A tunable-sensitivity phototransistor was designed and fabricated using a layer-transfer method (Methods, Fig. S1). The device is based on a MoS field effect transistor (FET) with an O-plasma-treated MoS/MoS diode inserted within the gate stack. Specifically, graphite was used to form the Ohmic contact, MoS as the channel, hexagonal boron nitride (h-BN) as the dielectric layer, and top and bottom protective layers. An O-plasma-treated MoS/MoS diodewas embedded between the h-BN dielectric layer and the bottom protective layer, serving as the photosensitive structure (Figs. 2a and S2). To elucidate the effects of O-plasma treatment on the composition and structure of MoS, comprehensive characterizations of the MoS samples before and after treatment were performed. Cross-sectional high-resolution transmission electron microscopy (HRTEM) of the h-BN/MoS/h-BN heterostructure reveals a well-defined layered structure. Mo and S are confined within the MoS layer, N aligns with h-BN, and oxygen is uniformly adsorbed on the cross-section (Fig. 2b). In contrast, the upper layer of the O-plasma-treated MoS becomes amorphous, confirmed by the presence of Mo, S, and O in energy dispersive X-ray spectroscopy (EDS) mapping (Fig. 2c). For pristine MoS, the S/Mo atomic ratio is approximately 1.95, consistent with the X-ray photoelectron spectroscopy (XPS) results (Figs. 2d and S3), and the oxygen content is about 32%, mainly originating from adsorbed oxygen (Fig. S3). After oxygen plasma treatment, the total oxygen content in the sample increases significantly to 68%, with about 36% of the oxygen incorporated as lattice oxygen. Meanwhile, the Mo and S contents are around 18% and 14%, respectively, corresponding to an approximate molecular formula of MoOS (Fig. 2e). Furthermore, plan-view HRTEM (Fig. S4), EDS (Fig. S4), and Raman characterizations (Fig. S5) further confirm the structural and compositional transition induced by O-plasma treatment.
XPS valence band spectra reveal that the valence band maximum shifts from ~0.8 eV below the Fermi level in pristine MoS₂ to ~1.2 eV after plasma treatment (Fig. S6). Ultraviolet photoelectron spectroscopy measurements further show an increase in work function from ~4.3 eV for pristine MoS to ~4.7 eV O-plasma-treated MoS. Theoretical calculations indicate that the bandgap widens with oxygen incorporation, increasing from 1.56 eV (MoS) to 1.8 eV (MoOS). Based on these experimental and theoretical results, schematic band diagrams for pristine and O-plasma-treated MoS were constructed (Fig. S7). The band alignment suggests the formation of a MoS/O-plasma-treated MoS n/n junction, which we experimentally validated by constructing both in-plane and vertical heterojunctions (Figs. S8 and S9). The working principle of the phototransistor with tunable sensitivity was shown in Fig. 2f, g. Under a negative gate bias (V), the O-plasma-treated MoS/MoS heterojunction is reverse-biased in the dark, causing most of the gate voltage to drop across the junction and keeping the MoS channel conductive (Figs. 2f and S10). Upon illumination, photogenerated carriers reduce the junction resistance, redistributing the gate voltage such that a larger portion is applied across the h-BN dielectric and MoS channel, thereby depleting the channel and turning the transistor off (Figs. 2g and S11). As the gate voltage becomes more negative, the voltage across the MoS channel increases further, enabling the device to shut off at lower light intensities, thus realizing tunable detection sensitivity. A new device symbol representing this structure is proposed (Fig. 2h).
Figure 2i shows the photoresponse behavior of the phototransistor. When the gate voltage is -2 V, the device shows only a slight current change at light intensities below 0.7 mW cm. However, within the light intensity range of 0.7-1.2 mW cm, the device exhibits a current change of nearly 10 times. In contrast, the current change in the O-plasma-treated MoS/MoS diode is only 1.6 times (Fig. S12). These demonstrate that our transistor can generate a non-linear relationship between photocurrent and light intensity, mimicking the retina system. To exclude the contribution of the MoS channel to the photocurrent, we conducted position-dependent illumination experiments, confirming that the dominant photoresponse originates from the O-plasma-treated MoS/MoS heterojunction (Fig. S13). In addition, the transistor behavior under light is different under different V, demonstrating a tunable photosensitivity by varying V.
Figure 3 shows the detailed optoelectronic performance of the tunable-sensitivity phototransistor. When a V of -9 V and a 100 ms light pulse are applied simultaneously, the device shows no obvious photoresponse when the light intensity is below 77 μW cm. However, when the light intensity further increases, the transistor current decreases abruptly. Once the light intensity surpasses 454 μW cm, the photoresponse of the device reaches saturation (Fig. 3a, b). This non-linear characteristic enables the device to effectively filter both strong-light and weak-light noise. In addition, by adjusting V, the device's response range to light intensity can be tuned. For instance, when V = -7 V, the device exhibits a light response range of 392-1061 μW cm; whereas, at -5 V, the response range is adjusted to 748-2122 μW cm. Overall, the device can precisely distinguish light intensities within the range of 77-50000 μW cm by changing V (Figs. 3b and S14).
To further demonstrate the device's ability to detect small changes in light intensity, we extracted the relationship between the current ratio and the light intensity ratio (P/P) at different V (Fig. 3c). When the light intensity changes by 3-5 times, the current ratio of the tunable-sensitivity phototransistor exceeds 10. In contrast, the current ratio of the conventional MoS phototransistor is about 5. This indicates that the tunable-sensitivity phototransistor has more than 1000 times higher capability in detecting small changes in light intensity compared to conventional photodetectors, and also outperforms previously reported gate-tunable phototransistors designed for contrast enhancement (Table S1). The tunable-sensitivity phototransistor also demonstrates a significantly higher responsivity compared to the conventional MoS phototransistor (Fig. 3d). Notably, as the light intensity increases, the responsivity of the tunable-sensitivity phototransistor gradually decreases, which is similar to that of the human retina (Fig. 3d). In contrast, the conventional MoS phototransistor lacked this characteristic (Figs. 3d and S15).
Figure 4a, b show a 3 × 3 photo sensor array based on tunable-sensitivity phototransistors, exhibiting good uniformity across all 9 transistors both in dark and under light conditions (Figs. 4c and S16). To demonstrate the ability to detect low-contrast targets, we input five sets of low-contrast signals of pattern "O" into both the conventional phototransistor array and the tunable-sensitivity phototransistor array. The light-to-background intensity ratio of pattern "O" ranges from 1.2 to 2.1. For the conventional phototransistor array, the output current ratio is only 1.3 when the light-to-background intensity ratio is 1.2, and increases modestly to 1.7 at a ratio of 2.1, which is insufficient to produce a clear image. In contrast, the tunable-sensitivity phototransistor array achieves a significantly higher current ratio of 3.4 under the same low-contrast condition (intensity ratio of 1.2), and up to 470 at an intensity ratio of 2.1, successfully enabling the recognition of a distinct "O" pattern. This demonstrates the superior performance of the tunable-sensitivity phototransistor in low-contrast target detection (Fig. 4d). Additionally, the developed array exhibited outstanding noise filtering capability. When an image "L" with increasing surrounding noise is input to the conventional phototransistor array, the output image "L" gradually becomes blurred as the noise intensity increases, due to the wide response range of the conventional detector. In contrast, under a gate voltage of -5 V, the tunable-sensitivity phototransistor array is selectively responsive to light intensities in the range of 748-2122 μW cm, making it immune to out-of-range light noise and allowing it to consistently produce a clear image (Fig. 4e).
The developed tunable-sensitivity phototransistor enables highly robust target recognition, demonstrated by integrating phototransistor arrays with an artificial neural network (ANN)-based intelligent machine vision system (Fig. 5a). An ANN model was employed to perform classification and recognition based on the image data generated by the phototransistor array. The ANN architecture consisted of an input layer, two hidden layers (the first with 128 neurons and the second with 64 neurons), and an output layer. The rectified linear unit (ReLU) was used as the activation function, and the cross-entropy function was used as the loss function. Conventional photodetectors capture all optical signals in the scene, making it challenging to distinguish low-contrast vehicle targets. On the other hand, by adjusting the gate voltage, our tunable-sensitivity phototransistor responds only to light signals within specific intensity ranges (Table S2). This capability allows accurate recognition of low-contrast vehicle targets in complex lighting conditions, whether under dim or bright lighting conditions, while effectively filtering out noise signals that interfere with target recognition.
For the low-contrast object recognition task, we selected 500 "bus" images from the CIFAR-100 dataset as positive samples, which were not included in the testing set. An additional 1000 images of other vehicles (e.g., motorcycles, bicycles, etc.) were selected as negative samples to establish a binary classification task (Methods). To evaluate the system's ability to recognize low-contrast targets, we used 500 images of buses, each with a resolution of 32 × 32 pixels, that were not included in the training set as the test dataset. The test dataset consists of five sub-datasets, each containing 100 images, with the average image contrast gradually decreasing across the datasets (Figs. 5b and S17). Figure 5c shows the recognition accuracy after 100 training epochs for each dataset. The machine vision system based on the tunable-sensitivity phototransistor achieved an accuracy exceeding 90%. In contrast, the system based on conventional photodetectors achieved an accuracy close to zero (Fig. 5c).
To evaluate the reliability of the tunable-sensitivity phototransistor-based machine vision system under complex lighting conditions, we introduced salt-and-pepper noise with densities of 5%, 10%, 20% and 30% to 500 test images (Figs. 5b and S18). These 2000 noisy images were then input into the system for recognition. Even with a noise density of 30%, the system's image recognition accuracy remained around 80% (Fig. 5d). These results highlight the superior robustness of our tunable-sensitivity phototransistor in low-contrast imaging, making it highly suitable for imaging tasks in complex lighting conditions.