a Schematic of the pixel structure, comprising laminated layers including a graphene top electrode, an electrolyte layer, and a back pixel electrode. Common voltage, V_G, is applied to the graphene layer to compensate unintentional doping, thereby enabling the device to operate near the Dirac point. The charge on the capacitor is regulated by the duration of gate voltage pulses applied to the gate line. b Photograph of the fabricated device consisting of an active-matrix array of 640 × 480 pixels. A binary voltage pattern (V_DH, V_DL) is produced by a chip-on-glass display driver controlled by an external microcontroller. c Photograph of the TFT back panel, displaying the sub-wavelength pixels, each consisting of a top ITO electrode connected to the pixel capacitor with a via through the polymer layer. The inset illustrates the unit cell with a size of 185 µm. A double back-gated a-Si thin-film transistor, with a channel length of 5 µm and a width of 75 µm, governs the charge/voltage on the pixel capacitor, which is formed between the middle capacitor electrode and the ground plane. d Scanning electron micrograph revealing the cross-section of the a-Si thin-film transistor. e Representative output (I_d vs V_d) and transfer (I_d vs V_g) characteristics of the n-type a-Si TFT. f Analysis of the pixel voltage’s dependence on the width of the applied gate pulses. The inset presents the circuit model of the pixel, incorporating the probe resistance (1 MΩ) and capacitance (16 pF), resulting in a 1.5 ms charging time. g Scanning electron microscope image of the surface, with charging contrast induced by the alternating voltages of V_DH and V_DL on the pixels.

a Schematic of the experimental setup used for measuring transmission pattern using a 64 × 64 pixels THz camera. b Representative images generated by the THz modulator array. c Spectrum of the THz modulation for 3 different devices with single, double, and three layers of graphene d Drain voltage dependence of the modulation. The inset shows the circuit model where C₀ and C_G represent the pixel capacitance and the voltage-dependent capacitance of the electrical double layer, respectively. Note that due to the dynamic charging of the pixel capacitor, the drain voltage does not directly appear on the pixel. e Variation of the transmission at 100 GHz after applying consecutive charging and discharging pulses. The inset shows the transition region with switching time around 1 ms.

To optimize THz transmission modulation, we fabricated modulators with single, double, and three layers of graphene. Figure 2c shows a comparison of the THz modulation spectrum derived from various devices using a time-domain THz spectrometer. The device with the double layer of graphene marked a significant enhancement over the one with a single layer; however, the introduction of a third layer resulted in considerable insertion loss without any improvement in modulation efficiency. This diminishing performance is presumably attributable to the inefficient gating of the subsequent graphene layers. All devices presented in this work were fabricated using double-layer graphene.

The drain voltage and the duration of the gate pulse are the two main control parameters of the modulator array. Figure 2d illustrates the dependence of the THz modulation (@1 THz) on the drain voltage. To quantify the charging time of the pixel, we applied pulses of V_DH and V_DL and monitored the transmission with a fast zero-bias Schottky diode attached to a horn antenna (Fig. 2e). We observed that the charging and discharging time can be reduced to 1 ms with short gate pulses of 30 µs (see the inset in Fig. 2e and Supporting materials Figs. S6–S8). To further understand the limitations of the device, we performed electromagnetic simulations of the device using modified Poisson–Nernst–Plank (PNP) equations to include polarization, diffusion, and steric effects of the ionic liquid electrolyte (See Supporting Materials Fig. S9)^30,31,32. We have assessed the time response of charging of pixel capacitors and contribution of steric effects and clustering dynamics to this charging process. In the numerical simulations, we transiently applied voltage to one pixel electrode and grounded the graphene layer while monitoring the charge accumulation. These results suggest that the steric effects in IL electrolyte determine the charging time of the pixel capacitor. By engineering the size of the ions, the switching speed of the modulator can be further improved. However, the complexity of the modulator introduces several technical factors that limit the update rate of the patterns on the device. The primary constraint is the communication capacity of the chip driver, which is restricted by the speed of the serial communication interface. Additionally, factors such as the row scan rate and electronic delays within the circuit further impact the device’s maximum achievable speed. Employing more advanced display drivers could significantly reduce the update time, bringing it closer to the intrinsic limit set by electrostatic gating (see Supporting Materials Fig. S10).

Reconfigurable THz reflection

Next, we investigate the reconfigurable THz reflectivity. To obtain large phase and intensity modulation, we have designed the device to operate around a reflection singularity, which emerges from the interference of multiple reflections from the top polymer layer (t₁ = 70 µm), graphene and TFT array³³. In this configuration, graphene behaves as a tuneable reflectivity mirror controlling the ratio of the interfering components (Fig. 3a). This multilayer structure behaves as a tuneable coupled cavity. To observe the resonance around at THz frequencies, we increased the thickness of the electrolyte layer to t₂ = 25 µm. This configuration enables continuous tunability of the resonance frequency between {f}_{\max } \sim \frac{1}{{t}_{1}} and {f}_{\min } \sim \frac{1}{{t}_{1}+{t}_{2}}. When graphene is reflective (@V_DH), the resonance frequency is determined by the top polymer coating (t₁ ~ 70 µm), however around the Dirac point (@V_DL), the graphene layer is transparent therefore the resonance is mainly defined by the total thickness (t₁ + t₂ ~ 95 µm). We mounted the device on a motorized xy-stage and measured the time-domain THz reflection spectrum using a 4-mirror reflection head. Figure 3b, c shows the variation of reflection amplitude and phase spectrum as the pixel voltage changes from V_DL = −2.4 V to V_DH = 10 V. During this process, the resonance frequency was observed to vary from 1.6 to 1.9 THz. In addition to the continuous tunability, the reflection spectrum goes through a perfect absorption indicating a topological switching clearly evidenced by a step-like jump in the phase spectrum (Fig. 3c). Besides the tuneable phase shift, this phase jump provides a modulation up to 2π. It is important to note that this phase jump is not a branch error in phase measurements, but a geometrical phase due to winding a singularity point. To characterize reconfigurable THz patterns, we generated binary images on the modulator array and performed a raster scan of the THz reflectivity to map these patterns. The spatial resolution of our THz spectrometer is approximately 600 µm, which is sufficient to resolve fine details in the THz reflectivity patterns. Figure 3d shows a reflectivity map (\,\Delta R=R\left({V}_{{DH}}\right)-R\left({V}_{{DL}}\right)) at 1.5 THz of checkerboard pattern generated by a binary image on the modulator. It is also possible to generate a grayscale THz pattern by defining a supercell by grouping a 3 × 3 array of pixels with binary voltage (see Supporting Materials Fig. S11). In this configuration, a 10-step grayscale value can be achieved by averaging the binary values of the pixels in the supercell. Figure 3e shows a reflection map of a fork diffraction grating obtained by superimposing linear diffraction grating pattern with a spiral phase structure. Besides its intricate structure, this pattern contains a dislocation with a topological charge, enabling diffracted beams with vortices, which can be used in various applications, like beam shaping and holography. To test the limits of our approach we also mapped a complex high-resolution pattern, a portrait of Queen Elizabeth II (Fig. 3f) showing the grayscale THz reflectivity.

Fig. 3: Reconfigurable THz reflective surfaces.

a Structure of the modulator showing the interfaces responsible for the multiple THz reflection resulting a tuneable resonance behavior. b, c shows the variation of the intensity and phase spectrum as the pixel voltages switched between V_DL = −2.4V–V_DH = 10 V and recorded with equal time intervals. The step-like phase modulation is due to a topological phase transition enabled by the reflection singularity. d–f show THz reflectivity map of a checkerboard, a grayscale fork diffraction grating and a grayscale portrait of Queen Elizabeth II. The inset shows the digitized binary image using 3 × 3 supercell. (By permission of © Royal Collection Enterprises Limited 2025 | Royal Collection Trust).

Single pixel THz imaging

To showcase a practical application of these large-scale modulators, we built a single-pixel camera capable of imaging concealed metallic objects. Figure 4a shows the experimental setup used for the system. The modulator array is illuminated by a collimated beam of 100 GHz light. The total transmitted light is collected and measured with a pyroelectric sensor. A metallic object placed in a paper envelope is positioned in front of the modulator. The core working principle of the camera is based on compressed sensing algorithm^13,34,35,36. This algorithm enables the reconstruction of the spatial information of the object (X, 1 × 1024 = 32 × 32 the vector representing the transmittance of the object) by measuring a series of structured transmission patterns generated by the modulator. Here, the shadow of the object on the modulator is reconstructed by changing the transition pattern of the modulator.

Fig. 4: Single-pixel millimeter wave camera.

a Experimental setup used for the single pixel camera consisting of an mm-wave source (100 GHz), illumination and collection optics, reconfigurable modulator array and a single pyroelectric sensor. The metallic object is placed in front of the modulator. b, c show the reconstructed images of a wrench and razor blade. The scale bars are 2 cm. d shows the convergence of the error function as a function of mask number used in the reconstruction algorithm.

We created a set of Hadamard transmission masks (M, 1024 × 1024 matrix containing the mask set) on the modulator and recorded the transmitted intensity (I, 1 × 1024 intensity array). Hadamard matrices are particularly well-suited for imaging applications due to their orthogonal rows, which yields minimal cross-talk between patterns. Using a Hadamard matrix of size 1024, we generated 1024 patterns of 32 × 32 images. Each mask encodes specific spatial information about the object, and the transmitted intensity was recorded for each pattern. The recorded intensity values were then used to reconstruct the object (X) through a computational decoding process that leverages the orthogonality of the Hadamard matrix. For each mask, we measured the differential modulation by changing the pixel voltage between V_DL and V_DH (see Supporting Materials Fig. S12). This approach generated positive or negative values, which is required to construct the orthogonal set of Hadamard masks to decode patterns as X = M⁻¹ I. Figure 4b, c shows reconstructed 32 × 32 pixel images of a wrench and a razor blade. It is important to note that the resolution of the imaging system is limited by the wavelength of the source (λ = 3 mm) used for the illumination. Notably, the cutout width of razor blade, ~2 mm, is visible in the reconstructed image. The graph in Fig. 4d shows the convergence of the image error function (χ²) with the mask number used in the reconstruction process. It should be noted that the modulation performance of our devices in reflection mode is higher than that in transmission mode. However, due to the current limitations of our time-domain THz system, we were unable to conduct single-pixel imaging in reflection mode. Our current setup primarily supports transmission mode, restricting our ability to fully explore and demonstrate the imaging potential of the reflection mode. Addressing these technical challenges in future studies would enable more detailed spectroscopic image images of objects.

Dynamic beam steering

Our devices enable dynamic beam steering and beam forming, showcasing possible uses for THz communication. To demonstrate this capability, we generated a dynamic diffraction pattern on the modulator and imaged the resulting diffracted beams. Figure 5 illustrates the experimental setup used for dynamic beam steering through the device using a binary diffraction pattern. In the experimental setup, a 6-in. Teflon lens with a focal length of 15 cm was used to focus the 100 GHz diffracted beams onto a THz camera. By changing the period of the diffraction patterns, we were able to control the direction of the diffracted beam. For instance, applying binary grating patterns with periods ranging from 8 mm to 20 mm resulted in distinct diffracted beams with tuneable angles, as shown in the images.

Fig. 5: Dynamic beam steering via diffraction.

a Experimental configuration used for the imaging of the diffracted beams from the device. The diffracted beams are focused on the camera using 6” Teflon lens with f = 15 cm. b Images of the diffracted beams generated by the modulator with binary grating pattern with period changing from 8 to 20 mm. c Images of the diffracted beams with orbital angular momentum of l = 1 and l = 2. The mask generated on modulator converts a Gaussian beam to a donut-shaped Laguerre–Gaussian beams with a phase vortex. d Calculated intensity profile of a Laguerre–Gaussian beams with l = 1 and l = 2.

Additionally, our device has the capability to generate complex THz beams with orbital angular momentum (OAM). When a fork diffraction pattern with a dislocation singularity was applied to the modulator (see Supporting materials Fig. S13), an input Gaussian beam was transformed into donut-shaped Laguerre–Gaussian beams with phase vortices corresponding to OAM values of l = ±l. The characteristic donut shape of the diffracted beam serves as evidence of the vortex and associated angular momentum. By varying the order of the dislocation (i.e., the charge of the singularity) in the fork diffraction pattern, we controlled the angular momentum of the diffracted beam, with Fig. 5c showcasing a diffraction pattern corresponding to l = ±2. For comparison, we calculated the intensity profiles of Laguerre–Gaussian beams, as shown in Fig. 5d. For comparison, we also calculated the intensity profiles of Laguerre–Gaussian beams, which are shown in Fig. 5d. As the angular momentum of the beam increases, the size of the vortex becomes larger, reflecting the expanding vortex structure. This demonstration highlights the ability of the device to dynamically manipulate beam profiles, enabling more complex optical operations.

As conclusion, by merging graphene modulators and the active-matrix TFT technology, we demonstrate large-area reconfigurable surfaces with sub-wavelength programmable elements that can manipulate THz- and mm-waves in real time. These devices can create reconfigurable spatio-temporal patterns of THz light both in transmission and reflection modes. We demonstrate that the reflection performance of these devices can be significantly improved by operating around a reflection singularity which creates large phase modulation. It should be noted that the observed phase modulation generated with these devices are associated with large insertion losses due to the tuneable THz absorption of the graphene layer. Our results suggest that these surfaces can be used as a spatial light modulator to dynamically steer THz light or to create complex beams, which is a critical capability required for next-generation communication systems. Furthermore, the proof-of-concept demonstration of the single-pixel THz imaging and the dynamic beam steering via diffraction highlight potential for applications in security, medical imaging, and industrial inspection, where non-invasive inspection methods are essential.