A flexible capacitive photoreceptor for the biomimetic retina

Biomimetic microelectronic devices are vital to human-inspired robotics and neuromorphic computing because the human visual system encodes illumination into neural spike trains via photoreceptors and downstream retinal circuitry. Rod cells, more numerous than cones, are responsible for low-light sensing and hyperpolarize under illumination, decreasing their firing rate; in essence, photoreceptors perform sensing and initial computation in the sensor. Existing artificial retina approaches often rely on photodiodes or photodetectors that require per-pixel biasing, incurring higher static power. Capacitive neural networks better emulate neural membrane behavior with lower static power than resistive/conductive approaches. There is therefore a need for a tunable photoreceptor device whose electrical membrane-like property is directly modulated by light intensity and wavelength, enabling artificial retina networks with efficient, perception-in-sensor operation. To address this, the authors propose a flexible, hybrid perovskite–ferroelectric polymer capacitive photoreceptor (CPR) whose capacitance changes proportionally with light, mimicking rod-cell behavior. They target visible-light sensitivity with spectral response aligned to human photopic vision, long-term ambient stability, and compatibility with spike-based encoding for neuromorphic vision; they further demonstrate digit recognition using an unsupervised spiking neural network driven by the CPR model.

Prior efforts in biomimetic eyes include hemispherical cameras using silicon nanomembranes with Si photodetectors, and perovskite nanowire array photoreceptors for image sensing. Dynamic vision sensors (DVS) also use photodiodes as photoreceptors, but these approaches demand per-pixel bias and lead to higher static power. Capacitors are commonly used to mimic cell membranes in CMOS neurons, and capacitive neural networks offer better neural emulation and lower static power than resistive/conductance-based implementations. Materials-wise, PZT thin films can show dielectric changes under UV light but are not responsive in the visible, limiting applicability. Visible-light photo-capacitors have been explored via dye-sensitized semiconducting nanoparticles, phosphors, photosensitive conjugated polymers, and plasmonic Ag nanowires. Hybrid halide perovskites are compelling for optoelectronics due to strong light absorption, long carrier lifetimes, low trap density, optical anisotropy, and high carrier mobility, powering advances in solar cells, photodetectors, and emitters; however, they suffer from moisture/oxygen instability. PVDF-based ferroelectric polymers, particularly PVDF-TrFE-CFE, provide high dielectric constant, low loss, relaxor ferroelectric behavior, and high charge–discharge efficiency, making them attractive dielectric media. Kapton (polyimide) is a robust flexible substrate with mature interconnect design for wearable and biocompatible electronics. These strands motivate combining perovskite optoelectronics with ferroelectric polymers to realize a light-tunable capacitor suited to neuromorphic vision.

A flexible capacitive photoreceptor (CPR) was fabricated as a metal–insulator–metal (MIM) capacitor using a perovskite–ferroelectric nanocomposite (PFNC) dielectric comprising methylammonium lead bromide (MAPbBr3) nanocrystals dispersed in the ferroelectric terpolymer PVDF-TrFE-CFE. The PFNC layer was sandwiched between a transparent indium tin oxide (ITO) top electrode and an aluminum-coated polyimide (Kapton) bottom electrode, forming arrays of MIM capacitors on a flexible substrate. Structural and optical characterization included cross-sectional SEM to visualize PFNC morphology, TEM to confirm MAPbBr3 nanocrystals embedded in the polymer, and XRD to verify the cubic MAPbBr3 phase within the polymer matrix. UV–Vis spectroscopy assessed absorption in PFNC versus pure polymer films, and photoluminescence (PL) spectra quantified emission and monitored stability. Device electro-optical measurements were conducted by illuminating the CPR with homogenized visible LEDs at specified peak wavelengths (violet ~403 nm, green ~520 nm, greenish-yellow ~560 nm, yellowish-orange ~590 nm, red ~630 nm) over various intensities using a custom setup. Impedance magnitude |Z| and phase were measured in the frequency range 1–100 kHz under dark and illuminated conditions. From impedance and phase, a frequency-independent pseudo-capacitance (Ca) was extracted. Nyquist plots were used to assess the frequency-independent behavior, and an equivalent circuit model using parallel RC ladder networks was adopted to capture fractional-order, light-sensitive impedance changes. Baseline capacitance density in the dark was determined. Long-term stability was evaluated by storing devices in ambient air (~23 °C, ~40% RH) and monitoring PL and absorbance after extended aging (e.g., 100 weeks), with overall nanocomposite stability reported up to ~129 weeks. Functionally, the CPR model (RC network with light-tunable parameters) was integrated with a low-power spike oscillator to generate spike trains with firing rate proportional to light intensity and wavelength, and the system’s neuromorphic sensing capability was demonstrated in simulation for handwritten digit (MNIST) recognition using an unsupervised trained spiking neural network.

• The PFNC exhibits strong UV–visible absorption extending to ~563 nm and robust photoluminescence characteristic of embedded MAPbBr3 nanocrystals; the band-edge cutoff is ~2.2 eV (λ ≈ 563 nm).
• CPR impedance decreases and pseudo-capacitance (Ca) increases with increasing light intensity; the response is frequency-independent over 1–100 kHz (pseudo-capacitive behavior).
• Spectral sensitivity peaks in the greenish-yellow regime, aligning with human photopic vision: greenish-yellow illumination produces larger capacitance changes than violet or red at comparable intensities.
• Under red illumination, even at ~1000 µW/cm², the change in Ca is negligible, consistent with minimal absorption in that band.
• The dark-condition capacitance density is ~1 nF/cm², aided by the ferroelectric polymer’s high-k properties.
• The CPR’s light-tunable dielectric behavior can be modeled by a parallel RC ladder (fractional-order) network reflecting increased dielectric conductivity due to photo-generated carriers.
• Stability: Devices stored in ambient air show no significant PL peak shift or linewidth broadening after ~100 weeks; overall composite stability is reported for ~129 weeks, attributed to hydrophobic PVDF-TrFE-CFE encapsulation mitigating moisture/oxygen effects.
• System-level demonstration: Using the CPR model with a low-power spike oscillator to encode light into spike trains, an unsupervised spiking neural network achieved 72.05% recognition accuracy on MNIST, evidencing applicability to neuromorphic vision.

The study addresses the need for low-power, photoreceptor-like sensing elements that directly convert luminance into a membrane-like electrical variable. By engineering a hybrid perovskite–ferroelectric composite within a MIM capacitor, the device’s capacitance becomes a monotonic function of incident light intensity and wavelength, mimicking rod-cell behavior where illumination modulates membrane potential and firing rate. The CPR’s peak sensitivity in the greenish-yellow aligns with human photopic sensitivity, reinforcing its biomimetic relevance. Unlike photodiode-based pixels requiring bias and incurring static power, the capacitive approach supports low static power and better membrane emulation for neuromorphic circuits. The demonstrated frequency-independent pseudo-capacitance across 1–100 kHz simplifies interfacing with spike oscillators to produce light-proportional spike trains. Robust ambient stability (up to ~129 weeks) alleviates a key challenge of perovskites, while flexibility enables conformal, hemispherical form factors akin to the retina. The simulated MNIST classification confirms that such sensors, when integrated into spiking neuromorphic pipelines, can perform perception-in-sensor tasks effectively.

This work introduces a flexible, visible-light-sensitive capacitive photoreceptor based on a MAPbBr3–PVDF-TrFE-CFE nanocomposite that translates illumination into capacitance changes with spectral sensitivity matching human photopic vision. The device exhibits frequency-independent pseudo-capacitance over 1–100 kHz, low static-power-friendly operation, and remarkable ambient stability due to polymer encapsulation. Modeling and circuit integration demonstrate conversion of light to spike trains and enable neuromorphic recognition (72.05% MNIST accuracy) in simulation, highlighting its promise for artificial retina networks and energy-efficient vision systems. Future work should pursue biocompatible, Pb-free perovskite alternatives; optimize nanocomposite composition for higher sensitivity and dynamic range (including improved red response); scale to high-density, hemispherical arrays; and realize on-chip hardware integration with low-power spiking processors for real-time applications.

• Material toxicity: The composite contains lead (Pb), rendering current devices non-biocompatible; Pb-free alternatives may reduce optical performance.
• Spectral limitations: Minimal response in the red band due to limited absorption up to ~563 nm.
• Measurement bandwidth: Characterization and pseudo-capacitance behavior were confined to 1–100 kHz; behavior outside this range was not reported.
• System demonstration: Neuromorphic performance was shown via simulations with a modeled device and oscillator rather than a fully integrated hardware system.
• Partial per-pixel details: While arrays are shown, comprehensive array-level uniformity, yield, and readout/biasing architectures were not detailed in the provided text.