Customization of silicon photodiodes (SiPDs) for specific target wavelengths has become increasingly important in the rapidly evolving domains of optical communication, medical devices and autonomous driving technologies. This study pioneers the use of machine learning for the structural optimization of SiPDs, leveraging simulation data from Synopsys TCAD to identify optimal configurations without the need for extensive physical prototyping. We aim to predict photodiode performance across a range of design parameters by employing XGBoost (XGB) and light gradient-boosting machine (LGBM) algorithms, thus streamlining the design and development process for high-performance, cost-effective devices. Our findings indicate that the LGBM algorithm provides a more accurate prediction of photodiode responsivity than XGB, as evidenced by lower root mean square error (RMSE) values. Specifically, the LGBM model achieved an RMSE of 0.282x10 ^-4 for datasets comprising 50% of the data, markedly outperforming the XGB model, which posted an RMSE of 0.754x10 ^-3 under comparable conditions. This work highlights the potential of employing machine learning in the optimization process of SiPDs, offering a path toward reducing the resources required for device design and optimization. The approach demonstrates the effectiveness of machine learning in accelerating the development of photonic devices, enabling more efficient and targeted optimization of photodiode structures for specific applications.

We demonstrate high responsivity in a polysilicon (Poly-Si) near-infrared photodetector by reducing the adverse effects of grain boundaries. The presence of grain boundaries is a serious problem in polysilicon since they act as traps and degrade the electrical characteristics. However, the relation between grain boundaries and photodetector performance has not yet been reported. Here, the effect of grain boundary on the resistivity and responsivity of Poly-Si devices is investigated. The resistivity of Poly-Si is higher than that of crystalline Si (c-Si) at the same doping concentration due to the potential barrier of the grain boundary, and its degradation of electrical characteristics is mitigated here by increasing the doping concentration. The photodetector with a highly-doped polysilicon layer exhibits a responsivity of 0.48 A/W at 900 nm, which is almost the same as that of c-Si photodetectors. In contrast, polysilicon devices with the same doping concentration as c-Si devices show degraded responsivity due to grain boundary traps. In addition, to improve the responsivity at target wavelengths of 900 nm and 1,064 nm, we deposited an indium-tin-oxide layer, which reduces the surface reflectance on the Poly-Si photodetector. The responsivity is improved by 22.9% and 50.0% at 900 nm and 1,064 nm respectively compared to devices without the ITO. The results here confirm the ability to fabricate low cost Poly-Si photodetectors with high responsivity which show similar or superior optical response compared to commercial c-Si devices.

We experimentally demonstrate Si-based electrolyte-gated transistors (EGTs) for detecting urea. The top-down-fabricated device exhibited excellent intrinsic characteristics, including a low subthreshold swing (SS) (~80 mV/dec) and a high on/off current ratio (~10⁷). The sensitivity, which varied depending on the operation regime, was analyzed with the urea concentrations ranging from 0.1 to 316 mM. The current-related response could be enhanced by reducing the SS of the devices, whereas the voltage-related response remained relatively constant. The urea sensitivity in the subthreshold regime was as high as 1.9 dec/pUrea, four times higher than the reported value. The extracted power consumption of 0.3 nW was extremely low compared to other FET-type sensors.