 980 + 300 = 1280 K, this time is calculated to be ts ≈ 2.6 s. It is then clear from Table 2 that the slow ramp rate offers a time window of 33.3 s, which amply accommodates the migration of Cu ions. In contrast, the fast ramp rate narrows the time window to 0.5 s, which is clearly insufficient for significant Cu⁺ ion movement. The intermediate ramp rate of 0.4 V/s offers a time window of 2.5 s, which can accommodate the movement of only a portion of ions, resulting in only partial reconfiguration of the filament. This dynamics of Cu electromigration is bound to affect the final morphology of the Cu filament. If the reset is performed at a fast ramp rate (e.g., rr = 2 V/s), the strengthening phase is skipped due to insufficient time for atomic rearrangement, and the ruptured filament may resemble the structure in Fig. 11a. If the ramp rate is slow, the filament undergoes a strengthening phase before rupture; this type of reset is likely to result in a ruptured filament shape similar to that in Fig. 11b. Consequently, the type of reset operation affects the morphology of the ruptured filament, which in turn influences the properties of the filament formed in the subsequent set operation. This behavior and its understanding may be useful for neuromorphic applications employing such devices.Fig. 11Filament’s schematic morphology after a reset operation using (a) a fast ramp rate compared to (b) a slow ramp rate. Filament’s schematic morphology after a reset operation using (a) a fast ramp rate compared to (b) a slow ramp rate. Conclusion In conclusion, this work provides a detailed physical and mathematical framework to explain the fundamental mechanisms behind the resistive strengthening phenomenon observed in Cu/TaOₓ/Pt ReRAM devices during aborted reset operations. Moving beyond the initial announcement of the effect, we have established a comprehensive model that successfully bridges the gap between experimental electrical characteristics and the nanoscale electro-thermal atomic processes governing filament dynamics. Our analysis demonstrates that the counterintuitive reduction in resistance, where Ron(ar) can be significantly smaller than Ron, on(+) on(-), is not a measurement artifact but a direct result of a deterministic atomic reconstruction of the conductive filament. This transformation is driven by the confluence of two key factors localized exclusively at the filament’s tip: extremely high electric fields, which can exceed 106 V/cm, and localized Joule heating, which elevates temperatures sufficiently to enable rapid Cu atom diffusion and ion mobility. The model reveals that under these conditions, a fragile, high-resistance truncated cone evolves into a more stable, lower-resistance double-cone or hourglass structure. A critical insight from this study is the introduction of a two-phase mechanism for the reset operation. The first, reinforcement phase, occurs at high currents near but below Ireset and is governed by electromigration and diffusivity, allowing the filament to reconfigure if sufficient time (ramp rate) is provided. The subsequent rupture phase occurs at a critical current and temperature, leading to the final local dissolution of the filament. The experimental validation using different voltage ramp rates is a key finding; it quantitatively proves that the timescale for ion mobility—calculated from fundamental diffusivity data and the Einstein relation—is the decisive factor controlling which phase dominates. Slow ramp rates (~ 0.03 V/s) provide a ~ 33 s window for complete reconstruction, while fast rates (~ 2.0 V/s) bypass it entirely, leading to immediate rupture. Furthermore, our geometrical optimization, accurately predicts the maximum possible reinforcement factor (~ 11) and its strong dependence on the initial filament’s tip sharpness. The excellent agreement between this theoretical limit and the experimental data range (1 to 11.3) provides compelling evidence for the model’s validity. This deeper understanding has significant implications. It resolves long-standing debates about filament morphology and reveals that the reset outcome is not a simple process but a complex competition between reinforcement and rupture. This new paradigm is crucial for device engineering. The ability to control filament morphology through operational parameters like ramp rate and compliance current opens new avenues for enhancing device endurance and tailoring switching characteristics. Finally, the ability to intentionally strengthen a filament, as demonstrated here, presents a fascinating novel functionality that could be exploited in multi-state memory or neuromorphic computing systems, where precise and stable control of synaptic weights is paramount.29 In neuromorphic computing, the resistance of a memristive device represents the synaptic weight—the strength of a connection between artificial neurons—where a key challenge is updating this weight in a precise, analog manner28,29,38. The discovered “strengthening reset” effect, wherein a compliance-current-limited reset operation reinforces a filament instead of rupturing it, provides a unique and elegant solution. This phenomenon fundamentally transforms the reset operation from a simple, binary ‘erase’ function into a sophisticated ‘sculpting’ tool. Traditionally, memristive synapses have been hampered by asymmetric weight updates: a SET pulse (positive bias) increases the weight, while a RESET pulse (negative bias) decreases it. Now, with the strengthening (aborted) reset operation (but not RESET), we have two methods for potentiation. The first is the standard potentiation via SET, using a positive voltage to form or grow the filament. The second, and more novel, is potentiation via a constrained RESET operation, which applies a negative voltage with a specific, low compliance current to reconstruct and strengthen a weak filament. This dual approach allows for a more symmetric and potentially linear tuning of the synaptic weight, as both voltage polarities can be leveraged for strengthening, depending on the Icc(-) parameter. Furthermore, this versatility directly enables the implementation of sophisticated learning rules like Spike-Timing-Dependent Plasticity (STDP), where a Long-Term Potentiation (LTP) event could be triggered by either a standard SET pulse or a carefully calibrated strengthening reset pulse, giving designers more tools to fine-tune the learning process. The effect’s pronounced nature in high-resistance filaments makes it particularly valuable for making fine, incremental adjustments to weak synapses, which is often the most challenging regime. Beyond weight updates, the strengthening reset can enhance stability and retention for faint memories; weak filaments are often unstable and prone to decay, but by periodically applying a mild, constrained reset operation pulse as a reinforcement mechanism, a neuromorphic system can actively stabilize these synaptic structures against forgetting. In summary, this effect provides neuromorphic circuit designers with an additional, highly controllable knob to finely craft and maintain synaptic weights, paving the way for more stable, linear, and biologically plausible neural networks.