Programmable ferroelectric bionic vision hardware with selective attention for high-precision image classification

Human vision handles large volumes of sensory information, and selective attention enables efficient processing by emphasizing salient regions and suppressing non-salient background. Implementing selective visual attention in hardware remains challenging because conventional CMOS-based approaches are bulky, complex, and computationally expensive, with sensing and processing separated. High-efficiency hardware attention requires large dynamic-range memory states. The study asks whether ferroelectric polarization can be leveraged to realize programmable, in-sensor selective attention for visual perception. By exploiting tunable remnant polarization in P(VDF-TrFE) to modulate photocarrier tunneling and retention, the authors aim to integrate sensing and processing in a compact device that selectively emphasizes pivotal optical inputs and suppresses others, improving downstream machine-vision tasks.

Prior hardware implementations of selective attention based on CMOS and conventional transistors have large footprints and high computational cost, and typically separate the sensory unit from processing, complicating synchronous handling of signals. Recent bioinspired optical sensing devices have integrated neuromorphic functions such as visual recognition and light adaptation into compact sensors, showing promise for optoelectronics, but generally lack selective attention capabilities. Ferroelectric materials like P(VDF-TrFE), with tunable remnant polarization and coercive voltage acting as a nonvolatile threshold, have been explored for nonvolatile memory and synaptic devices, suggesting potential to provide multi-state memory and modulate photonic synaptic behavior for attention-like processing.

The authors fabricate a ferroelectric bionic vision hardware (FeBVH) device using a bottom-gate, top-contact ferroelectric transistor architecture. PDVT-10 serves as the organic semiconductor channel. CdSe/ZnS quantum dots (QDs) are blended with PDVT-10 to provide photogeneration of carriers upon illumination. A P(VDF-TrFE) ferroelectric layer is used to provide programmable polarization; an Al2O3 layer is inserted to reduce trapping and improve hysteresis behavior. Device operation relies on photogenerated holes from QDs tunneling into the PDVT-10 HOMO under illumination, increasing channel conductance; recombination dynamics yield synaptic-like decay. Ferroelectric polarization modulates the energy barriers and carrier dynamics to tune volatility: positive polarization favors hole injection and impedes recombination (nonvolatile LTP-like behavior), while negative polarization impedes injection and accelerates decay (volatile STP-like behavior). Experimental characterization includes: (i) EPSC measurements under monochromatic light pulses of varying wavelengths (365, 520, 650 nm) and intensities (e.g., 1 µW/cm²) and durations (100–500 ms); (ii) fitting decay with an exponential to extract time constants; (iii) Kelvin probe force microscopy (KPFM) to measure surface potential of PDVT-10/QDs under illumination; (iv) absorption spectra of QDs, PDVT-10, and composite films; (v) double-sweep transfer characteristics under different wavelengths and gate biases, noting polarization effects beyond the coercive voltage; (vi) programming multiple polarization states using gate pulses (e.g., +32 to +50 V, and negative biases to -50 V) to create multiple nonvolatile current states and assess retention over 50 programming cycles; (vii) assessing paired-pulse facilitation (PPF) and multi-pulse current gain; (viii) constructing a 5×5 pixel array to encode and read out images under UV exposure (pattern "E"), measuring per-pixel EPSC under positive and negative polarization to evaluate selective attention via time-dependent signal-to-noise ratio (SNR); (ix) implementing image attention processing where mixed-wavelength stimuli are applied per pixel and time-evolving responses preferentially retain short-wavelength features; and (x) building a multiply-accumulate (MAC) neural network model using device conductance as inputs and drain voltages as weights to classify letter patterns (H, I, J, K, L) composed of multi-wavelength signals, training weights to minimize loss and evaluating classification accuracy with and without selective attention.

- Wavelength-selective photoresponse: 365 nm illumination elicits much larger EPSC (up to ~13 nA) and longer decay (τ ≈ 29.9 s) than longer wavelengths (e.g., 600–650 nm), consistent with higher UV absorption and increased surface potential under UV observed by KPFM. - Polarization-controlled volatility and memory strength: Positive polarization (e.g., +40 V) produces nonvolatile LTP-like behavior with memory factor ηM increasing from 34.5% to 47.1% as pulse duration increases (100–500 ms). Negative polarization (e.g., -40 V) yields volatile STP-like behavior with ηM ≈ 9.1%. - Multi-level nonvolatile states: By programming with different positive gate pulses (+32 to +50 V), the device achieves at least 10 distinct retention current states, repeatable over 50 cycles; under negative polarization, retention currents evolve differently with smaller decay constants. - Tunable gain and plasticity: EPSC can be modulated from ~2.5 nA to ~6.5 nA by pulse duration. Current gain under consecutive pulses reaches ~6.7 for positive polarization and is much smaller for negative polarization. PPF is enhanced under positive polarization and depressed under negative polarization, decreasing with larger inter-pulse intervals. - Array-level selective attention: In 5×5 arrays exposed to a UV "E" pattern, positive polarization enables clear retention of the signal with improved contrast over time, while negative or no polarization leads to rapid fading (oblivion). SNR defined as 20lg(S/N) shows retentional SNR at 40 s up to ~13.0 dB after +50 V polarization, with contrast differences up to ~13.12 dB between positive and negative polarization conditions. - Wavelength-dependent attention in image processing: Under positive polarization, short-wavelength features (e.g., blue/UV) are preferentially retained, enabling extraction of a blue "butterfly" from a scene also containing a green "leaf" as the longer-wavelength component decays faster. - High-accuracy classification with attention: A MAC-based sensory network using device conductance as inputs achieves 95.7% classification accuracy on multi-wavelength letter patterns (H, I, J, K, L) with selective attention, compared to 69.7% without attention. - Integration benefits: The device integrates sensing, signed-weight representation, and memory in a single element, reducing hardware overhead for in-sensor neuromorphic processing.

The results demonstrate that ferroelectric polarization in P(VDF-TrFE) can program the photonic synaptic behavior of PDVT-10/CdSe–ZnS QD transistors to implement selective attention at the device and array levels. Positive polarization lowers the effective barrier for hole tunneling and suppresses recombination, yielding nonvolatile retention and higher synaptic weights for salient (short-wavelength) stimuli; negative polarization produces rapid decay for non-salient inputs. This attention-like filtering enhances image contrast over time and selectively retains pivotal optical information in a compact, in-sensor architecture. At the system level, the wavelength-dependent retention enables image attention processing that improves machine vision performance, as evidenced by the substantial gain in classification accuracy (95.7% vs 69.7% without attention) in MAC-based networks. By combining sensing and computation with programmable multi-level memory states, the FeBVH addresses the challenge of bulky, power-intensive attention mechanisms in conventional hardware and offers a scalable route for neuromorphic vision systems.

The study introduces a programmable ferroelectric bionic vision hardware that emulates selective visual attention by leveraging ferroelectric polarization to tune photocarrier dynamics, volatility, and multi-level nonvolatile memory states. Devices show strong UV-selective responses, polarization-controlled LTP/STP, multi-state programmability, and array-level selective retention with high SNR. Incorporating these devices into sensory networks yields high-precision image classification (95.7% accuracy) and reduced hardware overhead via in-sensor computing. These advances enrich neuromorphic functions in bioinspired sensing devices and pave the way for future bioinspired optoelectronics and neuromorphic vision hardware.