Adaptive spatial hashing with dual-domain memristive hardware Authors: Dong Hoon Shin, Wonho Choi, Sunwoo Cheong, Néstor Ghenzi, Sung Keun Shim, Sung Ho Kim, Kunhee Son, Soo Hyung Lee, Byongwoo Park, Jonghoon Shin, Cheol Seong Hwang Published: 2026 ABSTRACT Locality-sensitive hashing is a technique for approximate similarity search, but fixed thresholds and analog encoding inefficiencies have limited its hardware implementation. This work introduces a dual-domain adaptive spatial hashing (DASH) architecture on a monolithic one-transistor-one-resistor active array operating in both digital and analog domains, with an integrated multifunctional memristor. DASH performs entropy-maximized random projection and data-driven bias adaptation in the analog domain, followed by efficient Hamming-distance computation in the digital domain. Using dual-domain vector-matrix multiplication, multidimensional inputs are compressed into binary hash codes, enabling compact similarity encoding within a unified memristive hardware platform. Experimental validation on three-dimensional-clustered synthetic data shows that DASH maintains spatial separability and improves bit entropy through adaptation. Large-scale simulations on a digit dataset demonstrate improved semantic preservation, similarity recall, and noise resilience compared to non-adaptive hashing. These results position DASH as a scalable, hardware-native solution for energy-efficient, locality-aware similarity search in edge and neuromorphic systems. Introduction Locality-sensitive hashing (LSH) has become a powerful method for assessing similarity in high-dimensional data spaces1,2. By mapping input vectors into discrete hash bins, LSH enables efficient and scalable retrieval for data mining tasks, such as approximate nearest-neighbor search, data clustering, and redundancy removal in large datasets, where similar inputs share higher bit similarity3,4. Although several software algorithms have been developed, implementing LSH directly in hardware remains challenging due to difficulties in efficient similarity computation, power and space constraints, and the need for adaptability to different data distributions. To overcome these limitations, interest in memory-centric computing frameworks that tightly integrate data storage and processing continues to grow. Among these approaches, memristor-based crossbar arrays5–7 are particularly promising for hardware implementation of LSH8–10. These arrays offer compact, nonvolatile memory capabilities with native support for highly parallel analog programmability and vector-matrix multiplication (VMM). However, memristive devices exhibit a trade-off between stochastic and deterministic switching behaviors, which limits the range of operations that can be performed efficiently within a single hardware system11–13. In LSH, where concepts such as randomly generated hyperplanes are used to divide high-dimensional space, controlled stochasticity is especially vital for effective entropy-based projection. Because the memristor’s programmed conductance directly influences the semantics of the VMM operation, the diversity and flexibility of supported computations are closely tied to device-level characteristics14,15. Fully utilizing these properties requires hardware-software co-design, where device-level features directly support algorithmic functions. However, conventional LSH typically relies solely on fixed, randomly selected hyperplanes without adjustment. Furthermore, while previous memristive hashing implementations have demonstrated in-memory search capabilities using fixed random projections, they often lacked the adaptability required to handle skewed or complex data distributions efficiently8,10. This work shows a dual-domain adaptive spatial hashing (DASH) structure built on a monolithic one-transistor-one-resistor (1T1R) active array composed of high-performance n-type thin-film transistors (TFTs) and multimodal resistive-switching memristors. DASH utilizes the dual-domain operational capabilities of the 1T1R arrays to perform analog-domain VMM for continuous distance calculation and spatial thresholding, as well as digital-domain VMM for logic-based hash evaluations. The system introduces an adaptation step that actively fine-tunes hyperplanes based on the statistical features of the input data, enabling a hardware-friendly generation-to-search pipeline via data-driven bias adaptation16. Supplementary Fig. 1 illustrates how the framework exploits stochastic switching for hyperplane creation and precise conductance adjustments for adaptive thresholding. The core memristive platform and its measurement system were developed to support multi-domain, data-driven computation. Benchmark tests and large-scale simulations confirm that DASH preserves spatial structure, improves similarity resolution, and enhances retrieval accuracy, highlighting its potential as a scalable, efficient solution for in-memory similarity search. Results Dual-domain memristive crossbar array Figure 1 summarizes the physical implementation and structural analysis of the dual-domain 1T1R hardware platform for the proposed DASH. Figure 1a shows the optical microscopy images of the fabricated 16 × 16 1T1R array. Monolithic integration was achieved by connecting the drain terminal of each n-type TFT to the bottom electrode of a memristor, forming 1T1R cells. The memristor used in this work is a Ta/HfO2/RuO2 (THR) device, chosen for its multi-phase reset behavior and its suitability for both analog and binary data representation17. The TFT, made with a 4.5-nm-thick sputtered In2O3 channel, provides sufficient drive current to switch the THR memristor while also serving as a selector to suppress leakage current in the array. Figure 1b illustrates the schematic diagram of the integrated 1T1R structure. The gate, source, and drain of the TFT are overlaid with a memristor layer formed on top of the drain, with sputter-deposited Al2O3 used as the interlayer dielectric (ILD) to enable monolithic integration. Figures 1c–f show the cross-sectional transmission electron microscopy (TEM) images and energy-dispersive X-ray spectroscopy (EDS) analyses of both the TFT and THR memristors, confirming the layer alignment and structural consistency with the device design. The methods section and Supplementary Fig. 2 outline the complete fabrication process of the 1T1R array, including its structural analysis and ILD.Fig. 1Structural and materials characterization of the dual-domain 1T1R array platform.a Optical microscopic images of the monolithically integrated 16 × 16 1T1R array composed of In2O3 thin-film transistors and Ta/HfO2/RuO2 memristors. Scale bars represent 1 mm for the array and 30 μm for the magnified image. b, Schematic illustration of the 1T1R stack, where the drain of the TFT connects to the bottom electrode of the memristor, separated by an Al2O3 interlayer dielectric. c–d Cross-sectional TEM images of the integrated device showing the TFT and memristor stack. e, f EDS elemental mapping confirms uniform spatial distribution and integrity of each layer. Scale bars in (c–f) represent 10 nm. Structural and materials characterization of the dual-domain 1T1R array platform. a Optical microscopic images of the monolithically integrated 16 × 16 1T1R array composed of In2O3 thin-film transistors and Ta/HfO2/RuO2 memristors. Scale bars represent 1 mm for the array and 30 μm for the magnified image. b, Schematic illustration of the 1T1R stack, where the drain of the TFT connects to the bottom electrode of the memristor, separated by an Al2O3 interlayer dielectric. c–d Cross-sectional TEM images of the integrated device showing the TFT and memristor stack. e, f EDS elemental mapping confirms uniform spatial distribution and integrity of each layer. Scale bars in (c–f) represent 10 nm. Figure 2 summarizes the electrical performance of the fabricated dual-domain 1T1R array. Figure 2a shows the transfer characteristics of the In2O3 TFT, displaying standard n-type behavior with a positive threshold voltage (Vth) of ~ 1.3 V, as illustrated in the linear-scale inset. Figure 2b demonstrates the device-to-device (D2D) uniformity across 100 TFT devices, with all devices showing nA-scale off currents at a gate voltage (Vg) of 0 V. This low off-state leakage ensures favorable cell selectivity in the array by suppressing sneak-path currents. Figure 2c presents the output characteristics, highlighting strong gate controllability and a low transresistance (<500 Ω) in the on state, confirming the high drive capability required for THR memristor programming. Supplementary Fig. 3 provides additional TFT performance metrics. The multimodal switching properties of the THR memristor, which can exhibit both stochastic and deterministic behavior depending on the switching conditions, have been thoroughly explained elsewhere17. Fig. 2d–i display the switching behavior and variability of the integrated THR memristor under various operational modes. Figure 2d shows the direct current (d.c.) current-voltage (I–V) sweep of a pristine THR device under gate-controlled compliance (Vg,set = 1.5 V, Vg,reset = 3.0 V), measured while varying the Vte. Specifically, a lower Vg,set is applied to prevent overshoot current and irreversible breakdown by limiting the drain current through TFT compliance. In contrast, a higher Vg,reset is employed to ensure sufficient current flow to effectively disrupt the conducting filaments during the reset process. A sharp transition between the high resistance state (HRS) and the low resistance state (LRS) is observed, resulting in a memory window that is larger than two orders of magnitude, making it suitable for digital operations (binary logic). Conversely, as shown in Fig. 2e, applying partial reset voltages (e.g., Vte = – 1.5 V in a three-terminal setup) yields intermediate resistance states (IRS) with a wider variation. Therefore, the device can be operated in a mode that causes stochastic conductance changes for hyperplane projection or enables the programming of specific conductance levels for threshold tuning. These flexible operating modes form the basis for DASH’s hybrid computing approach. Supplementary Fig. 4 presents additional memristor reliability metrics, including switching endurance and state retention.Fig. 2Electrical characterization of TFT and THR devices for dual-domain operation.a Transfer characteristics of a representative In2O3 TFT showing typical n-type behavior with positive threshold voltage. The inset shows the transfer curve plotted on a linear scale. b Device-to-device uniformity in off-current across 100 TFTs confirms strong array selectivity. The inset shows a magnified plot of the region from − 0.5 V to 0.5 V. c Output characteristics with different gate voltages of the TFT. d Digital domain operation of THR memristor, showing abrupt bipolar switching with > 100 × memory window. e Partial reset operation yields intermediate resistance states (IRS), enabling analog programmability and stochastic dispersion under analog domain operation. f Extracted conductance values from individual cells under one-shot mapping with varying reset voltages (Vte) ranging from – 1.5 V to – 2.4 V, measured at Vte = 0.1 V and Vg = 3 V. Devices exhibit a broad distribution of intermediate states between well-defined LRS (~ 1 mS) and HRS ( ~ 1 μS) levels. g Write-and-verify based target mapping from 100 μS to 1000 μS, demonstrating fine-grained, overlap-free programming accuracy across the dynamic range. h One-shot mapping results at Vte = – 1.8 V in an 8 × 8 subarray, showing randomly distributed conductance states for stochastic projection. i Target mapping using iterative programming yields precise and uniform conductance levels across the same subarray. Electrical characterization of TFT and THR devices for dual-domain operation. a Transfer characteristics of a representative In2O3 TFT showing typical n-type behavior with positive threshold voltage. The inset shows the transfer curve plotted on a linear scale. b Device-to-device uniformity in off-current across 100 TFTs confirms strong array selectivity. The inset shows a magnified plot of the region from − 0.5 V to 0.5 V. c Output characteristics with different gate voltages of the TFT. d Digital domain operation of THR memristor, showing abrupt bipolar switching with > 100 × memory window. e Partial reset operation yields intermediate resistance states (IRS), enabling analog programmability and stochastic dispersion under analog domain operation. f Extracted conductance values from individual cells under one-shot mapping with varying reset voltages (Vte) ranging from – 1.5 V to – 2.4 V, measured at Vte = 0.1 V and Vg = 3 V. Devices exhibit a broad distribution of intermediate states between well-defined LRS (~ 1 mS) and HRS ( ~ 1 μS) levels. g Write-and-verify based target mapping from 100 μS to 1000 μS, demonstrating fine-grained, overlap-free programming accuracy across the dynamic range. h One-shot mapping results at Vte = – 1.8 V in an 8 × 8 subarray, showing randomly distributed conductance states for stochastic projection. i Target mapping using iterative programming yields precise and uniform conductance levels across the same subarray. Figure 2f summarizes the conductance states achieved under different reset Vte from – 1.5 V to –2.4 V. All state readouts were performed at Vte = 0.1 V, with Vg set to 3 V for the selected cell during programming. As the magnitude of the reset Vte increases, the devices show various intermediate conductance values between the two separated extremes of LRS ( ~ 1 mS) and HRS ( ~ 1 μS). The LRS and HRS have minimal variations due to their saturated switching behavior17. However, the various IRSs show significant conductance dispersion, especially under partial reset conditions. Such a variation is usually undesirable in typical memristor applications, such as VMM calculation. However, such a property enables one-shot mapping of hyperplanes, where a single reset pulse in the intermediate regime can produce randomly distributed conductance states across the array. Such variability acts as a natural source of entropy for stochastic projection in DASH. Supplementary Fig. 5 provides a detailed statistical analysis of this switching dispersion across different operating regions. In contrast, it is also necessary to assign a specific target conductance value to the THR devices, which can be achieved in the conductance region close to LRS. Figure 2g shows that iterative write-and-verify operations6,18, targeting conductance levels from 100 μS to 1000 μS, allow precise conductance tuning without overlap between programmed states. This IRS-to-LRS region was specifically chosen to ensure I-V linearity and minimize read noise for accurate VMM17. The detailed procedure for the write-and-verify programming method is described in the Methods section. Lastly, Fig. 2h, i demonstrate the mapping behavior in an 8 × 8 1T1R subarray. One-shot mapping at Vte = – 1.8 V (Fig. 2h) results in a broad D2D spread of conductance values, which is beneficial for randomized hashing. In contrast, target mapping (Fig. 2i) produces tightly controlled levels across the array with minimal variation. Also, Supplementary Figs. 6 and 7 present the switching behavior of the THR under incremental pulsing and the corresponding array-level reproducibility for both switching modes. These results confirm the reliability and dual functionality of the proposed 1T1R platform, validating its ability to support both stochastic projection and deterministic thresholding modes required for DASH. Crossbar array operation for adaptive spatial hashing The DASH enhances conventional LSH by introducing an adaptation step that actively adjusts the randomly generated hyperplanes to better align with the structure of the input data. This method not only enhances spatial representation but also enables compact, efficient in-memory computation when implemented on specialized hardware. To deploy DASH in hardware, the system must support three essential functions: generating random hyperplanes, adapting thresholds, and processing hash data of binarized similarity representations. Figure 3 illustrates the schematic diagram of how DASH operates, as well as the hardware-level testing of the dual-domain 1T1R platform that enables its implementation.Fig. 3System architecture and hardware validation of DASH on a 1T1R memristive platform.a Photograph of the custom-built PCB integrating an MCU, multi-channel DAC, and TIA for dual-domain operation of the 1T1R array. Electrical access to the array is achieved via a custom probe card, ensuring stable subarray interfacing. b Schematic representation of the DASH data flow, showing dual-mode operation: analog domain VMM for projection and thresholding, and digital domain VMM for Hamming distance-based similarity matching. c Conceptual overview of the analog domain DASH process. Random hyperplanes are implemented as differential-conductance pairs across two adjacent memristor columns, with input vectors applied as voltage stimuli to perform projection via an analog VMM. d Illustration of the digital domain DASH operation. Binary hash codes are stored as differential states in paired cells, and the Hamming distance between input queries and stored entries is computed through current-mode digital VMM. e End-to-end DASH processing pipeline. Query vectors are projected and binarized in the analog domain with adaptive biasing, then matched to stored codes in the digital domain via parallel estimation of Hamming distances. f Histogram of differential conductance values measured from randomly initialized cell pairs (one-shot mapping), demonstrating symmetrical stochasticity suitable for random hyperplane generation. g Analog I–V characteristics of target-mapped conductance states, exhibiting linear response necessary for accurate analog VMM. h Analog VMM validation using a 4 × 2 subarray programmed with target conductances. Measured dot-product outputs align with ideal projections, confirming the accuracy of analog inference. i Digital domain validation of bitwise HD computation. The output current varies with the mismatch count across binary-coded differential pairs, with 100 readout cycles confirming consistent behavior. Horizontal bars indicate the mean values. j Full digital domain HD readout from 8-bit query vectors. Output currents form discrete, non-overlapping clusters for HD = 0–8, validating the feasibility of parallel similarity comparison. Horizontal bars indicate the mean values. System architecture and hardware validation of DASH on a 1T1R memristive platform. a Photograph of the custom-built PCB integrating an MCU, multi-channel DAC, and TIA for dual-domain operation of the 1T1R array. Electrical access to the array is achieved via a custom probe card, ensuring stable subarray interfacing. b Schematic representation of the DASH data flow, showing dual-mode operation: analog domain VMM for projection and thresholding, and digital domain VMM for Hamming distance-based similarity matching. c Conceptual overview of the analog domain DASH process. Random hyperplanes are implemented as differential-conductance pairs across two adjacent memristor columns, with input vectors applied as voltage stimuli to perform projection via an analog VMM. d Illustration of the digital domain DASH operation. Binary hash codes are stored as differential states in paired cells, and the Hamming distance between input queries and stored entries is computed through current-mode digital VMM. e End-to-end DASH processing pipeline. Query vectors are projected and binarized in the analog domain with adaptive biasing, then matched to stored codes in the digital domain via parallel estimation of Hamming distances. f Histogram of differential conductance values measured from randomly initialized cell pairs (one-shot mapping), demonstrating symmetrical stochasticity suitable for random hyperplane generation. g Analog I–V characteristics of target-mapped conductance states, exhibiting linear response necessary for accurate analog VMM. h Analog VMM validation using a 4 × 2 subarray programmed with target conductances. Measured dot-product outputs align with ideal projections, confirming the accuracy of analog inference. i Digital domain validation of bitwise HD computation. The output current varies with the mismatch count across binary-coded differential pairs, with 100 readout cycles confirming consistent behavior. Horizontal bars indicate the mean values. j Full digital domain HD readout from 8-bit query vectors. Output currents form discrete, non-overlapping clusters for HD = 0–8, validating the feasibility of parallel similarity comparison. Horizontal bars indicate the mean values. Figure 3a shows an image of a custom-printed circuit board (PCB) designed for dual-domain 1T1R array operation. The system includes a microcontroller unit (MCU) that interfaces with a host computer, a multi-channel digital-to-analog converter (DAC) used to apply voltages to the gate and drain terminals of the 1T1R devices, and a high-resolution transimpedance amplifier (TIA) to measure the output current at the source terminals. Electrical connections to the array are made using a custom-fabricated probe card with contact interfaces, ensuring accurate access to subarrays and reliable measurements. Supplementary Fig. 8 shows the complete measurement setup of the PCB. To enable both analog- and digital-domain VMM, the hardware system follows the data flow shown schematically in Fig. 3b. In both domains, computation occurs through the application of voltage and the readout of current, but the operational mode determines the meaning of the mapped conductance within the array. When random conductance distributions are implemented via one-shot mapping, or when specific conductance values are programmed using target mapping, analog-domain VMM is performed by applying analog input voltages and reading the accumulated output current. Conversely, when binary data representing HRS and LRS are programmed into the array, digital-domain VMM is enabled by applying binary voltages (0 V and 0.1 V, corresponding to 0 and 1) and computing bitwise similarity via parallel current summation, which is subsequently used to estimate the Hamming distance (HD). Figures 3c–e illustrate the step-by-step process of the DASH system. The analog process begins by creating random hyperplanes, which are physically represented as differential-conductance pairs in the memristor array (Fig. 3c). Each hyperplane is implemented with two neighboring columns, where the differential conductance between paired memristors determines each element of the hyperplane vector16. An input vector (such as a three-dimensional (3D) coordinate) is normalized and applied to the array as voltage inputs. The resulting analog VMM operation produces current outputs that indicate the signed projections of the input vector onto each hyperplane, thereby showing the input’s relative position in space with respect to each hyperplane16. However, because these hyperplanes are inherently random, the resulting partitions may not effectively reflect the structure of the data space. To improve the discriminative power of spatial hashing, a hyperplane adaptation step is added. For each hyperplane, the median of the current responses across all input samples is calculated, capturing the central tendency of the spatial distribution. Supplementary Note 1 provides the detailed mathematical formalism regarding this median-based bias calculation and its entropy-maximizing effect. This median value is then stored in a pair of memristor cells as a differential conductance bias, forming an additional row vector within the array. While this median-based approach minimizes peripheral complexity, the use of dedicated bias rows allows the architecture to scale naturally toward multi-threshold or hierarchical methods, providing the flexibility to handle more complex, multimodal data distributions. By re-evaluating the VMM operation with this bias term, the hyperplane threshold is adaptively shifted to better align with the data distribution. The final current output, either positive or negative, is thresholded to generate a binary hash code for each input, thereby compressing high-dimensional analog data into compact, discriminative digital representations. In the digital domain, shown in Fig. 3d, the binary hash codes generated by the DASH process are physically mapped onto the memristor array by programming each bit into a differential pair of cells with one representing logic ‘0’ and the other ‘1’. Multiple hash codes, created by several hyperplanes per data point, are written simultaneously across the array to improve robustness and fault tolerance in high-dimensional similarity evaluation. Once stored, the similarity is determined by measuring the HD between a query and stored entries. This is achieved by applying two-level voltage inputs (e.g., 0 V and 0.1 V) that correspond to the binary query code. The output current then indicates the number of mismatches, or bit differences, between the query and stored codes. Because each mismatch adds linearly to the output current, the array inherently performs a parallel and non-Euclidean comparison across all bits, enabling efficient and scalable digital VMM-based similarity search. Figure 3e illustrates the overall search algorithm flow of DASH. A query sample is first hashed through the analog domain process, which involves voltage-based projection onto hyperplanes and bias adjustment. The resulting binary code is then transferred to the digital domain, where a digital VMM computes the HD between the query and stored items. Based on this similarity, DASH returns the most relevant matches. This hierarchical analog-to-digital architecture facilitates approximate nearest neighbor search in high-dimensional data spaces. Figures 3f–j illustrate the hardware-level validation of core DASH operations, which was tested across multiple subarrays within the fabricated 1T1R array. Each subfigure includes an inset that schematically illustrates the mapping configuration of active cells in the corresponding subarray, ensuring clarity and consistency in interpreting the operation mode. Figure 3f presents a histogram of differential conductance obtained from 100 independent programming trials on a single differential cell pair. Initially, in the LRS, the devices were reset with Vte = – 1.5 V and Vg = 3 V to induce stochastic IRSs. The resulting differential conductance values are statistically centered around zero with a sufficient standard deviation, confirming the reproducibility of voltage-dependent Gaussian sampling. This stochastic behavior supports the feasibility of using such distributions to form high-dimensional random hyperplanes. Figure 3g shows the I-V characteristics of analog domain readout from target-mapped conductance values. A well-separated linear I-V responses confirm the linear superposition behavior required for analog VMM, making this mode well-suited for analog similarity computation. Figure 3h further validates analog VMM. A 4 × 2 subarray was programmed with target conductance values, and multiple instances of analog VMM were executed. The measured current outputs coincide well with the desired value from theoretical dot-product computation, demonstrating high fidelity of analog mapping and inference across physically distinct cells. In the digital domain, Fig. 3i illustrates how a two-cell HD calculation is acquired16. As shown in the inset, two memristor cells were programmed into binary states (HRS in orange for ‘0’, LRS in purple for ‘1’) and paired with two different input voltages (V1, V2) representing the bits to be compared. Over 100 read cycles, the output current magnitude directly correlated with the HD between the bit pairs, confirming the bitwise comparison function using differential conductance. Lastly, Fig. 3j extends this digital operation to a larger subarray, in which 8-bit binary input vectors are mapped to 16 memristor cells in a single row. Input voltages for each bit were applied simultaneously, and the total output current was used to measure HD, which ranged from 0 to 8. The resulting current levels formed distinct clusters for each HD value, with no overlap, demonstrating that the 1T1R array can compute the HD in parallel. Overall, these results confirm the practical feasibility of implementing DASH’s unit operations, including one-shot stochastic projection, analog domain adaptation, and VMM, as well as digital HD calculation, in a single memristive hardware platform. The array’s dual-mode operation and reconfigurability enable DASH to flexibly switch between analog and digital similarity paradigms depending on the task. Experimental validation of adaptive spatial hashing Figure 4 presents the experimental validation of the proposed DASH algorithm using a fabricated 16 × 16 1T1R array. A complete analog-to-digital spatial hashing pipeline was implemented, in which 16 synthetic 3D input points, forming four clusters, were projected onto eight randomly generated hyperplanes, followed by bias adaptation and digital readout.Fig. 4Experimental validation of DASH using a fabricated 1T1R array and a synthetic 3D dataset.a Three-stage illustration of DASH processing. Left: 16 synthetic 3D points grouped into four spatial clusters. Middle: binary classification using a representative hyperplane without adaptation, showing poor intra-cluster separability. Right: improved binary segmentation after applying bias adaptation via DASH. b Hyperplane matrix composed of eight 3D vectors encoded as a 3 × 8 conductance matrix (in Siemens), physically mapped using differential memristor pairs. The color scale bar represents conductance in Siemens (S). c Analog domain VMM output currents obtained by applying normalized input voltages (± 0.2 V) to the array. Some hyperplanes exhibit sign bias, which limits binary discriminability. The color scale bar indicates output current in Amperes (A). d Median-based bias adaptation: the median current for each hyperplane is mapped to a differential conductance and integrated as an additional bias row. The color scale bar represents conductance in Siemens (S). e Adapted analog VMM results after incorporating bias terms, showing more symmetric current distributions around 0 A. The color scale bar indicates output current in Amperes (A). f Binary codes obtained from sign-based thresholding of analog outputs, read in the digital domain at 0.1 V, demonstrating distinct and well-separated patterns. The color scale bar indicates output current in Amperes (A). g Histogram of output currents from pairwise HD computation across all binary codes. Discrete, non-overlapping current levels for HD = 0 to 8 confirm the reliability of digital-domain HD computation. h, i Distance matrices from the original 3D data (Euclidean) and DASH-encoded binary data (Hamming), showing preserved cluster structure. j Scatter plot of ED vs HD across all point pairs; correlation coefficient = 0.761 confirms locality preservation in binary space. k Bitwise entropy comparison before and after DASH across hyperplanes. Post-adaptation entropy increases indicate improved binary balance and expressivity. Experimental validation of DASH using a fabricated 1T1R array and a synthetic 3D dataset. a Three-stage illustration of DASH processing. Left: 16 synthetic 3D points grouped into four spatial clusters. Middle: binary classification using a representative hyperplane without adaptation, showing poor intra-cluster separability. Right: improved binary segmentation after applying bias adaptation via DASH. b Hyperplane matrix composed of eight 3D vectors encoded as a 3 × 8 conductance matrix (in Siemens), physically mapped using differential memristor pairs. The color scale bar represents conductance in Siemens (S). c Analog domain VMM output currents obtained by applying normalized input voltages (± 0.2 V) to the array. Some hyperplanes exhibit sign bias, which limits binary discriminability. The color scale bar indicates output current in Amperes (A). d Median-based bias adaptation: the median current for each hyperplane is mapped to a differential conductance and integrated as an additional bias row. The color scale bar represents conductance in Siemens (S). e Adapted analog VMM results after incorporating bias terms, showing more symmetric current distributions around 0 A. The color scale bar indicates output current in Amperes (A). f Binary codes obtained from sign-based thresholding of analog outputs, read in the digital domain at 0.1 V, demonstrating distinct and well-separated patterns. The color scale bar indicates output current in Amperes (A). g Histogram of output currents from pairwise HD computation across all binary codes. Discrete, non-overlapping current levels for HD = 0 to 8 confirm the reliability of digital-domain HD computation. h, i Distance matrices from the original 3D data (Euclidean) and DASH-encoded binary data (Hamming), showing preserved cluster structure. j Scatter plot of ED vs HD across all point pairs; correlation coefficient = 0.761 confirms locality preservation in binary space. k Bitwise entropy comparison before and after DASH across hyperplanes. Post-adaptation entropy increases indicate improved binary balance and expressivity. Figure 4a illustrates the evolution of spatial encoding in three steps: the initial distribution of 16 clustered points (left), the binary mapping obtained by a representative hyperplane without adaptation (middle), and the enhanced binary separability after applying DASH (right). Without adaptation, the hyperplane cannot distinguish between points that are differently clustered, assigning the same binary label despite spatial differences. After adaptation, more precise spatial segmentation becomes apparent, as evidenced by the more distinct binary outputs. Figure 4b displays the initial hyperplane matrix, where each 3D hyperplane is represented as a 3 × 8 matrix and mapped into differential conductance values (in Siemens). Figure 4c shows the resulting output currents in the analog domain VMM when normalized voltage vectors (scaled to ± 0.2 V) representing input points are applied to the array. These currents reveal the relative positions of the points concerning each hyperplane. Notably, several hyperplanes show biased response distributions, and all output currents have the same sign, indicating their appropriate binary expressions are improbable. Figure 4d shows the hyperplane matrix after the bias adaptation step, which was used to address the imbalance mentioned above. For each hyperplane, the median current across all input points is calculated, capturing the primary trend of the spatial projections across the dataset. This median value is then mapped to an equivalent differential conductance, which is physically implemented as an extra bias row in the memristor array. The updated hyperplane matrix reflects this correction. Figure 4e displays the adapted VMM results after adding the bias terms. The current distributions become more symmetric around zero, improving each hyperplane’s ability to produce balanced binary outputs. Based on the sign of these currents, binary codes are generated, as shown in Fig. 4f, where the digital-domain results (read at 0.1 V) clearly display distinct binary patterns across the data points. Although near-zero differential currents in boundary regions (white regions in Fig. 4e) may occasionally lead to stochastic bit flips due to measurement noise, such occurrences are limited to ambiguous samples at the decision boundary and do not compromise the global semantic structure. Figure 4g presents a histogram of output currents from pairwise comparisons of binary codes, categorized by their HD, which evaluates the quality of the binary encoding. All possible HD values (from 0 to 8) are observed without overlap in the output current distributions, confirming the discrete and reliable nature of digital-domain computations. This separation indicates that the memristor array can perform quantized similarity estimation through physical HD calculations. Figures 4h–k evaluate the spatial accuracy and expressiveness of DASH. Figure 4h shows the Euclidean distance (ED) matrix derived from the original 3D coordinates, while Fig. 4i presents the HD matrix based on the binary codes. Despite the dimensionality reduction and binarization, the cluster structures are well preserved. Figure 4j provides a quantitative comparison of ED and HD via a scatter plot, revealing a positive correlation (coefficient = 0.761), indicating that spatial similarity in the original domain is maintained after binarization. This positive correlation suggests that bit differences (HD) still reflect geometric distances (ED), confirming that DASH preserves the relative structure of the data even in binary space. Finally, Fig. 4k compares the bitwise entropy before and after adaptation across all hyperplanes. The consistent increase in entropy after DASH confirms not only improved separability but also a more balanced use of binary states, thereby enhancing the representational capacity of the hash codes. Supplementary Figs. 9 and 10 further validate this result by extending the analysis to other 3D datasets with different numbers of clusters, demonstrating DASH’s scalability and robustness across diverse spatial data distributions. Performance evaluation of adaptive spatial hashing To assess the feasibility and effectiveness of the proposed DASH system, a comprehensive set of simulations was performed using the Digits dataset19, which contains 8 × 8 greyscale images of handwritten digits from 0 to 9. Figure 5a illustrates the overall processing procedure, and the Method section provides a detailed explanation of the simulation process. Each image was flattened into a 64-dimensional vector and projected into an N-bit binary space using a set of random hyperplanes (N = 16, 32, 64, 128, 256). Both adaptive and non-adaptive biasing methods were tested to evaluate how bias calibration affects semantic preservation and retrieval accuracy. In this task, each query image was encoded via DASH and compared to the original dataset using HD to identify the closest digit. For adaptive DASH, each hyperplane’s bias was set to the median of the non-adaptive projection values, effectively centering the bit distribution. Unlike traditional LSH, which requires numerous hyperplanes for accuracy, DASH enhances the expressiveness of each hyperplane through adaptation, enabling accurate representation with fewer bits2,20–22. This feature decreases hardware burden, as DASH supports input scaling in the row direction while maximizing projection efficiency in the column direction through data-driven hyperplane adaptation.Fig. 5Simulation-based evaluation of DASH using the Digits dataset.a Schematic procedure of DASH implementation for 8 × 8 greyscale digit images. Each image is flattened into a 64-dimensional vector and projected onto N random hyperplanes (N = 16–256), followed by binary encoding via either non-adaptive or adaptive (DASH) biasing. b–d Pairwise distance matrices comparing 10 randomly sampled digit vectors. b Cosine distance from the original 64D feature space. c HD between binary encodings using non-adaptive hyperplanes. d HD with DASH, showing improved preservation of semantic structure compared to (c). e–g Two-dimensional t-SNE visualizations of the digit data. e Original 64D features. f Non-adaptive binary embeddings, showing dispersed intra-class and blurred inter-class boundaries. g DASH-based binary embeddings, recovering compact and well-separated clusters akin to (e). h–j Evaluation of robustness under random bit-flip perturbations. h Pearson correlation between corrupted and original binary hash codes as a function of the bit-flip ratio. i Recall@5 and (j) Precision@15 for query retrieval using binary HD search. DASH outperforms the non-adaptive method, especially under higher bit-flip ratios, demonstrating superior stability, discriminability, and search performance. Simulation-based evaluation of DASH using the Digits dataset. a Schematic procedure of DASH implementation for 8 × 8 greyscale digit images. Each image is flattened into a 64-dimensional vector and projected onto N random hyperplanes (N = 16–256), followed by binary encoding via either non-adaptive or adaptive (DASH) biasing. b–d Pairwise distance matrices comparing 10 randomly sampled digit vectors. b Cosine distance from the original 64D feature space. c HD between binary encodings using non-adaptive hyperplanes. d HD with DASH, showing improved preservation of semantic structure compared to (c). e–g Two-dimensional t-SNE visualizations of the digit data. e Original 64D features. f Non-adaptive binary embeddings, showing dispersed intra-class and blurred inter-class boundaries. g DASH-based binary embeddings, recovering compact and well-separated clusters akin to (e). h–j Evaluation of robustness under random bit-flip perturbations. h Pearson correlation between corrupted and original binary hash codes as a function of the bit-flip ratio. i Recall@5 and (j) Precision@15 for query retrieval using binary HD search. DASH outperforms the non-adaptive method, especially under higher bit-flip ratios, demonstrating superior stability, discriminability, and search performance. Figures 5b–d compare the pairwise distance matrices of 10 randomly sampled digit vectors before and after binarization, which assess whether the binary projection preserved the semantic structure of the data23,24. Fig. 5b shows the cosine distance matrix derived from the original 64-dimensional vectors, illustrating true similarities among digits. Figures 5c, d display the corresponding HD for the non-adaptive and adaptive (DASH) binary encodings, respectively. The same color map scale was used to highlight differences. Figure 5d shows a clear relationship between cosine distance and HD is observed under DASH, although the metric spaces differ, i.e., continuous cosine versus discrete HD. In contrast, the non-adaptive version does not reflect the same semantic grouping, emphasizing the importance of adaptation for meaningful binarization. This shows that DASH-encoded binary vectors preserve not only class membership but also pairwise similarity trends originally defined in cosine space. Figures 5e–g present two-dimensional t-distributed stochastic neighbor embedding (t-SNE) visualizations of data in different representational spaces14,25,26. t-SNE is a nonlinear dimensionality reduction method that projects high-dimensional data onto a 2D plane while maintaining local neighborhood relationships, making it useful for visually assessing class separability25,26. In these plots, colored contours depict kernel density, which is the estimated boundary around each digit class, highlighting the compactness of intra-class clusters and the clear separation between classes. Ideally, with the original 64-dimensional vectors, each digit class forms a compact, well-separated cluster enclosed by distinct boundaries. When vectors are converted to binary codes without adjustment, the neighborhood structure partially deteriorates: intra-class points scatter more widely, and class boundaries become blurred and overlapping. This result indicates that the non-adaptive projection does not fully preserve the intrinsic geometry of the original space. In contrast, the DASH-based binary codes recover much of this structure. The adaptive bias centering step adjusts the activation probability of each bit, resulting in clusters that are not only more compact but also better separated by clearer boundaries that resemble the ideal case. It should be noted that the absolute cluster positions differ between the original and binary embeddings. This difference occurs because t-SNE preserves local neighborhoods rather than global geometry. Therefore, the critical point to note is that DASH maintains coherent intra-class grouping and distinct inter-class separation, as in the ideal case. Supplementary Figs. 11 and 12 show t-SNE visualizations for different hyperplane configurations, confirming that this property remains consistent across various projection dimensions. Additionally, Supplementary Figs. 13–16 provide a qualitative assessment of the generated hash codes for both non-adaptive and DASH cases. These results include visualizations of hash bit activation probability maps, thresholded binary signatures, and quantitative metrics, such as distance correlation with the original space, bit entropy, unique code ratio, intra- and inter-class HDs, and separability ratio. Overall, these analyses demonstrate that the adaptive bias calibration in DASH produces more balanced and discriminative binary codes, improving both the diversity and fidelity of the hash space. Finally, Fig. 5h–j compare the impact of the bit-flip ratio on the robustness of the proposed DASH framework versus the conventional non-adaptive approach under random bit-flip perturbations. In this test, the bit-flip ratio, p, ranged from 0 to 0.20, meaning each bit in the binary hash code had a p chance of flipping from 0 to 1 or vice versa to test the algorithmic resilience of the DASH. Each data point represents the average of 10 independent trials. Figure 5h shows the Pearson correlation between hash codes. At low bit-flip probabilities (0 ≤ p ≤ 0.05), the non-adaptive method exhibits a slightly higher correlation, as in nearly ideal conditions, the fixed mapping of the non-adaptive method preserves the original hash structure, resulting in a closer match to the ground truth. In contrast, the DASH method makes a small remapping adjustment even with minimal perturbation, which can slightly decay correlation in this range. However, as the bit-flip ratio increases (p ≥ 0.10), the DASH method demonstrates better resistance to degradation and eventually surpasses the non-adaptive baseline, especially in a hardware-based setup. Figures 5i, j evaluate the query performance of the proposed DASH framework using two standard retrieval metrics: Recall@5 and Precision@1527,28. Recall@5 measures the proportion of relevant items successfully retrieved within the top 5 results for a given query, reflecting the system’s ability to capture most correct matches in a small candidate set. Precision@15 measures the fraction of relevant items among the top 15 results, indicating how accurately the retrieved set represents the intended target class without including irrelevant samples. Across all tested conditions, the DASH-based adaptive biasing strategy consistently outperforms the conventional non-adaptive approach, closely approaching the ideal baseline derived from exact cosine similarity. This result means that DASH not only retrieves a larger fraction of relevant items within the top results but also ensures that the retrieved set remains accurate, even when operating in a compressed binary space under significant bit-flip perturbations. Such improvement directly translates into better search reliability and semantic preservation, ensuring that the nearest-neighbor relationships in the original feature space are more faithfully reflected even after binarization. These results demonstrate that DASH’s adaptive calibration generates binary codes with enhanced discriminability and robustness, rendering it more suitable for practical similarity search applications where accurate coverage (recall) and purity (precision) of results are crucial. Discussion This study presents DASH, a dual-domain adaptive spatial hashing system that physically implements entropy generation, analog adaptation, and digital similarity estimation within a single memristive platform. Unlike previous LSH methods, DASH enables co-optimized entropy generation and data-driven adaptation through stochastic and analog device behaviors within the given 1T1R crossbar hardware. While the dual-domain operation and bias adaptation introduce procedural steps, the hardware overhead is minimized by sharing the same peripheral circuitry and 1T1R array for both domains. Supplementary Table 1 presents a qualitative comparison that highlights the architectural efficiency of DASH relative to conventional analog-only and digital-only implementations. It is shown that DASH can dynamically reshape randomly sampled hyperplanes to reflect the data structure better through bias adaptation, thereby enhancing both separability and entropy across binary outputs. At the system level, DASH supports two complementary computing domains. The analog domain enables precise spatial mapping through VMM operations, while the digital domain performs logic-level similarity estimation using HD. These dual-domain operations are enabled by accurate control of memristor switching modes and reliable integration with thin-film transistors, as confirmed by extensive electrical testing and in situ measurements. The complete end-to-end validation from spatial data input to digital Hamming similarity calculation is experimentally demonstrated using a monolithic 16 × 16 1T1R array. This physical verification confirms the feasibility of deploying DASH as a hardware-native search primitive. Although the implemented system is based on a modest-scale array, the ability to perform the entire processing pipeline, from analog projection and adaptation to digital similarity computation, on a single physical platform highlights the completeness and practical viability of the proposed method. Furthermore, since DASH relies on standard VMM-based operations and programmable conductance states, the same architecture and computation protocol are likely to scale naturally to larger and more complex crossbar systems without requiring structural changes. Significantly, the same hardware, featuring a unified array structure and shared measurement infrastructure, can support both analog and digital processing modes, regardless of differences in device states or mapping methods. This dual-domain compatibility demonstrates the flexibility and scalability of the proposed system, enabling it to adapt smoothly across various computational modes while maintaining a consistent physical framework. Building on this scalability, simulation studies further demonstrate DASH’s effectiveness on high-dimensional data through a digit image search task. DASH not only maintains semantic structure across binary projections but also compensates for hardware limitations during spatial binarization, confirming its algorithmic advantage over conventional methods. The robustness of the system performance to moderate deviations from the exact median bias, indicating tolerance to device variability, was also verified. Supplementary Figs. 17–20 show comprehensive analysis results on the impact of hyperplane scaling, comparative evaluations against different hashing algorithms, and validation on more complex datasets. Furthermore, a rigorous bottom-up performance evaluation was conducted to quantify the energy and latency advantages of the DASH architecture. As detailed in Supplementary Note 2, DASH achieves a remarkable energy efficiency gain of up to 200.50× compared to conventional processor-based platforms. This performance leap is attributed to the dual-domain operational strategy, which minimizes data movement, and to the use of one-shot mapping for core hashing operations, effectively mitigating the von Neumann bottleneck. Supplementary Note 2 and Supplementary Tables 2–4 provide the step-by-step benchmarking methodology and a comprehensive breakdown of the five operational phases. Overall, DASH exemplifies a co-designed hardware–algorithm framework capable of data-adaptive and similarity-aware processing. Its architectural flexibility and proven scalability across various data domains indicate broad applicability in edge-AI systems, neuromorphic platforms, and in-memory associative computing. Methods Device fabrication The monolithically integrated 1T1R array was fabricated through the following process. The process began with the formation of the THR memristor on SiO2 (100 nm)/Si substrate. A 5 nm Ti adhesion layer, 50 nm Pt passivation layer, and 20 nm RuO2 bottom electrode were deposited on a substrate via reactive sputtering (SNTEK, CDS 5000), and patterned using standard photolithography followed by lift-off. A 4 nm HfO2 switching layer was deposited by atomic layer deposition (ALD, CN-1, Plus 200) using tetrakis(dimethylamino)hafnium and O3 at a substrate temperature of 210 °C. The top electrode, composed of 20 nm Ta and 50 nm Pt layers, was deposited via sputtering (SNTEK, CDS 5000) and defined by an additional lift-off step, thereby completing the memristor structure. Following the fabrication of memristors, array-level integration of TFTs was carried out to enable a 1T1R configuration. A 50 nm interlayer dielectric (Al2O3) was deposited using RF sputtering (Sorona, SRN120) to isolate the memristor from the overlying TFT electrically. A 40 nm Ni gate electrode was deposited by electron-beam evaporation (Sorona, SRN200), followed by an 8.5 nm HfO2 gate dielectric via ALD. The active channel layer of 4.5 nm In2O3 was deposited by RF sputtering (Sorona, SRN120) and patterned using HCl-based wet etching. Contact vias between the TFT and memristor were opened by selectively etching the gate dielectric and ILD using buffered HF, and subsequently filled with Ni to form vertical interconnections. Finally, a 50 nm Ni layer was deposited and patterned to form the TFT source and drain terminals, thereby completing the vertically stacked 1T1R active array. All fabrication processes were conducted at substrate temperatures below 400 °C, enabling COMS-compatible monolithic integration. Electrical measurements The I–V characteristics of the THR devices were measured using a semiconductor parameter analyzer (SPA, HP 4155 A), with the top electrode biased and the bottom electrode grounded throughout the sweep. For analog-mode conductance tuning, a write-and-verify protocol was employed to reach the target conductance precisely. Programming pulses were delivered to the top electrode, using positive voltages for set and negative voltages for reset operations. Each write cycle consisted of 10 pulses under Vg = 2 V, starting at Vte = ± 0.3 V. If the conductance remained outside the desired range, the pulse amplitude was increased stepwise by 0.05 V per iteration. In cases of overshoot, the pulse polarity was reversed, and the pulse resumed from ± 0.3 V using the same incremental approach. After each pulse set, the device conductance was monitored via read operations with Vg = 2 V and Vte = 0.1 V. The tuning process concluded when the conductance fell within ± 5 μS of the target. For array-level measurements, the 16 × 16 active array was mounted on a custom-designed probe card interfaced with an array measurement PCB. The test system for the experimental demonstration consisted of a host PC, a custom analog and digital mixed circuit mounted on a PCB. The host PC managed the measurement interface, controlled the MCU connected to the board, and processed the measurement results received via serial communication. An ARM Cortex-based MCU processed control commands from the host PC and subsequently transmitted them to the board’s integrated circuit (IC) components according to their intended purposes. The board included regulators for stable positive and negative power supplies to IC components and multiplexers (MUXs) for selecting devices and their corresponding terminals. Two 8-channel DACs were used to bias independent voltages to the gate and drain of the device, respectively, and to generate adjustable pulses. Current outputs from the source electrode were converted to voltages and then delivered to the MCU’s internal 12-bit ADC. Various amplifiers on the board were connected with adjustable passive resistors and utilized as transimpedance and inverting amplifiers. The host PC interface, the control code, and the embedded code for the MCU were all written in C. Array simulation for digit dataset hashing The DASH algorithm was evaluated using the scikit-learn digits dataset19, consisting of 8 × 8 greyscale images of handwritten digits (0–9). Each image was flattened into a 64-dimensional vector and L2-normalized. For each hyperplane dimensionality N, a 65 × 2 N crossbar array was simulated, with weights sampled directly from experimental differential conductance distributions to account for hardware non-idealities. These hyperplanes were then implemented in the analog-domain VMM, where input vectors were applied as voltages, and the resulting output currents were measured. The assumed array size scaled proportionally with the required number of hyperplanes. Binary codes were generated either by fixed thresholding at zero (non-adaptive) or by subtracting the median projection value per bit before thresholding (adaptive/DASH). The similarity between binary codes was computed using the normalized HD. To examine the preservation of semantic structure in the binary domain, we applied t-distributed Stochastic Neighbor Embedding (t-SNE) to both the original vectors and the binarized codes. Specifically, the top 100 samples (10 per digit class) were embedded into a two-dimensional space using the cosine distance metric as the input similarity measure, a perplexity of 30, and an early exaggeration factor of 12. This nonlinear dimensionality reduction preserves local neighborhood relationships, enabling qualitative comparison of cluster separation between the original and binary spaces. Performance evaluation included correlation with the original cosine distances, analysis of intra-class and inter-class distances, bit entropy, unique code ratio, and separability ratio. Retrieval accuracy was quantified using Recall@5 and Precision@15 across varying random bit-flip ratios (p = 0–0.20), where each bit had probability p of flipping from 0 to 1 or vice versa. Recall@K and Precision@K were computed as: Recall@K = Number of relevant items retrieved within top K / Total number of relevant items Precision@K = Number of relevant items retrieved within top K / K In this work, Recall@5 measures the fraction of relevant items present in the top 5 retrieved results, while Precision@15 measures the proportion of relevant items among the top 15 results. All simulations were performed using custom Python scripts built with NumPy, scikit-learn, and Matplotlib, ensuring identical processing pipelines in both adaptive and non-adaptive modes for a fair comparison. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file