Silicon Nanowire Field-Effect Transistors (NW FETs) often encounter difficulties in preserving long-term stability and sensitivity when operating in complex biological fluids. In this work, diamond-like carbon (DLC) films ranging from 1 nm to 14 nm were introduced to modify the surface silicon dioxide layers with the aim of enhancing their long-term stability. The DLC-covered structures exhibited excellent chemical and physical reliability. The mechanism of long-term stability was explored using impedance spectroscopy. Moreover, DLC-covered NW FET samples in C-reactive protein (CRP) solutions showed a sensitivity of 100 mV per decade, which is promising for biosensor applications.
Recent advances in data-driven modelling have positioned Machine Learning (ML) as a potential alternative to conventional Thermionic Emission (TE) based approaches for extracting diode parameters from measured current–voltage (IV) characteristics. However, purely statistical optimization may produce parameter estimates that lack physical consistency with the underlying semiconductor transport mechanisms. This issue is particularly important for Schottky barrier diodes (SBDs), where parameters such as the Schottky barrier height (ΦB), ideality factor (η), saturation current (Is) and series resistance (Rs) are governed by well-established semiconductor physics. In this paper, Pd/Al-doped ZnO/p-Si SBDs are investigated using both physics-based characterization and data-driven regression frameworks. The performance of Artificial Neural Networks (ANN), Kernel Ridge Regression (KRR), Gaussian Process Regression (GPR) and Support Vector Regression (SVR) is systematically compared with conventional TE-based parameter extraction. The objective is to determine whether ML models can provide accurate parameter estimation while maintaining consistency with the physical characteristics of the proposed SBD.Al-doped ZnO (AZO) thin films (~150 nm) were deposited on p-Si by RF magnetron sputtering and annealed at 100 °C and 300 °C and Pd Schottky contacts were formed on the AZO surface. Structural (XRD), morphological (SEM/AFM), compositional (EDX), optical (UV–Vis) and electrical (IV and CV) characterizations were carried out, and the junction parameters η, Is and ΦB were extracted experimentally using TE theory, giving η ≈ 4.6 and 4.1 and ΦB = 0.63 and 0.86 eV for the 300 °C and 100 °C devices respectively. Characterization and modelling benchmarks were executed for both annealing temperatures to assess model adaptability under severe tracking constraints.For the 300 °C annealed device, GPR conserves the physical reality with η = 2.37 and ΦB = 0.90 eV, closely mirroring the TE reference targets (η = 2.34, ΦB = 0.91 eV). SVR achieved a low statistical error (MSE = 0.009) but at the expense of physical validity, yielding a non-physical ideality factor (η = 15.19). For the 100 °C annealed device, tracking complexity increased because of lower current magnitudes; SVR again gave the lowest numerical error (MSE = 0.008, η = 1.75) but a physically inconsistent parameter set, whereas GPR maintained structural consistency (ΦB = 0.91 eV, with Is matching the reference value) despite an overprediction of the ideality factor (η = 12.25). ANN and KRR configurations show non-physical oscillations and forward-bias current overshoots for both devices. This deviation is attributed to the algorithmic omission of Rs and interfacial trap-state boundary conditions. This study shows that exceptional statistical correlation (R > 0.999) can coexist with structurally invalid physical models. Consequently, it also underscores the critical necessity of transitioning from pure statistical frameworks to Physics Informed Neural Networks (PINNs), which explicitly encode thermionic-emission boundaries to prevent mathematical overfitting in Wide Band Gap (WBG) semiconductor diagnostics.
We propose a charge-based small-signal impedance model for the trap states in the IZO thin-film transistors (TFTs) that represents the underlying complex physical processes by a concise equivalent circuit, thereby facilitating the decoupling of free-electron and trap-state contributions, the extraction of trap-state parameters and unified analysis of trap-related frequency-dependent capacitance-conductance-voltage characteristics. In this model, a relaxation equation for trapped charge is derived based on the Shockley-Read-Hall (SRH) theory. Combining the voltage distribution relation and charge conservation law, an impedance expression including the trap-state response, free-electron response, and gate insulator capacitance is obtained. With series resistance further taken into account, the derived impedance expression is represented by an equivalent circuit, which links the macroscopic device response with free-electron modulation, trap occupation variation, and trap-state relaxation. Furthermore, multi-frequency CV and G-V measurements of IZO TFTs validate the proposed model, and the simulated results agree well with the experimental data over the experimentally investigated frequency range of 1 kHz–1 MHz. The results demonstrate that the proposed model can effectively describe the trap-state response in IZO TFTs and provide a physical framework for trap-state parameter extraction and frequency-dependent analysis in oxide TFTs.
This work presents a neural network compact model for gate leakage in InP based pseudomorphic high electron mobility transistors (pHEMTs), trained exclusively on measured data from two epitaxial technologies: a conventional single-δ-doped device and an advanced double-δ-doped device. A unified dataset combining all Ig–VGS curves for both structures is used to train a modest size multilayer perceptron that maps gate–source voltage, drain–source voltage, and a structure flag to log10Ig. An ablation study confirms that gate length and the structure flag are perfectly collinear in the present two device dataset, leading to the adoption of a 3 input model VGSVDSs. The resulting model attains a global test RMSE of about 0.12 decades, while accurately reproducing measured leakage characteristics across multiple drain biases and gate bias regions. Bias space error maps confirm that errors remain small over most of the operating space, particularly for the advanced double-δ-doped device. By avoiding explicit regime partitioning and technology specific analytical formulations. The proposed ANN-based model provides an accurate data-driven description of the measured reverse-bias gate-leakage characteristics within the investigated bias range. Its value lies in fitting two measured InP pHEMT technologies within one framework and without manual regime partitioning; however, its accuracy has not been directly benchmarked against a separately extracted physics-based analytical model on the same dataset, and additional measured data would be required to extend it to other device structures.
A shared voltage-domain global shutter CMOS image sensor (CIS) with intermediate charge storage nodes (SNx) and 2 × 2 charge-domain binning is presented to support full-resolution mode and provide projected signal-to-noise ratio (SNR) improvement in binning mode. The proposed pixel groups four pinned photodiodes (PPDs) into a shared cell with a common floating diffusion (FD), while staged charge transfer through the SNx decouples exposure termination from sequential readout. TCAD simulations show controlled charge transfer and good linearity in full-resolution and binning modes. In binning mode, the simulated FD response achieves a 3.89× enhancement, close to the ideal 4× scaling. Post-layout extraction gives an FD capacitance of 1.6 fF and a conversion gain (CG) of 100 μV/e−. A programmable gain amplifier (PGA) with auto-zero calibration and capacitor-reused correlated multiple sampling (CMS) further supports low-noise readout. At ×8 PGA gain, the simulated total read noise decreases from 2.41 e− in full-resolution mode to 1.89 e− in CMS binning mode, corresponding to a projected maximum SNR gain of 4.95× under low-light conditions, where read noise far exceeds photon shot noise. As photon shot noise becomes dominant at higher signal levels, the projected SNR gain approaches 1.97 × .
