Tuesday, November 17, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Charge-Trap Memcapacitor Array Enables Linear Capacitive Vector-Matrix Multiplication for In-Memory Computing

November 17, 2026
in Science News
Cassandra Pierce
By Cassandra Pierce Scienmag Editorial Profile - Systems Neuroscience
Reading Time: 6 mins read
0
Charge-Trap Memcapacitor Array Enables Linear Capacitive Vector-Matrix Multiplication for In-Memory Computing

Charge-Trap Memcapacitor Array Enables Linear Capacitive Vector-Matrix Multiplication for In-Memory Computing

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Engineers have fabricated and experimentally demonstrated a vertically stackable charge-trap-flash memcapacitor array capable of executing linear charge-domain vector–matrix multiplication with an error below 0.227 percent. Developing energy-efficient hardware platforms for modern artificial intelligence workloads has become increasingly urgent as neural network model parameters expand into hundreds of billions of values. In contemporary computer systems relying on conventional von Neumann computing architectures, arithmetic throughput is frequently constrained not by raw processing speed but by the latency and substantial energy overhead associated with transferring data across deep memory hierarchies between high-bandwidth memory, standard dynamic random-access memory, and non-volatile secondary storage. In-memory computing architectures aim to alleviate these data-movement bottlenecks by collocating storage and arithmetic operations within the physical memory array itself. While conventional in-memory computing demonstrations predominantly rely on resistance-based memory devices that sum continuous electrical currents during read operations, such conductance-domain approaches encounter severe thermal management challenges and static power dissipation when scaled into dense three-dimensional configurations.

To overcome the fundamental trade-off between integration density and analog computing parallelism, the researchers developed a capacitive readout architecture fabricated from charge-trap-flash memory stacks derived from commercial three-dimensional NAND flash manufacturing technology. Traditional three-dimensional NAND flash memory provides ultra-high vertical integration density across hundreds of tiers, but its conventional serially connected cell string architecture historically prevented the parallel analog operations required for high-throughput vector–matrix multiplication workloads. Rather than using conductance-weighted current summation governed by Ohm’s law, the memcapacitor encodes synaptic connection weights in physical capacitance states and performs computation through transient displacement-charge accumulation governed by the fundamental relation where charge equals capacitance multiplied by voltage. Because output signals are generated strictly as transient displacement charges during voltage transitions rather than sustained conductive current flows through the memory elements, the capacitive approach intrinsically circumvents steady-state direct-current power dissipation and mitigates the severe heat accumulation commonly encountered in dense three-dimensional crossbars.

A critical structural advance introduced in the fabricated memcapacitor array is the continuous in-situ-doped N+ bottom readout electrode placed beneath a lightly doped active poly-Si channel layer. In earlier charge-trap memcapacitor architectures, attempts to read stored states encountered an acute resistance–depletion trade-off because the same semiconductor layer governed both lateral conduction and electrostatic capacitance modulation. In the newly demonstrated design, dopant gas modulation during fabrication established a sharp retrograde doping profile that decouples the lateral readout resistance from the depletion-based capacitance window. The underlying N+ region provides an uninterrupted, highly conductive lateral readout path while preserving the lightly doped channel zone necessary for electrostatic gate-induced depletion. To evaluate structural variations, the team compared n-type and p-type channel bodies, discovering that the n-type body exhibited dramatically improved array-wide uniformity for both maximum and minimum capacitance states because high electron mobility suppresses resistive voltage drops along extended word-line and bit-line paths.

The physical fabrication process for the planar memcapacitor arrays, implemented in both 12 by 24 and 24 by 48 crossbar configurations, began on a buried oxide layer followed by retrograde poly-Si channel deposition. The functional gate dielectric stack adopted a metal-organic chemical vapor deposited TiN word-line electrode positioned above an alumina, silicon nitride, and silicon dioxide stack, forming an integrated TANOS memory structure. Scanning electron microscopy confirmed a clean periodic arrangement across the completed 24 by 48 array with a gate spacing of 500 nanometers and precise layer thicknesses consisting of 50 nanometers of titanium nitride, 8.1 nanometers of aluminum oxide, 6.0 nanometers of silicon nitride, and 4.2 nanometers of silicon dioxide. Ion implantation steps established carrier reservoirs, where bottom phosphorus-doped N+ regions supply electrons during programming and boron-doped P+ regions provide minority-carrier holes required for erase operations. Electrical characterization utilized a multi-pin probe card, switching matrices, and semiconductor parameter analyzers, while array-level transient charge readouts were monitored through external operational amplifier circuits.

Pristine memory cells across the fabricated 24 by 48 array exhibited typical n-type metal-oxide-semiconductor capacitor curves with remarkable array-wide device uniformity, demonstrating an exceptionally low normalized standard deviation of less than 0.37 percent. Analog synaptic weight modulation was accomplished via incremental step-pulse programming, where 41 consecutive pulses ranging from 8 to 12 volts in 0.1-volt increments induced gradual, monotonic shifts in capacitance-voltage characteristics through progressive electron trapping in the silicon nitride layer. Bidirectional tuning was demonstrated using incremental step-pulse erase pulses that induced hole-assisted detrapping, yielding symmetric analog potentiation and depression characteristics. During programming and erasing operations, unselected word lines and bit lines were held at intermediate pass voltages that prevented unintended carrier channel formation. Monitoring output displacement charges confirmed that neighboring unselected cells remained entirely undisturbed, verifying robust cell-level selectivity across the crossbar matrix without accidental programmatic write disturbance.

Employing closed-loop program-and-verify pulsing schemes, the researchers successfully generated 16 distinct analog charge states spanning 9.5 picocoulombs to 11.0 picocoulombs in 0.1-picocoulomb increments, achieving a precise 4-bit synaptic weight resolution. Achieving these target states required an average of 45 program-and-verify iterations, resulting in a mean absolute programming error of 0.0178 picocoulombs with 95 percent of absolute errors remaining below 0.0465 picocoulombs. Extended retention measurements confirmed that all 16 programmed charge levels remained clearly distinguishable over 10,000 seconds at room temperature, while high-capacitance and low-capacitance state evaluations at 85 degrees Celsius exhibited stable memory window preservation with predictable, minor thermal detrapping drift. Furthermore, transient read endurance testing consisting of one million continuous read cycles revealed no measurable drift or degradation in output charge, confirming that the transient displacement-charge readout mechanism imposes negligible electrical stress on the gate dielectric stack.

The computational fidelity of the fabricated hardware was validated through array-scale image mapping and parallel vector–matrix multiplication readouts. A 4-bit grayscale image mapped directly onto the 24 by 48 array demonstrated uniform write-verify convergence and disturbance-free bit alterations across all 1,152 physical cells. To verify analog summation accuracy, the displacement charge of individual cells was measured independently and compared against the total charge accumulated when multiple cells were activated simultaneously. The measured multi-cell displacement charge followed the ideal linear sum of single-cell contributions with an error below 0.227 percent. Readout was accomplished by transferring displacement charge during voltage discharge into an external reference capacitor connected to an operational amplifier. By integrating transient currents across a 1-microsecond transition window, single-cell read energies were extracted at approximately 117 picojoules for high-capacitance erased states and 42.8 picojoules for low-capacitance programmed states.

To assess vertical scaling potential, three-dimensional technology computer-aided design simulations evaluated gate-all-around cylindrical charge-trap memcapacitors integrated along common vertical bit-line channels mimicking hundreds of physical 3D NAND layers. The simulated structures demonstrated negligible layer-to-layer electrostatic variation and tight spatial carrier confinement during write operations. However, the simulations revealed that reducing gate lengths and intercell spacing to modern 3D NAND dimensions introduces systematic verification-to-inference electrostatic mismatches. Because single cells are biased individually during program verification while all cells are simultaneously biased during parallel inference, variations in intercell silicon surface potential alter the effective channel accumulation area beneath selected word lines. This intercell interference caused parallel vector–matrix multiplication output charge to deviate from verified values, introducing a 10.58 percent inference error when intercell spacing was reduced to 30 nanometers.

To eliminate this scaling-induced distortion, the team devised an intercell trapped-charge stabilization scheme. Prior to weight programming, a single global initialization pulse is applied across all word lines to uniformly inject electrons into the intercell silicon nitride regions, permanently pinning the underlying silicon surface in depletion regardless of subsequent operational bias conditions. Because the intercell regions remain consistently depleted during both individual verification and parallel inference phases, the effective capacitive accumulation area per cell becomes identical in both modes. TCAD simulations indicated that this intercell trapped-charge method suppresses vector–matrix multiplication error by 84.6 percent at 30-nanometer spacing, lowering overall computation error to just 1.63 percent. Experimental validation on fabricated 500-nanometer test structures confirmed that this error reduction remained stable between 76.7 percent and 78.2 percent over 10,000 seconds without observable electrostatic degradation.

Finally, the researchers evaluated system-level pattern recognition performance using a hybrid hardware–software spiking neural network implementing a modified VGG-8 architecture trained on the CIFAR-10 image classification dataset. While preceding convolutional and dense layers operated in software simulation, the final 24 by 10 classification layer was mapped directly onto physical 24 by 10 memcapacitor arrays using differential pairs to encode signed synaptic weights. Under a 10 percent hardware weight-transfer error margin, the hybrid spiking network achieved an 88.01 percent classification accuracy, which is within 0.16 percentage points of the 88.17 percent accuracy achieved by pure software spiking networks. Although peripheral circuitry, high programming voltages between 8 and 12 volts, and transient charging times must be further optimized, geometric scaling projections based on intrinsic cell capacitance indicate read energy potentials in the femtojoule range, establishing vertically stackable memcapacitors as a highly viable platform for dense, low-power neural computing hardware.

Subject of Research: Vertically stackable charge-trap memcapacitors for capacitive in-memory vector-matrix multiplication

Article Title: Capacitive in-memory vector–matrix multiplication computing with charge-trap memcapacitors

Article References: Yu, J., Hwang, H., Kim, H., & Choi, W. Y. (2026). Capacitive in-memory vector–matrix multiplication computing with charge-trap memcapacitors. Nano Convergence, 13(1), Article 42. https://doi.org/10.1186/s40580-026-00575-9

Image Credits: AI Generated

DOI: 10.1186/s40580-026-00575-9

Keywords: in-memory computing, memcapacitor, vector-matrix multiplication, charge-trap flash, 3D NAND, neuromorphic hardware, capacitive readout, spiking neural networks, displacement charge, TANOS stack, crossbar array, intercell trapped charge, analog computing, deep learning accelerators, hardware neural networks, closed-loop programming

Cite Scienmag News

Cassandra Pierce. (November 17, 2026). Charge-Trap Memcapacitor Array Enables Linear Capacitive Vector-Matrix Multiplication for In-Memory Computing. Scienmag. https://scienmag.com/charge-trap-memcapacitor-array-enables-linear-capacitive-vector-matrix-multiplication-for-in-memory-computing/

Cassandra Pierce. "Charge-Trap Memcapacitor Array Enables Linear Capacitive Vector-Matrix Multiplication for In-Memory Computing." Scienmag, 17 November 2026, https://scienmag.com/charge-trap-memcapacitor-array-enables-linear-capacitive-vector-matrix-multiplication-for-in-memory-computing/. Accessed 17 November 2026.

Cassandra Pierce. "Charge-Trap Memcapacitor Array Enables Linear Capacitive Vector-Matrix Multiplication for In-Memory Computing." Scienmag. November 17, 2026. https://scienmag.com/charge-trap-memcapacitor-array-enables-linear-capacitive-vector-matrix-multiplication-for-in-memory-computing/

Tags: 3D NANDanalog computingcapacitive readoutcharge-trap flashclosed-loop programmingcrossbar arraydeep learning acceleratorsdisplacement chargehardware neural networksin-memory computingintercell trapped chargememcapacitorneuromorphic hardwarespiking neural networksTANOS stackvector matrix multiplication
Share26Tweet16
Previous Post

Bayesian Framework Improves Wind Turbine Fault Detection Precision-Recall Metric in Simulation

Next Post

Artificial Intelligence Enhances Reliability Assessment and Design in Advanced Electronics Packaging

Related Posts

Maize Income Linked to Lower Poverty Measures Among Bangladeshi Smallholders
Earth Science

Maize Income Linked to Lower Poverty Measures Among Bangladeshi Smallholders

November 17, 2026
Low birth weight accounts for 6.8 million disability-adjusted life years in Bangladesh
Medicine

Low birth weight accounts for 6.8 million disability-adjusted life years in Bangladesh

November 17, 2026
Preoperative HALP Index Nonlinearly Predicts Biochemical Recurrence Following Radical Prostatectomy
Medicine

Preoperative HALP Index Nonlinearly Predicts Biochemical Recurrence Following Radical Prostatectomy

November 17, 2026
Rice Protein Complex Enables Plants to Detect and Resist Fungi
Science News

Rice Protein Complex Enables Plants to Detect and Resist Fungi

November 17, 2026
Treat-to-Target Strategies Advance Clinical Outcome Assessment in Juvenile Systemic Lupus
Science News

Treat-to-Target Strategies Advance Clinical Outcome Assessment in Juvenile Systemic Lupus

November 17, 2026
Real-World Review Finds Most Patients Persist on Anti-CGRP Migraine Therapies
Medicine

Real-World Review Finds Most Patients Persist on Anti-CGRP Migraine Therapies

November 17, 2026
Next Post
Artificial Intelligence Enhances Reliability Assessment and Design in Advanced Electronics Packaging

Artificial Intelligence Enhances Reliability Assessment and Design in Advanced Electronics Packaging

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Maize Income Linked to Lower Poverty Measures Among Bangladeshi Smallholders
  • Low birth weight accounts for 6.8 million disability-adjusted life years in Bangladesh
  • Preoperative HALP Index Nonlinearly Predicts Biochemical Recurrence Following Radical Prostatectomy
  • Rice Protein Complex Enables Plants to Detect and Resist Fungi

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading